From adecd18bf733d9758237e21b5b8213588a8719f0 Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Mon, 16 Feb 2026 23:58:15 +1100
Subject: [PATCH 1/9] feat(bench): add page splitting/copying benchmarks and
 markdown report

Add benchmarks for page splitting, copying, and merging (#26).
Synthetic 100-page and 2000-page PDFs are generated from sample.pdf
and cached to disk for reuse.

New benchmark suites:
- splitting.bench.ts: single-page extraction, full split, batch extract
- copying.bench.ts: cross-doc copy, duplication, merging
- comparison.bench.ts: head-to-head vs pdf-lib for all of the above

Report generation:
- scripts/bench-report.ts transforms vitest JSON output to markdown
- reports/benchmarks.md committed to repo, updated by CI
- .github/workflows/bench.yml runs weekly + on push to main
---
 .github/workflows/bench.yml    |  56 +++++++
 .gitignore                     |   3 +
 benchmarks/comparison.bench.ts | 148 ++++++++++++++++++-
 benchmarks/copying.bench.ts    |  94 ++++++++++++
 benchmarks/fixtures.ts         |  95 +++++++++++-
 benchmarks/splitting.bench.ts  | 119 +++++++++++++++
 package.json                   |   1 +
 reports/benchmarks.md          | 235 +++++++++++++++++++++++++++++
 scripts/bench-report.ts        | 262 +++++++++++++++++++++++++++++++++
 9 files changed, 1010 insertions(+), 3 deletions(-)
 create mode 100644 .github/workflows/bench.yml
 create mode 100644 benchmarks/copying.bench.ts
 create mode 100644 benchmarks/splitting.bench.ts
 create mode 100644 reports/benchmarks.md
 create mode 100644 scripts/bench-report.ts

diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml
new file mode 100644
index 0000000..cae6a1f
--- /dev/null
+++ b/.github/workflows/bench.yml
@@ -0,0 +1,56 @@
+name: Benchmarks
+
+on:
+  # Run on pushes to main (to keep report up to date)
+  push:
+    branches: [main]
+  # Run weekly on Mondays at 06:00 UTC
+  schedule:
+    - cron: "0 6 * * 1"
+  # Allow manual trigger
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: write
+
+jobs:
+  bench:
+    name: Run Benchmarks
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Install dependencies
+        run: bun install --frozen-lockfile
+
+      - name: Run benchmarks and generate report
+        run: bun run bench:report
+
+      - name: Upload JSON results
+        uses: actions/upload-artifact@v4
+        with:
+          name: bench-results
+          path: reports/bench-results.json
+          retention-days: 90
+
+      - name: Commit updated report
+        run: |
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+
+          git add reports/benchmarks.md
+
+          if git diff --staged --quiet; then
+            echo "No changes to benchmark report"
+          else
+            git commit -m "docs: update benchmark report"
+            git push
+          fi
diff --git a/.gitignore b/.gitignore
index 0d003c2..fd26d33 100644
--- a/.gitignore
+++ b/.gitignore
@@ -52,5 +52,8 @@ debug/
 fixtures/benchmarks/
 fixtures/private/
 
+# Benchmark JSON results (machine-specific)
+reports/bench-results.json
+
 # Temporary files
 tmp/
diff --git a/benchmarks/comparison.bench.ts b/benchmarks/comparison.bench.ts
index 255a57d..355c33d 100644
--- a/benchmarks/comparison.bench.ts
+++ b/benchmarks/comparison.bench.ts
@@ -9,10 +9,12 @@ import { PDFDocument } from "pdf-lib";
 import { bench, describe } from "vitest";
 
 import { PDF } from "../src";
-import { loadFixture, getHeavyPdf } from "./fixtures";
+import { getHeavyPdf, getSynthetic100, getSynthetic2000, loadFixture } from "./fixtures";
 
-// Pre-load fixture
+// Pre-load fixtures
 const pdfBytes = await getHeavyPdf();
+const synthetic100 = await getSynthetic100();
+const synthetic2000 = await getSynthetic2000();
 
 describe("Load PDF", () => {
   bench("libpdf", async () => {
@@ -119,3 +121,145 @@ describe("Load, modify, and save PDF", () => {
     await pdf.save();
   });
 });
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Page splitting comparison (issue #26)
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Extract single page from 100-page PDF", () => {
+  bench("libpdf", async () => {
+    const pdf = await PDF.load(synthetic100);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+
+  bench("pdf-lib", async () => {
+    const pdf = await PDFDocument.load(synthetic100);
+    const newDoc = await PDFDocument.create();
+    const [page] = await newDoc.copyPages(pdf, [0]);
+    newDoc.addPage(page);
+    await newDoc.save();
+  });
+});
+
+describe("Split 100-page PDF into single-page PDFs", () => {
+  bench(
+    "libpdf",
+    async () => {
+      const pdf = await PDF.load(synthetic100);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    "pdf-lib",
+    async () => {
+      const pdf = await PDFDocument.load(synthetic100);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const newDoc = await PDFDocument.create();
+        const [page] = await newDoc.copyPages(pdf, [i]);
+        newDoc.addPage(page);
+        await newDoc.save();
+      }
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
+
+describe(`Split 2000-page PDF into single-page PDFs (${(synthetic2000.length / 1024 / 1024).toFixed(1)}MB)`, () => {
+  bench(
+    "libpdf",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 0, iterations: 1, time: 0 },
+  );
+
+  bench(
+    "pdf-lib",
+    async () => {
+      const pdf = await PDFDocument.load(synthetic2000);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const newDoc = await PDFDocument.create();
+        const [page] = await newDoc.copyPages(pdf, [i]);
+        newDoc.addPage(page);
+        await newDoc.save();
+      }
+    },
+    { warmupIterations: 0, iterations: 1, time: 0 },
+  );
+});
+
+describe("Copy 10 pages between documents", () => {
+  bench("libpdf", async () => {
+    const source = await PDF.load(synthetic100);
+    const dest = PDF.create();
+    const indices = Array.from({ length: 10 }, (_, i) => i);
+    await dest.copyPagesFrom(source, indices);
+    await dest.save();
+  });
+
+  bench("pdf-lib", async () => {
+    const source = await PDFDocument.load(synthetic100);
+    const dest = await PDFDocument.create();
+    const indices = Array.from({ length: 10 }, (_, i) => i);
+    const pages = await dest.copyPages(source, indices);
+
+    for (const page of pages) {
+      dest.addPage(page);
+    }
+
+    await dest.save();
+  });
+});
+
+describe("Merge 2 x 100-page PDFs", () => {
+  bench(
+    "libpdf",
+    async () => {
+      const merged = await PDF.merge([synthetic100, synthetic100]);
+      await merged.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    "pdf-lib",
+    async () => {
+      const doc1 = await PDFDocument.load(synthetic100);
+      const doc2 = await PDFDocument.load(synthetic100);
+      const merged = await PDFDocument.create();
+
+      const pages1 = await merged.copyPages(doc1, doc1.getPageIndices());
+
+      for (const page of pages1) {
+        merged.addPage(page);
+      }
+
+      const pages2 = await merged.copyPages(doc2, doc2.getPageIndices());
+
+      for (const page of pages2) {
+        merged.addPage(page);
+      }
+
+      await merged.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
diff --git a/benchmarks/copying.bench.ts b/benchmarks/copying.bench.ts
new file mode 100644
index 0000000..034b6c2
--- /dev/null
+++ b/benchmarks/copying.bench.ts
@@ -0,0 +1,94 @@
+/**
+ * PDF page-copying and merging benchmarks.
+ *
+ * Tests the performance of copying pages between documents and merging
+ * multiple PDFs. These operations are closely related to splitting
+ * (issue #26) and represent the other side of the workflow.
+ */
+
+import { bench, describe } from "vitest";
+
+import { PDF } from "../src";
+import { getSynthetic100, loadFixture, mediumPdfPath } from "./fixtures";
+
+// Pre-load fixtures
+const mediumPdf = await loadFixture(mediumPdfPath);
+const synthetic100 = await getSynthetic100();
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Page copying
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Copy pages between documents", () => {
+  bench("copy 1 page", async () => {
+    const source = await PDF.load(mediumPdf);
+    const dest = PDF.create();
+    await dest.copyPagesFrom(source, [0]);
+    await dest.save();
+  });
+
+  bench("copy 10 pages from 100-page PDF", async () => {
+    const source = await PDF.load(synthetic100);
+    const dest = PDF.create();
+    const indices = Array.from({ length: 10 }, (_, i) => i);
+    await dest.copyPagesFrom(source, indices);
+    await dest.save();
+  });
+
+  bench(
+    "copy all 100 pages",
+    async () => {
+      const source = await PDF.load(synthetic100);
+      const dest = PDF.create();
+      const indices = Array.from({ length: 100 }, (_, i) => i);
+      await dest.copyPagesFrom(source, indices);
+      await dest.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Self-copy (page duplication)
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Duplicate pages within same document", () => {
+  bench("duplicate page 0", async () => {
+    const pdf = await PDF.load(mediumPdf);
+    await pdf.copyPagesFrom(pdf, [0]);
+    await pdf.save();
+  });
+
+  bench("duplicate all pages (double the document)", async () => {
+    const pdf = await PDF.load(mediumPdf);
+    const indices = Array.from({ length: pdf.getPageCount() }, (_, i) => i);
+    await pdf.copyPagesFrom(pdf, indices);
+    await pdf.save();
+  });
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Merging
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Merge PDFs", () => {
+  bench("merge 2 small PDFs", async () => {
+    const merged = await PDF.merge([mediumPdf, mediumPdf]);
+    await merged.save();
+  });
+
+  bench("merge 10 small PDFs", async () => {
+    const sources = Array.from({ length: 10 }, () => mediumPdf);
+    const merged = await PDF.merge(sources);
+    await merged.save();
+  });
+
+  bench(
+    "merge 2 x 100-page PDFs",
+    async () => {
+      const merged = await PDF.merge([synthetic100, synthetic100]);
+      await merged.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
diff --git a/benchmarks/fixtures.ts b/benchmarks/fixtures.ts
index b18e21c..4cf11f2 100644
--- a/benchmarks/fixtures.ts
+++ b/benchmarks/fixtures.ts
@@ -1,12 +1,16 @@
 /**
  * Benchmark fixture helpers.
  *
- * Provides utilities for loading PDF fixtures for benchmarks.
+ * Provides utilities for loading PDF fixtures for benchmarks,
+ * including synthetic large PDFs built by copying pages from
+ * existing fixtures.
  */
 
 import { existsSync, mkdirSync, writeFileSync } from "node:fs";
 import { readFile } from "node:fs/promises";
 
+import { PDF } from "../src";
+
 // Heavy PDF - downloaded on first run (~10MB)
 const HEAVY_PDF_PATH = "fixtures/benchmarks/cc-journalists-guide.pdf";
 const HEAVY_PDF_URL =
@@ -15,6 +19,11 @@ const HEAVY_PDF_URL =
 // Fallback large PDF - use existing fixture from pdfbox malformed tests (2MB)
 const LARGE_PDF_FALLBACK = "fixtures/malformed/pdfbox/PDFBOX-3947.pdf";
 
+// Synthetic PDFs - generated on first run, cached locally
+const SYNTHETIC_DIR = "fixtures/benchmarks";
+const SYNTHETIC_100_PATH = `${SYNTHETIC_DIR}/synthetic-100p.pdf`;
+const SYNTHETIC_2000_PATH = `${SYNTHETIC_DIR}/synthetic-2000p.pdf`;
+
 /**
  * Load a fixture file as bytes.
  */
@@ -65,6 +74,90 @@ export async function getLargePdf(): Promise<Uint8Array> {
   return loadFixture(LARGE_PDF_FALLBACK);
 }
 
+/**
+ * Build a synthetic PDF with the given number of pages by copying
+ * pages from sample.pdf. Each page gets unique text to simulate
+ * real-world content variation.
+ */
+async function buildSyntheticPdf(pageCount: number): Promise<Uint8Array> {
+  const sourceBytes = await loadFixture(mediumPdfPath);
+  const source = await PDF.load(sourceBytes);
+  const sourcePageCount = source.getPageCount();
+
+  // Start by copying the source pages
+  const pdf = await PDF.load(sourceBytes);
+
+  // Copy pages from source repeatedly until we reach the target count
+  const pagesNeeded = pageCount - sourcePageCount;
+
+  if (pagesNeeded > 0) {
+    // Build an array of source page indices to copy in bulk
+    const indices: number[] = [];
+
+    for (let i = 0; i < pagesNeeded; i++) {
+      indices.push(i % sourcePageCount);
+    }
+
+    await pdf.copyPagesFrom(source, indices);
+  }
+
+  // Add unique text to each page so content varies
+  for (let i = 0; i < pdf.getPageCount(); i++) {
+    const page = pdf.getPage(i);
+
+    if (page) {
+      page.drawText(`Page ${i + 1} of ${pageCount}`, {
+        x: 50,
+        y: 20,
+        font: "Helvetica",
+        size: 8,
+      });
+    }
+  }
+
+  return pdf.save();
+}
+
+/**
+ * Get or create a synthetic PDF cached to disk.
+ */
+async function getOrCreateSynthetic(path: string, pageCount: number): Promise<Uint8Array> {
+  if (existsSync(path)) {
+    return loadFixture(path);
+  }
+
+  console.log(`Building synthetic ${pageCount}-page PDF...`);
+  const start = performance.now();
+
+  const bytes = await buildSyntheticPdf(pageCount);
+
+  mkdirSync(SYNTHETIC_DIR, { recursive: true });
+  writeFileSync(path, bytes);
+
+  const elapsed = ((performance.now() - start) / 1000).toFixed(1);
+  const size = (bytes.length / 1024 / 1024).toFixed(1);
+
+  console.log(`Cached ${pageCount}-page PDF to ${path} (${size}MB) in ${elapsed}s`);
+
+  return bytes;
+}
+
+/**
+ * Get a synthetic 100-page PDF.
+ * Built by copying pages from sample.pdf. Cached on disk after first build.
+ */
+export async function getSynthetic100(): Promise<Uint8Array> {
+  return getOrCreateSynthetic(SYNTHETIC_100_PATH, 100);
+}
+
+/**
+ * Get a synthetic 2000-page PDF.
+ * Built by copying pages from sample.pdf. Cached on disk after first build.
+ */
+export async function getSynthetic2000(): Promise<Uint8Array> {
+  return getOrCreateSynthetic(SYNTHETIC_2000_PATH, 2000);
+}
+
 // Pre-load common fixtures
 export const smallPdfPath = "fixtures/basic/rot0.pdf";
 export const mediumPdfPath = "fixtures/basic/sample.pdf";
diff --git a/benchmarks/splitting.bench.ts b/benchmarks/splitting.bench.ts
new file mode 100644
index 0000000..23f5708
--- /dev/null
+++ b/benchmarks/splitting.bench.ts
@@ -0,0 +1,119 @@
+/**
+ * PDF page-splitting benchmarks.
+ *
+ * Tests the performance of splitting a PDF into individual single-page PDFs.
+ * This is the primary benchmark requested in issue #26 for users who need
+ * to split 2000+ page documents at high throughput.
+ *
+ * Scenarios:
+ * - Extract single page (baseline)
+ * - Split 100-page PDF into individual pages
+ * - Split 2000-page PDF into individual pages
+ */
+
+import { bench, describe } from "vitest";
+
+import { PDF } from "../src";
+import { getSynthetic100, getSynthetic2000, loadFixture, mediumPdfPath } from "./fixtures";
+
+// Pre-load fixtures outside benchmarks to isolate I/O from measurements
+const mediumPdf = await loadFixture(mediumPdfPath);
+const synthetic100 = await getSynthetic100();
+const synthetic2000 = await getSynthetic2000();
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Single page extraction (baseline)
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Extract single page", () => {
+  bench("extractPages (1 page from small PDF)", async () => {
+    const pdf = await PDF.load(mediumPdf);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+
+  bench("extractPages (1 page from 100-page PDF)", async () => {
+    const pdf = await PDF.load(synthetic100);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+
+  bench("extractPages (1 page from 2000-page PDF)", async () => {
+    const pdf = await PDF.load(synthetic2000);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Full split: every page into its own PDF
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Split into single-page PDFs", () => {
+  bench(
+    `split 100-page PDF (${(synthetic100.length / 1024 / 1024).toFixed(1)}MB)`,
+    async () => {
+      const pdf = await PDF.load(synthetic100);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    `split 2000-page PDF (${(synthetic2000.length / 1024 / 1024).toFixed(1)}MB)`,
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 0, iterations: 1, time: 0 },
+  );
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Batch extraction: extract ranges of pages
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Batch page extraction", () => {
+  bench(
+    "extract first 10 pages from 2000-page PDF",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const indices = Array.from({ length: 10 }, (_, i) => i);
+      const extracted = await pdf.extractPages(indices);
+      await extracted.save();
+    },
+    { warmupIterations: 1, iterations: 5 },
+  );
+
+  bench(
+    "extract first 100 pages from 2000-page PDF",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const indices = Array.from({ length: 100 }, (_, i) => i);
+      const extracted = await pdf.extractPages(indices);
+      await extracted.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    "extract every 10th page from 2000-page PDF (200 pages)",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const indices = Array.from({ length: 200 }, (_, i) => i * 10);
+      const extracted = await pdf.extractPages(indices);
+      await extracted.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
diff --git a/package.json b/package.json
index b511e67..230f00b 100644
--- a/package.json
+++ b/package.json
@@ -46,6 +46,7 @@
   },
   "scripts": {
     "bench": "vitest bench",
+    "bench:report": "bun run scripts/bench-report.ts",
     "build": "tsdown",
     "docs:build": "bun run --cwd apps/docs build",
     "docs:dev": "bun run --cwd apps/docs dev",
diff --git a/reports/benchmarks.md b/reports/benchmarks.md
new file mode 100644
index 0000000..b0b0010
--- /dev/null
+++ b/reports/benchmarks.md
@@ -0,0 +1,235 @@
+# Benchmark Report
+
+> Generated on 2026-02-16 at 12:50:00 UTC
+>
+> System: darwin | Apple M4 Pro (12 cores) | 24GB RAM | Bun 1.3.5
+
+---
+
+## Contents
+
+- [Comparison](#comparison)
+- [Copying](#copying)
+- [Drawing](#drawing)
+- [Forms](#forms)
+- [Loading](#loading)
+- [Saving](#saving)
+- [Splitting](#splitting)
+
+## Comparison
+
+### Load PDF
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |   895.0 |  1.12ms |  1.59ms | ±1.07% |     448 |
+| pdf-lib   |    36.7 | 27.21ms | 29.03ms | ±1.90% |      19 |
+
+- **libpdf** is 24.35x faster than pdf-lib
+
+### Create blank PDF
+
+| Benchmark | ops/sec | Mean |   p99 |    RME | Samples |
+| :-------- | ------: | ---: | ----: | -----: | ------: |
+| libpdf    |   38.6K | 26us |  45us | ±0.84% |  19,283 |
+| pdf-lib   |   10.3K | 97us | 461us | ±1.82% |   5,172 |
+
+- **libpdf** is 3.73x faster than pdf-lib
+
+### Add 10 pages
+
+| Benchmark | ops/sec |  Mean |   p99 |    RME | Samples |
+| :-------- | ------: | ----: | ----: | -----: | ------: |
+| libpdf    |   19.1K |  52us |  87us | ±0.91% |   9,562 |
+| pdf-lib   |    6.3K | 158us | 770us | ±2.55% |   3,173 |
+
+- **libpdf** is 3.01x faster than pdf-lib
+
+### Draw 50 rectangles
+
+| Benchmark | ops/sec |   Mean |    p99 |    RME | Samples |
+| :-------- | ------: | -----: | -----: | -----: | ------: |
+| pdf-lib   |    2.2K |  458us | 1.80ms | ±3.80% |   1,093 |
+| libpdf    |   627.4 | 1.59ms | 2.30ms | ±1.44% |     314 |
+
+- **pdf-lib** is 3.48x faster than libpdf
+
+### Load and save PDF
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |   909.3 |  1.10ms |  1.48ms | ±0.88% |     456 |
+| pdf-lib   |    22.0 | 45.45ms | 58.21ms | ±6.77% |      11 |
+
+- **libpdf** is 41.33x faster than pdf-lib
+
+### Load, modify, and save PDF
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |    37.9 | 26.38ms | 33.50ms | ±5.76% |      20 |
+| pdf-lib   |    23.1 | 43.25ms | 44.89ms | ±1.52% |      12 |
+
+- **libpdf** is 1.64x faster than pdf-lib
+
+### Extract single page from 100-page PDF
+
+| Benchmark | ops/sec |   Mean |    p99 |    RME | Samples |
+| :-------- | ------: | -----: | -----: | -----: | ------: |
+| libpdf    |   503.4 | 1.99ms | 3.10ms | ±1.55% |     252 |
+| pdf-lib   |   155.9 | 6.41ms | 7.44ms | ±1.53% |      79 |
+
+- **libpdf** is 3.23x faster than pdf-lib
+
+### Split 100-page PDF into single-page PDFs
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |    35.3 | 28.29ms | 29.79ms | ±1.35% |      18 |
+| pdf-lib   |    35.0 | 28.58ms | 32.30ms | ±3.07% |      18 |
+
+- **libpdf** is 1.01x faster than pdf-lib
+
+### Copy 10 pages between documents
+
+| Benchmark | ops/sec |   Mean |     p99 |    RME | Samples |
+| :-------- | ------: | -----: | ------: | -----: | ------: |
+| libpdf    |   334.0 | 2.99ms |  3.55ms | ±1.12% |     168 |
+| pdf-lib   |   103.7 | 9.64ms | 14.95ms | ±3.70% |      52 |
+
+- **libpdf** is 3.22x faster than pdf-lib
+
+### Merge 2 x 100-page PDFs
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |    47.1 | 21.23ms | 24.31ms | ±2.08% |      24 |
+| pdf-lib   |    22.9 | 43.64ms | 47.48ms | ±2.23% |      12 |
+
+- **libpdf** is 2.06x faster than pdf-lib
+
+## Copying
+
+### Copy pages between documents
+
+| Benchmark                       | ops/sec |    Mean |     p99 |    RME | Samples |
+| :------------------------------ | ------: | ------: | ------: | -----: | ------: |
+| copy 1 page                     |    2.3K |   429us |   735us | ±1.26% |   1,166 |
+| copy 10 pages from 100-page PDF |   344.1 |  2.91ms |  3.57ms | ±1.12% |     173 |
+| copy all 100 pages              |    92.3 | 10.84ms | 13.86ms | ±1.96% |      47 |
+
+- **copy 1 page** is 6.78x faster than copy 10 pages from 100-page PDF
+- **copy 1 page** is 25.27x faster than copy all 100 pages
+
+### Duplicate pages within same document
+
+| Benchmark                                 | ops/sec |  Mean |   p99 |    RME | Samples |
+| :---------------------------------------- | ------: | ----: | ----: | -----: | ------: |
+| duplicate all pages (double the document) |    2.2K | 461us | 798us | ±0.89% |   1,086 |
+| duplicate page 0                          |    2.2K | 464us | 758us | ±0.77% |   1,078 |
+
+- **duplicate all pages (double the document)** is 1.01x faster than duplicate page 0
+
+### Merge PDFs
+
+| Benchmark               | ops/sec |    Mean |     p99 |    RME | Samples |
+| :---------------------- | ------: | ------: | ------: | -----: | ------: |
+| merge 2 small PDFs      |    1.4K |   712us |  1.12ms | ±1.00% |     702 |
+| merge 10 small PDFs     |   254.5 |  3.93ms |  5.73ms | ±2.08% |     128 |
+| merge 2 x 100-page PDFs |    48.6 | 20.58ms | 26.32ms | ±3.33% |      25 |
+
+- **merge 2 small PDFs** is 5.51x faster than merge 10 small PDFs
+- **merge 2 small PDFs** is 28.89x faster than merge 2 x 100-page PDFs
+
+## Drawing
+
+| Benchmark                           | ops/sec |   Mean |    p99 |    RME | Samples |
+| :---------------------------------- | ------: | -----: | -----: | -----: | ------: |
+| draw 100 lines                      |   399.8 | 2.50ms | 2.74ms | ±0.51% |     200 |
+| draw 100 rectangles                 |   360.3 | 2.78ms | 3.55ms | ±1.22% |     181 |
+| draw 100 circles                    |   279.7 | 3.58ms | 4.42ms | ±1.33% |     140 |
+| draw 100 text lines (standard font) |   259.5 | 3.85ms | 4.29ms | ±0.61% |     130 |
+| create 10 pages with mixed content  |   193.0 | 5.18ms | 6.40ms | ±1.35% |      97 |
+
+- **draw 100 lines** is 1.11x faster than draw 100 rectangles
+- **draw 100 lines** is 1.43x faster than draw 100 circles
+- **draw 100 lines** is 1.54x faster than draw 100 text lines (standard font)
+- **draw 100 lines** is 2.07x faster than create 10 pages with mixed content
+
+## Forms
+
+| Benchmark         | ops/sec |   Mean |    p99 |    RME | Samples |
+| :---------------- | ------: | -----: | -----: | -----: | ------: |
+| read field values |   702.6 | 1.42ms | 1.85ms | ±0.97% |     352 |
+| get form fields   |   677.3 | 1.48ms | 2.36ms | ±1.45% |     339 |
+| flatten form      |   198.8 | 5.03ms | 5.84ms | ±1.28% |     100 |
+| fill text fields  |   155.1 | 6.45ms | 7.36ms | ±1.25% |      78 |
+
+- **read field values** is 1.04x faster than get form fields
+- **read field values** is 3.53x faster than flatten form
+- **read field values** is 4.53x faster than fill text fields
+
+## Loading
+
+| Benchmark              | ops/sec |   Mean |    p99 |    RME | Samples |
+| :--------------------- | ------: | -----: | -----: | -----: | ------: |
+| load small PDF (888B)  |   38.7K |   26us |   37us | ±1.10% |  19,336 |
+| load medium PDF (19KB) |   23.8K |   42us |   54us | ±0.84% |  11,904 |
+| load form PDF (116KB)  |    1.6K |  639us | 1.07ms | ±0.96% |     782 |
+| load heavy PDF (9.9MB) |   909.7 | 1.10ms | 1.46ms | ±0.81% |     455 |
+
+- **load small PDF (888B)** is 1.62x faster than load medium PDF (19KB)
+- **load small PDF (888B)** is 24.73x faster than load form PDF (116KB)
+- **load small PDF (888B)** is 42.51x faster than load heavy PDF (9.9MB)
+
+## Saving
+
+| Benchmark                          | ops/sec |   Mean |    p99 |    RME | Samples |
+| :--------------------------------- | ------: | -----: | -----: | -----: | ------: |
+| save unmodified (19KB)             |   22.2K |   45us |   63us | ±0.78% |  11,103 |
+| incremental save (19KB)            |    6.9K |  144us |  373us | ±1.14% |   3,461 |
+| save with modifications (19KB)     |    2.4K |  422us |  799us | ±0.97% |   1,185 |
+| save heavy PDF (9.9MB)             |   850.9 | 1.18ms | 1.58ms | ±0.91% |     426 |
+| incremental save heavy PDF (9.9MB) |   494.7 | 2.02ms | 2.37ms | ±0.74% |     248 |
+
+- **save unmodified (19KB)** is 3.21x faster than incremental save (19KB)
+- **save unmodified (19KB)** is 9.37x faster than save with modifications (19KB)
+- **save unmodified (19KB)** is 26.10x faster than save heavy PDF (9.9MB)
+- **save unmodified (19KB)** is 44.89x faster than incremental save heavy PDF (9.9MB)
+
+## Splitting
+
+### Extract single page
+
+| Benchmark                                | ops/sec |    Mean |     p99 |    RME | Samples |
+| :--------------------------------------- | ------: | ------: | ------: | -----: | ------: |
+| extractPages (1 page from small PDF)     |    2.2K |   452us |   931us | ±1.63% |   1,106 |
+| extractPages (1 page from 100-page PDF)  |   536.3 |  1.86ms |  3.04ms | ±1.54% |     269 |
+| extractPages (1 page from 2000-page PDF) |    24.7 | 40.43ms | 42.46ms | ±2.46% |      13 |
+
+- **extractPages (1 page from small PDF)** is 4.12x faster than extractPages (1 page from 100-page PDF)
+- **extractPages (1 page from small PDF)** is 89.36x faster than extractPages (1 page from 2000-page PDF)
+
+### Split into single-page PDFs
+
+| Benchmark                   | ops/sec |     Mean |      p99 |    RME | Samples |
+| :-------------------------- | ------: | -------: | -------: | -----: | ------: |
+| split 100-page PDF (0.1MB)  |    32.2 |  31.02ms |  35.44ms | ±2.72% |      17 |
+| split 2000-page PDF (0.9MB) |     1.8 | 550.66ms | 550.66ms | ±0.00% |       1 |
+
+- **split 100-page PDF (0.1MB)** is 17.75x faster than split 2000-page PDF (0.9MB)
+
+### Batch page extraction
+
+| Benchmark                                              | ops/sec |    Mean |     p99 |    RME | Samples |
+| :----------------------------------------------------- | ------: | ------: | ------: | -----: | ------: |
+| extract first 10 pages from 2000-page PDF              |    23.3 | 42.90ms | 46.80ms | ±3.59% |      12 |
+| extract first 100 pages from 2000-page PDF             |    20.1 | 49.79ms | 52.72ms | ±2.98% |      11 |
+| extract every 10th page from 2000-page PDF (200 pages) |    18.6 | 53.74ms | 59.30ms | ±3.19% |      10 |
+
+- **extract first 10 pages from 2000-page PDF** is 1.16x faster than extract first 100 pages from 2000-page PDF
+- **extract first 10 pages from 2000-page PDF** is 1.25x faster than extract every 10th page from 2000-page PDF (200 pages)
+
+---
+
+_Results are machine-dependent. Use for relative comparison only._
diff --git a/scripts/bench-report.ts b/scripts/bench-report.ts
new file mode 100644
index 0000000..d16617d
--- /dev/null
+++ b/scripts/bench-report.ts
@@ -0,0 +1,262 @@
+/**
+ * Benchmark report generator.
+ *
+ * Runs vitest bench with JSON output, then transforms the results
+ * into a markdown report saved to reports/benchmarks.md.
+ *
+ * Usage:
+ *   bun run scripts/bench-report.ts
+ *   bun run scripts/bench-report.ts --json-only    # Just dump JSON, skip markdown
+ *   bun run scripts/bench-report.ts --from-json results.json  # Generate from existing JSON
+ */
+
+import { execSync } from "node:child_process";
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { cpus, platform, totalmem } from "node:os";
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Types for vitest bench JSON output
+// ─────────────────────────────────────────────────────────────────────────────
+
+interface BenchmarkResult {
+  name: string;
+  rank: number;
+  rme: number;
+  hz: number;
+  min: number;
+  max: number;
+  mean: number;
+  p75: number;
+  p99: number;
+  p995: number;
+  p999: number;
+  sampleCount: number;
+  median: number;
+}
+
+interface BenchmarkGroup {
+  fullName: string;
+  benchmarks: BenchmarkResult[];
+}
+
+interface BenchmarkFile {
+  filepath: string;
+  groups: BenchmarkGroup[];
+}
+
+interface BenchmarkOutput {
+  files: BenchmarkFile[];
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Helpers
+// ─────────────────────────────────────────────────────────────────────────────
+
+function formatHz(hz: number): string {
+  if (hz >= 1000) {
+    return `${(hz / 1000).toFixed(1)}K`;
+  }
+
+  if (hz >= 1) {
+    return hz.toFixed(1);
+  }
+
+  return hz.toFixed(3);
+}
+
+function formatTime(ms: number): string {
+  if (ms >= 1000) {
+    return `${(ms / 1000).toFixed(2)}s`;
+  }
+
+  if (ms >= 1) {
+    return `${ms.toFixed(2)}ms`;
+  }
+
+  return `${(ms * 1000).toFixed(0)}us`;
+}
+
+function formatRme(rme: number): string {
+  return `\u00b1${rme.toFixed(2)}%`;
+}
+
+function getSystemInfo(): string {
+  const cpu = cpus()[0];
+  const cpuModel = cpu?.model ?? "Unknown CPU";
+  const cpuCount = cpus().length;
+  const mem = (totalmem() / 1024 / 1024 / 1024).toFixed(0);
+  const os = platform();
+  const runtime = `Bun ${process.versions.bun ?? "unknown"}`;
+
+  return `${os} | ${cpuModel} (${cpuCount} cores) | ${mem}GB RAM | ${runtime}`;
+}
+
+/**
+ * Extract a short file label from a benchmark filepath.
+ * e.g. "/Users/.../benchmarks/loading.bench.ts" -> "Loading"
+ */
+function fileLabel(filepath: string): string {
+  const match = filepath.match(/([^/]+)\.bench\.ts$/);
+
+  if (!match) {
+    return filepath;
+  }
+
+  const name = match[1];
+
+  return name.charAt(0).toUpperCase() + name.slice(1);
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Markdown generation
+// ─────────────────────────────────────────────────────────────────────────────
+
+function generateMarkdown(data: BenchmarkOutput): string {
+  const lines: string[] = [];
+  const now = new Date();
+  const dateStr = now.toISOString().split("T")[0];
+  const timeStr = now.toISOString().split("T")[1].split(".")[0];
+
+  lines.push("# Benchmark Report");
+  lines.push("");
+  lines.push(`> Generated on ${dateStr} at ${timeStr} UTC`);
+  lines.push(`>`);
+  lines.push(`> System: ${getSystemInfo()}`);
+  lines.push("");
+  lines.push("---");
+  lines.push("");
+
+  // Table of contents
+  lines.push("## Contents");
+  lines.push("");
+
+  for (const file of data.files) {
+    const label = fileLabel(file.filepath);
+    const anchor = label.toLowerCase().replace(/\s+/g, "-");
+    lines.push(`- [${label}](#${anchor})`);
+  }
+
+  lines.push("");
+
+  // Each file becomes a section
+  for (const file of data.files) {
+    const label = fileLabel(file.filepath);
+    lines.push(`## ${label}`);
+    lines.push("");
+
+    for (const group of file.groups) {
+      // If the group name differs from the file-level name, add a subheading
+      const groupName = group.fullName.replace(/^benchmarks\/[^>]+> /, "").trim();
+      const isTopLevel = group.benchmarks.length > 0 && !groupName.includes(" > ");
+
+      // Check if this group has a describe() wrapper (indicated by " > " in fullName)
+      const describeName = group.fullName.includes(" > ")
+        ? group.fullName.split(" > ").slice(1).join(" > ")
+        : null;
+
+      if (describeName) {
+        lines.push(`### ${describeName}`);
+        lines.push("");
+      }
+
+      // Build the results table
+      lines.push("| Benchmark | ops/sec | Mean | p99 | RME | Samples |");
+      lines.push("|:---|---:|---:|---:|---:|---:|");
+
+      // Sort by rank
+      const sorted = [...group.benchmarks].sort((a, b) => a.rank - b.rank);
+
+      for (const bench of sorted) {
+        const name = bench.name;
+        const hz = formatHz(bench.hz);
+        const mean = formatTime(bench.mean);
+        const p99 = formatTime(bench.p99);
+        const rme = formatRme(bench.rme);
+        const samples = bench.sampleCount.toLocaleString();
+
+        lines.push(`| ${name} | ${hz} | ${mean} | ${p99} | ${rme} | ${samples} |`);
+      }
+
+      lines.push("");
+
+      // Add comparison summary for groups with multiple benchmarks
+      if (sorted.length >= 2) {
+        const fastest = sorted[0];
+        const rest = sorted.slice(1);
+
+        for (const slower of rest) {
+          const ratio = (fastest.hz / slower.hz).toFixed(2);
+          lines.push(`- **${fastest.name}** is ${ratio}x faster than ${slower.name}`);
+        }
+
+        lines.push("");
+      }
+    }
+  }
+
+  // Footer
+  lines.push("---");
+  lines.push("");
+  lines.push("*Results are machine-dependent. Use for relative comparison only.*");
+  lines.push("");
+
+  return lines.join("\n");
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Main
+// ─────────────────────────────────────────────────────────────────────────────
+
+const args = process.argv.slice(2);
+const jsonOnly = args.includes("--json-only");
+const fromJsonIdx = args.indexOf("--from-json");
+
+const jsonPath = "reports/bench-results.json";
+const mdPath = "reports/benchmarks.md";
+
+mkdirSync("reports", { recursive: true });
+
+let data: BenchmarkOutput;
+
+if (fromJsonIdx !== -1 && args[fromJsonIdx + 1]) {
+  // Generate markdown from an existing JSON file
+  const inputPath = args[fromJsonIdx + 1];
+
+  if (!existsSync(inputPath)) {
+    console.error(`File not found: ${inputPath}`);
+    process.exit(1);
+  }
+
+  data = JSON.parse(readFileSync(inputPath, "utf-8")) as BenchmarkOutput;
+  console.log(`Loaded benchmark results from ${inputPath}`);
+} else {
+  // Run benchmarks and capture JSON
+  console.log("Running benchmarks...\n");
+
+  try {
+    execSync(`bun run bench -- --outputJson ${jsonPath}`, {
+      stdio: "inherit",
+      timeout: 600_000, // 10 minute timeout
+    });
+  } catch (error) {
+    console.error("Benchmark run failed");
+    process.exit(1);
+  }
+
+  if (!existsSync(jsonPath)) {
+    console.error(`Expected JSON output at ${jsonPath} but file not found`);
+    process.exit(1);
+  }
+
+  data = JSON.parse(readFileSync(jsonPath, "utf-8")) as BenchmarkOutput;
+  console.log(`\nBenchmark JSON saved to ${jsonPath}`);
+}
+
+if (jsonOnly) {
+  process.exit(0);
+}
+
+// Generate and write markdown report
+const md = generateMarkdown(data);
+writeFileSync(mdPath, md);
+console.log(`Benchmark report saved to ${mdPath}`);

From 2fe6842ba2b6d3724c4708a6a9b40e0a262e2fac Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 10:51:02 +1100
Subject: [PATCH 2/9] perf(copier): make ObjectCopier fully synchronous
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

ObjectCopier does zero I/O — every method was async but never awaited
anything asynchronous. Removing async/await eliminates microtask
scheduling overhead on every recursive call in the deep-copy graph walk.

Benchmarks show ~15% improvement on full-split workloads:
- 100-page split: 31.6ms → 27.3ms (1.16x)
- 2000-page split: 582.5ms → 506.6ms (1.15x)
---
 reports/baselines/baseline.json             | 224 ++++++++++++++++++++
 reports/baselines/comparison-sync-copier.md |  28 +++
 reports/baselines/sync-copier.json          | 224 ++++++++++++++++++++
 src/api/pdf.ts                              |   6 +-
 src/document/object-copier.test.ts          |  50 ++---
 src/document/object-copier.ts               |  46 ++--
 6 files changed, 529 insertions(+), 49 deletions(-)
 create mode 100644 reports/baselines/baseline.json
 create mode 100644 reports/baselines/comparison-sync-copier.md
 create mode 100644 reports/baselines/sync-copier.json

diff --git a/reports/baselines/baseline.json b/reports/baselines/baseline.json
new file mode 100644
index 0000000..bae2d4c
--- /dev/null
+++ b/reports/baselines/baseline.json
@@ -0,0 +1,224 @@
+{
+  "files": [
+    {
+      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
+      "groups": [
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
+          "benchmarks": [
+            {
+              "id": "-1072438084_0_0",
+              "name": "extractPages (1 page from small PDF)",
+              "rank": 1,
+              "rme": 1.598586232947195,
+              "samples": [],
+              "totalTime": 500.1328439999987,
+              "min": 0.3573750000000473,
+              "max": 2.4449999999999363,
+              "hz": 2215.411391778147,
+              "period": 0.45138343321299523,
+              "mean": 0.45138343321299523,
+              "variance": 0.015017269876550877,
+              "sd": 0.12254497083336745,
+              "sem": 0.0036815068475241517,
+              "df": 1107,
+              "critical": 1.96,
+              "moe": 0.007215753421147338,
+              "p75": 0.4394999999999527,
+              "p99": 0.8506250000000364,
+              "p995": 0.9532910000000356,
+              "p999": 1.7487909999999829,
+              "sampleCount": 1108,
+              "median": 0.4120829999999387
+            },
+            {
+              "id": "-1072438084_0_1",
+              "name": "extractPages (1 page from 100-page PDF)",
+              "rank": 2,
+              "rme": 1.14326877952979,
+              "samples": [],
+              "totalTime": 500.9256660000017,
+              "min": 1.502375000000029,
+              "max": 2.347500000000082,
+              "hz": 560.9614740722809,
+              "period": 1.782653615658369,
+              "mean": 1.782653615658369,
+              "variance": 0.03038259004746126,
+              "sd": 0.17430602412843127,
+              "sem": 0.010398225630092402,
+              "df": 280,
+              "critical": 1.96,
+              "moe": 0.02038052223498111,
+              "p75": 1.8196669999999813,
+              "p99": 2.319291000000021,
+              "p995": 2.3412920000000668,
+              "p999": 2.347500000000082,
+              "sampleCount": 281,
+              "median": 1.7196670000000722
+            },
+            {
+              "id": "-1072438084_0_2",
+              "name": "extractPages (1 page from 2000-page PDF)",
+              "rank": 3,
+              "rme": 2.6124445200935758,
+              "samples": [],
+              "totalTime": 531.2322489999992,
+              "min": 37.85783300000003,
+              "max": 44.21758299999965,
+              "hz": 24.471405914214404,
+              "period": 40.864019153846094,
+              "mean": 40.864019153846094,
+              "variance": 3.120363595900889,
+              "sd": 1.7664550930892324,
+              "sem": 0.4899264933798263,
+              "df": 12,
+              "critical": 2.179,
+              "moe": 1.0675498290746415,
+              "p75": 42.14979199999971,
+              "p99": 44.21758299999965,
+              "p995": 44.21758299999965,
+              "p999": 44.21758299999965,
+              "sampleCount": 13,
+              "median": 40.22212499999978
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
+          "benchmarks": [
+            {
+              "id": "-1072438084_1_0",
+              "name": "split 100-page PDF (0.1MB)",
+              "rank": 1,
+              "rme": 1.8856591767197517,
+              "samples": [],
+              "totalTime": 506.2218300000013,
+              "min": 30.098917000000256,
+              "max": 34.55699999999979,
+              "hz": 31.60669700870063,
+              "period": 31.638864375000082,
+              "mean": 31.638864375000082,
+              "variance": 1.2540684871776795,
+              "sd": 1.119851993424881,
+              "sem": 0.27996299835622024,
+              "df": 15,
+              "critical": 2.131,
+              "moe": 0.5966011494971053,
+              "p75": 32.05224999999973,
+              "p99": 34.55699999999979,
+              "p995": 34.55699999999979,
+              "p999": 34.55699999999979,
+              "sampleCount": 16,
+              "median": 31.37264550000009
+            },
+            {
+              "id": "-1072438084_1_1",
+              "name": "split 2000-page PDF (0.9MB)",
+              "rank": 2,
+              "rme": 0,
+              "samples": [],
+              "totalTime": 582.5043749999995,
+              "min": 582.5043749999995,
+              "max": 582.5043749999995,
+              "hz": 1.716725303565318,
+              "period": 582.5043749999995,
+              "mean": 582.5043749999995,
+              "variance": 0,
+              "sd": 0,
+              "sem": 0,
+              "df": 0,
+              "critical": 12.71,
+              "moe": 0,
+              "p75": 582.5043749999995,
+              "p99": 582.5043749999995,
+              "p995": 582.5043749999995,
+              "p999": 582.5043749999995,
+              "sampleCount": 1,
+              "median": 582.5043749999995
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
+          "benchmarks": [
+            {
+              "id": "-1072438084_2_0",
+              "name": "extract first 10 pages from 2000-page PDF",
+              "rank": 1,
+              "rme": 3.094676818514817,
+              "samples": [],
+              "totalTime": 514.4472479999995,
+              "min": 39.471832999999606,
+              "max": 46.71299999999974,
+              "hz": 23.326006790107296,
+              "period": 42.87060399999996,
+              "mean": 42.87060399999996,
+              "variance": 4.360044973374979,
+              "sd": 2.0880720709245115,
+              "sem": 0.602774486117803,
+              "df": 11,
+              "critical": 2.201,
+              "moe": 1.3267066439452846,
+              "p75": 43.80012500000066,
+              "p99": 46.71299999999974,
+              "p995": 46.71299999999974,
+              "p999": 46.71299999999974,
+              "sampleCount": 12,
+              "median": 42.56633349999993
+            },
+            {
+              "id": "-1072438084_2_1",
+              "name": "extract first 100 pages from 2000-page PDF",
+              "rank": 2,
+              "rme": 3.2071455847378436,
+              "samples": [],
+              "totalTime": 509.16504200000236,
+              "min": 48.13025000000016,
+              "max": 54.86754199999996,
+              "hz": 19.639997201535987,
+              "period": 50.91650420000023,
+              "mean": 50.91650420000023,
+              "variance": 5.211578755405502,
+              "sd": 2.282888248558282,
+              "sem": 0.7219126509076774,
+              "df": 9,
+              "critical": 2.262,
+              "moe": 1.6329664163531663,
+              "p75": 52.17775000000074,
+              "p99": 54.86754199999996,
+              "p995": 54.86754199999996,
+              "p999": 54.86754199999996,
+              "sampleCount": 10,
+              "median": 51.348021000000244
+            },
+            {
+              "id": "-1072438084_2_2",
+              "name": "extract every 10th page from 2000-page PDF (200 pages)",
+              "rank": 3,
+              "rme": 4.394613298511672,
+              "samples": [],
+              "totalTime": 529.0494579999995,
+              "min": 54.37287500000002,
+              "max": 64.11266699999942,
+              "hz": 17.011642038200534,
+              "period": 58.78327311111106,
+              "mean": 58.78327311111106,
+              "variance": 11.294649634264848,
+              "sd": 3.360751349663472,
+              "sem": 1.120250449887824,
+              "df": 8,
+              "critical": 2.306,
+              "moe": 2.5832975374413225,
+              "p75": 60.755959000000075,
+              "p99": 64.11266699999942,
+              "p995": 64.11266699999942,
+              "p999": 64.11266699999942,
+              "sampleCount": 9,
+              "median": 57.71983300000011
+            }
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/reports/baselines/comparison-sync-copier.md b/reports/baselines/comparison-sync-copier.md
new file mode 100644
index 0000000..8752df2
--- /dev/null
+++ b/reports/baselines/comparison-sync-copier.md
@@ -0,0 +1,28 @@
+# Benchmark Comparison: Sync ObjectCopier
+
+## Change
+
+Removed async/await from all ObjectCopier methods. The copier does zero I/O —
+every method was async but never awaited anything asynchronous. Removing the
+async overhead eliminates microtask scheduling on every recursive call in the
+deep-copy graph walk.
+
+## Results
+
+| Benchmark                                              | Baseline (ms) | Sync (ms) | Speedup |
+| :----------------------------------------------------- | ------------: | --------: | ------: |
+| extractPages (1 page from small PDF)                   |          0.45 |      0.40 |   1.14x |
+| extractPages (1 page from 100-page PDF)                |          1.78 |      1.74 |   1.03x |
+| extractPages (1 page from 2000-page PDF)               |         40.86 |     41.04 |   1.00x |
+| split 100-page PDF (0.1MB)                             |         31.64 |     27.35 |   1.16x |
+| split 2000-page PDF (0.9MB)                            |        582.50 |    506.61 |   1.15x |
+| extract first 10 pages from 2000-page PDF              |         42.87 |     43.21 |   0.99x |
+| extract first 100 pages from 2000-page PDF             |         50.92 |     53.87 |   0.95x |
+| extract every 10th page from 2000-page PDF (200 pages) |         58.78 |     56.65 |   1.04x |
+
+## Key Takeaways
+
+- **100-page split**: 31.6ms -> 27.3ms (1.16x faster)
+- **2000-page split**: 582.5ms -> 506.6ms (1.15x faster)
+- Single-page extraction from small PDFs: ~14% faster (0.45ms -> 0.40ms)
+- Batch extraction noise is within margin of error for single runs
diff --git a/reports/baselines/sync-copier.json b/reports/baselines/sync-copier.json
new file mode 100644
index 0000000..7b6bd2f
--- /dev/null
+++ b/reports/baselines/sync-copier.json
@@ -0,0 +1,224 @@
+{
+  "files": [
+    {
+      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
+      "groups": [
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
+          "benchmarks": [
+            {
+              "id": "-1072438084_0_0",
+              "name": "extractPages (1 page from small PDF)",
+              "rank": 1,
+              "rme": 0.8797965308418093,
+              "samples": [],
+              "totalTime": 500.17104999999685,
+              "min": 0.3400409999999283,
+              "max": 0.8508749999998599,
+              "hz": 2517.1388867868463,
+              "period": 0.39727644956314284,
+              "mean": 0.39727644956314284,
+              "variance": 0.004003720204126598,
+              "sd": 0.0632749571641625,
+              "sem": 0.0017832777658714494,
+              "df": 1258,
+              "critical": 1.96,
+              "moe": 0.0034952244211080407,
+              "p75": 0.3927090000001954,
+              "p99": 0.74350000000004,
+              "p995": 0.7761249999998654,
+              "p999": 0.8387500000000045,
+              "sampleCount": 1259,
+              "median": 0.38120900000012625
+            },
+            {
+              "id": "-1072438084_0_1",
+              "name": "extractPages (1 page from 100-page PDF)",
+              "rank": 2,
+              "rme": 0.9765406302412275,
+              "samples": [],
+              "totalTime": 500.3271310000009,
+              "min": 1.584292000000005,
+              "max": 2.372624999999971,
+              "hz": 575.6233914886368,
+              "period": 1.737246982638892,
+              "mean": 1.737246982638892,
+              "variance": 0.021576654723250537,
+              "sd": 0.14688994085113705,
+              "sem": 0.008655572771994156,
+              "df": 287,
+              "critical": 1.96,
+              "moe": 0.016964922633108545,
+              "p75": 1.756249999999909,
+              "p99": 2.2702919999999267,
+              "p995": 2.3456670000000486,
+              "p999": 2.372624999999971,
+              "sampleCount": 288,
+              "median": 1.6839165000000094
+            },
+            {
+              "id": "-1072438084_0_2",
+              "name": "extractPages (1 page from 2000-page PDF)",
+              "rank": 3,
+              "rme": 2.1980090898940152,
+              "samples": [],
+              "totalTime": 533.5146669999999,
+              "min": 38.59916599999997,
+              "max": 43.44508300000007,
+              "hz": 24.366715301568274,
+              "period": 41.039589769230766,
+              "mean": 41.039589769230766,
+              "variance": 2.227890706742653,
+              "sd": 1.4926120416044664,
+              "sem": 0.4139760961830687,
+              "df": 12,
+              "critical": 2.179,
+              "moe": 0.9020539135829065,
+              "p75": 42.321292000000085,
+              "p99": 43.44508300000007,
+              "p995": 43.44508300000007,
+              "p999": 43.44508300000007,
+              "sampleCount": 13,
+              "median": 40.792207999999846
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
+          "benchmarks": [
+            {
+              "id": "-1072438084_1_0",
+              "name": "split 100-page PDF (0.1MB)",
+              "rank": 1,
+              "rme": 1.9707618954593826,
+              "samples": [],
+              "totalTime": 519.5589599999998,
+              "min": 25.109875000000102,
+              "max": 29.522958000000017,
+              "hz": 36.569478081948596,
+              "period": 27.34520842105262,
+              "mean": 27.34520842105262,
+              "variance": 1.2500641058887851,
+              "sd": 1.11806265740735,
+              "sem": 0.25650116506237763,
+              "df": 18,
+              "critical": 2.101,
+              "moe": 0.5389089477960554,
+              "p75": 28.18224999999984,
+              "p99": 29.522958000000017,
+              "p995": 29.522958000000017,
+              "p999": 29.522958000000017,
+              "sampleCount": 19,
+              "median": 27.39729099999977
+            },
+            {
+              "id": "-1072438084_1_1",
+              "name": "split 2000-page PDF (0.9MB)",
+              "rank": 2,
+              "rme": 0,
+              "samples": [],
+              "totalTime": 506.605583,
+              "min": 506.605583,
+              "max": 506.605583,
+              "hz": 1.9739221863253724,
+              "period": 506.605583,
+              "mean": 506.605583,
+              "variance": 0,
+              "sd": 0,
+              "sem": 0,
+              "df": 0,
+              "critical": 12.71,
+              "moe": 0,
+              "p75": 506.605583,
+              "p99": 506.605583,
+              "p995": 506.605583,
+              "p999": 506.605583,
+              "sampleCount": 1,
+              "median": 506.605583
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
+          "benchmarks": [
+            {
+              "id": "-1072438084_2_0",
+              "name": "extract first 10 pages from 2000-page PDF",
+              "rank": 1,
+              "rme": 2.7205867827906096,
+              "samples": [],
+              "totalTime": 518.4931240000024,
+              "min": 40.28308300000026,
+              "max": 45.694250000000466,
+              "hz": 23.143990623104088,
+              "period": 43.20776033333353,
+              "mean": 43.20776033333353,
+              "variance": 3.422865540761208,
+              "sd": 1.850098792162518,
+              "sem": 0.5340775178412157,
+              "df": 11,
+              "critical": 2.201,
+              "moe": 1.175504616768516,
+              "p75": 44.60529199999928,
+              "p99": 45.694250000000466,
+              "p995": 45.694250000000466,
+              "p999": 45.694250000000466,
+              "sampleCount": 12,
+              "median": 43.13712450000003
+            },
+            {
+              "id": "-1072438084_2_1",
+              "name": "extract first 100 pages from 2000-page PDF",
+              "rank": 2,
+              "rme": 6.90077055615189,
+              "samples": [],
+              "totalTime": 538.6501679999983,
+              "min": 48.26045899999917,
+              "max": 65.11799999999948,
+              "hz": 18.56492505540262,
+              "period": 53.86501679999983,
+              "mean": 53.86501679999983,
+              "variance": 27.003718600062566,
+              "sd": 5.196510232844978,
+              "sem": 1.6432808220162056,
+              "df": 9,
+              "critical": 2.262,
+              "moe": 3.717101219400657,
+              "p75": 54.15062500000022,
+              "p99": 65.11799999999948,
+              "p995": 65.11799999999948,
+              "p999": 65.11799999999948,
+              "sampleCount": 10,
+              "median": 53.544791499999974
+            },
+            {
+              "id": "-1072438084_2_2",
+              "name": "extract every 10th page from 2000-page PDF (200 pages)",
+              "rank": 3,
+              "rme": 3.996294175601932,
+              "samples": [],
+              "totalTime": 509.82821200000035,
+              "min": 52.99541700000009,
+              "max": 62.211042000000816,
+              "hz": 17.65300504790424,
+              "period": 56.64757911111115,
+              "mean": 56.64757911111115,
+              "variance": 8.673642398208653,
+              "sd": 2.9451048195622263,
+              "sem": 0.9817016065207421,
+              "df": 8,
+              "critical": 2.306,
+              "moe": 2.2638039046368315,
+              "p75": 58.50329199999942,
+              "p99": 62.211042000000816,
+              "p995": 62.211042000000816,
+              "p999": 62.211042000000816,
+              "sampleCount": 9,
+              "median": 56.19641700000011
+            }
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/src/api/pdf.ts b/src/api/pdf.ts
index 1ac4296..d34fc69 100644
--- a/src/api/pdf.ts
+++ b/src/api/pdf.ts
@@ -1698,6 +1698,7 @@ export class PDF {
    * const [duplicate] = await pdf.copyPagesFrom(pdf, [0], { insertAt: 1 });
    * ```
    */
+  // oxlint-disable-next-line typescript/require-await -- Public async API kept for backward compat; ObjectCopier is sync.
   async copyPagesFrom(
     source: PDF,
     indices: number[],
@@ -1729,7 +1730,7 @@ export class PDF {
         throw new Error(`Source page ${index} not found`);
       }
 
-      const copiedPageRef = await copier.copyPage(srcPage.ref);
+      const copiedPageRef = copier.copyPage(srcPage.ref);
       copiedRefs.push(copiedPageRef);
     }
 
@@ -1824,6 +1825,7 @@ export class PDF {
    * }
    * ```
    */
+  // oxlint-disable-next-line typescript/require-await -- Public async API kept for backward compat; ObjectCopier is sync.
   async embedPage(source: PDF, pageIndex: number): Promise<PDFEmbeddedPage> {
     const srcPage = source.getPage(pageIndex);
 
@@ -1842,7 +1844,7 @@ export class PDF {
     let resources: PdfDict;
 
     if (srcResources) {
-      const copied = await copier.copyObject(srcResources);
+      const copied = copier.copyObject(srcResources);
 
       // This is guaranteed by our checks above
       resources = copied instanceof PdfDict ? copied : new PdfDict();
diff --git a/src/document/object-copier.test.ts b/src/document/object-copier.test.ts
index fae721f..b62deab 100644
--- a/src/document/object-copier.test.ts
+++ b/src/document/object-copier.test.ts
@@ -23,7 +23,7 @@ describe("ObjectCopier", () => {
 
       const copier = new ObjectCopier(source, dest);
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       // The copied page should be registered in dest
       expect(copiedPageRef).toBeInstanceOf(PdfRef);
@@ -47,7 +47,7 @@ describe("ObjectCopier", () => {
 
       const copier = new ObjectCopier(source, dest);
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
 
@@ -73,7 +73,7 @@ describe("ObjectCopier", () => {
       expect(srcPage.has("Parent")).toBe(true);
 
       const copier = new ObjectCopier(source, dest);
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
       expect(copiedPage.has("Parent")).toBe(false);
@@ -92,7 +92,7 @@ describe("ObjectCopier", () => {
       expect(srcPage.has("Annots")).toBe(true);
 
       const copier = new ObjectCopier(source, dest, { includeAnnotations: false });
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
       expect(copiedPage.has("Annots")).toBe(false);
@@ -111,7 +111,7 @@ describe("ObjectCopier", () => {
       expect(srcPage.has("Annots")).toBe(true);
 
       const copier = new ObjectCopier(source, dest);
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
       expect(copiedPage.has("Annots")).toBe(true);
@@ -138,7 +138,7 @@ describe("ObjectCopier", () => {
       const copier = new ObjectCopier(source, dest);
       const fakeRef = PdfRef.of(99999, 0);
 
-      await expect(copier.copyPage(fakeRef)).rejects.toThrow(/not found/);
+      expect(() => copier.copyPage(fakeRef)).toThrow(/not found/);
     });
 
     it("copies page Resources", async () => {
@@ -150,7 +150,7 @@ describe("ObjectCopier", () => {
 
       const copier = new ObjectCopier(source, dest);
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
 
@@ -181,9 +181,9 @@ describe("ObjectCopier", () => {
       const str = PdfString.fromString("hello");
 
       // Primitives should be returned as-is (they're immutable)
-      expect(await copier.copyObject(name)).toBe(name);
-      expect(await copier.copyObject(num)).toBe(num);
-      expect(await copier.copyObject(str)).toBe(str);
+      expect(copier.copyObject(name)).toBe(name);
+      expect(copier.copyObject(num)).toBe(num);
+      expect(copier.copyObject(str)).toBe(str);
     });
 
     it("creates new instance for arrays", async () => {
@@ -195,7 +195,7 @@ describe("ObjectCopier", () => {
 
       const arr = new PdfArray([PdfNumber.of(1), PdfNumber.of(2), PdfName.of("Test")]);
 
-      const copied = await copier.copyObject(arr);
+      const copied = copier.copyObject(arr);
       expect(copied).toBeInstanceOf(PdfArray);
       expect(copied).not.toBe(arr);
       expect(copied.length).toBe(3);
@@ -218,7 +218,7 @@ describe("ObjectCopier", () => {
         Key2: PdfString.fromString("value"),
       });
 
-      const copied = await copier.copyObject(dict);
+      const copied = copier.copyObject(dict);
       expect(copied).toBeInstanceOf(PdfDict);
       expect(copied).not.toBe(dict);
 
@@ -241,7 +241,7 @@ describe("ObjectCopier", () => {
         Array: innerArr,
       });
 
-      const copied = await copier.copyObject(outerDict);
+      const copied = copier.copyObject(outerDict);
 
       // Outer should be new
       expect(copied).not.toBe(outerDict);
@@ -269,7 +269,7 @@ describe("ObjectCopier", () => {
       const copier = new ObjectCopier(source, dest);
 
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       // Refs should be different
       expect(copiedPageRef.objectNumber).not.toBe(srcPageRef.objectNumber);
@@ -290,8 +290,8 @@ describe("ObjectCopier", () => {
 
       // Copy same page twice - internal resources should be deduplicated
       const srcPageRef = source.getPage(0)!.ref;
-      const copied1 = await copier.copyPage(srcPageRef);
-      const copied2 = await copier.copyPage(srcPageRef);
+      const copied1 = copier.copyPage(srcPageRef);
+      const copied2 = copier.copyPage(srcPageRef);
 
       // Page refs are different (each copyPage registers a new page)
       expect(copied1.objectNumber).not.toBe(copied2.objectNumber);
@@ -321,7 +321,7 @@ describe("ObjectCopier", () => {
       const srcPageRef = source.getPage(0)!.ref;
 
       // This should not throw due to circular references
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
       expect(copiedPageRef).toBeInstanceOf(PdfRef);
 
       // Page should be valid
@@ -342,7 +342,7 @@ describe("ObjectCopier", () => {
 
       const copier = new ObjectCopier(source, dest);
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
 
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
 
@@ -365,7 +365,7 @@ describe("ObjectCopier", () => {
       const srcPageRef = source.getPage(0)!.ref;
 
       // Should not throw - streams are decoded and re-encoded
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
       const copiedPage = dest.getObject(copiedPageRef) as PdfDict;
 
       expect(copiedPage.getName("Type")?.value).toBe("Page");
@@ -381,7 +381,7 @@ describe("ObjectCopier", () => {
       const originalPageRef = pdf.getPage(0)!.ref;
 
       const copier = new ObjectCopier(pdf, pdf);
-      const duplicatedRef = await copier.copyPage(originalPageRef);
+      const duplicatedRef = copier.copyPage(originalPageRef);
 
       // Refs should be different
       expect(duplicatedRef.objectNumber).not.toBe(originalPageRef.objectNumber);
@@ -411,7 +411,7 @@ describe("ObjectCopier", () => {
       // Copy a page
       const copier = new ObjectCopier(source, dest);
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
       dest.insertPage(destOriginalCount, copiedPageRef);
 
       // Save
@@ -444,7 +444,7 @@ describe("ObjectCopier", () => {
       for (let i = 0; i < source.getPageCount(); i++) {
         const copier = new ObjectCopier(source, dest);
         const srcPageRef = source.getPage(i)!.ref;
-        const copiedPageRef = await copier.copyPage(srcPageRef);
+        const copiedPageRef = copier.copyPage(srcPageRef);
         dest.insertPage(dest.getPageCount(), copiedPageRef);
       }
 
@@ -468,7 +468,7 @@ describe("ObjectCopier", () => {
       for (let i = 0; i < originalCount; i++) {
         const copier = new ObjectCopier(pdf, pdf);
         const srcPageRef = pdf.getPage(i)!.ref;
-        const copiedPageRef = await copier.copyPage(srcPageRef);
+        const copiedPageRef = copier.copyPage(srcPageRef);
         // Insert duplicate after the original
         pdf.insertPage(i * 2 + 1, copiedPageRef);
       }
@@ -493,7 +493,7 @@ describe("ObjectCopier", () => {
       // Copy form page with annotations
       const copier = new ObjectCopier(source, dest, { includeAnnotations: true });
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
       dest.insertPage(dest.getPageCount(), copiedPageRef);
 
       const savedBytes = await dest.save();
@@ -516,7 +516,7 @@ describe("ObjectCopier", () => {
       // Copy from encrypted source
       const copier = new ObjectCopier(source, dest);
       const srcPageRef = source.getPage(0)!.ref;
-      const copiedPageRef = await copier.copyPage(srcPageRef);
+      const copiedPageRef = copier.copyPage(srcPageRef);
       dest.insertPage(dest.getPageCount(), copiedPageRef);
 
       const savedBytes = await dest.save();
diff --git a/src/document/object-copier.ts b/src/document/object-copier.ts
index 78c8c15..7e5d719 100644
--- a/src/document/object-copier.ts
+++ b/src/document/object-copier.ts
@@ -10,6 +10,8 @@
  * - Flattens inherited page attributes during page copy
  * - Smart stream handling: raw bytes if unencrypted, re-encode if encrypted
  * - Circular reference detection
+ *
+ * All operations are synchronous — there is no I/O involved in object copying.
  */
 
 import type { PDF } from "#src/api/pdf";
@@ -46,7 +48,7 @@ const INHERITABLE_PAGE_ATTRS = ["Resources", "MediaBox", "CropBox", "Rotate"] as
  * @example
  * ```typescript
  * const copier = new ObjectCopier(sourcePdf, destPdf);
- * const copiedPageRef = await copier.copyPage(sourcePageRef);
+ * const copiedPageRef = copier.copyPage(sourcePageRef);
  * destPdf.insertPage(0, copiedPageRef);
  * ```
  */
@@ -79,7 +81,7 @@ export class ObjectCopier {
    * @param srcPageRef Reference to the page in source document
    * @returns Reference to the copied page in destination document
    */
-  async copyPage(srcPageRef: PdfRef): Promise<PdfRef> {
+  copyPage(srcPageRef: PdfRef): PdfRef {
     const srcPage = this.source.getObject(srcPageRef);
 
     if (!(srcPage instanceof PdfDict)) {
@@ -98,7 +100,7 @@ export class ObjectCopier {
 
         if (inherited) {
           // Deep copy the inherited value
-          const copied = await this.copyObject(inherited);
+          const copied = this.copyObject(inherited);
           cloned.set(key, copied);
         }
       }
@@ -125,7 +127,7 @@ export class ObjectCopier {
     cloned.delete("Parent");
 
     // Deep copy all values in the cloned dict, remapping refs
-    const copiedPage = await this.copyDictValues(cloned);
+    const copiedPage = this.copyDictValues(cloned);
 
     // Register in destination and return ref
     return this.dest.register(copiedPage);
@@ -134,25 +136,25 @@ export class ObjectCopier {
   /**
    * Deep copy any PDF object, remapping references to destination.
    */
-  async copyObject<T extends PdfObject>(obj: T): Promise<T> {
+  copyObject<T extends PdfObject>(obj: T): T {
     if (obj instanceof PdfRef) {
       // oxlint-disable-next-line typescript/no-unsafe-type-assertion
-      return (await this.copyRef(obj)) as unknown as T;
+      return this.copyRef(obj) as unknown as T;
     }
 
     if (obj instanceof PdfStream) {
       // oxlint-disable-next-line typescript/no-unsafe-type-assertion
-      return (await this.copyStream(obj)) as unknown as T;
+      return this.copyStream(obj) as unknown as T;
     }
 
     if (obj instanceof PdfDict) {
       // oxlint-disable-next-line typescript/no-unsafe-type-assertion
-      return (await this.copyDict(obj)) as unknown as T;
+      return this.copyDict(obj) as unknown as T;
     }
 
     if (obj instanceof PdfArray) {
       // oxlint-disable-next-line typescript/no-unsafe-type-assertion
-      return (await this.copyArray(obj)) as unknown as T;
+      return this.copyArray(obj) as unknown as T;
     }
 
     // Primitives (PdfName, PdfNumber, PdfString, PdfBool, PdfNull)
@@ -166,7 +168,7 @@ export class ObjectCopier {
    * Handles circular references by registering a placeholder before
    * recursively copying the referenced object's contents.
    */
-  private async copyRef(ref: PdfRef): Promise<PdfRef> {
+  private copyRef(ref: PdfRef): PdfRef {
     const key = `${ref.objectNumber}:${ref.generation}`;
 
     // Already copied (or being copied)?
@@ -206,7 +208,7 @@ export class ObjectCopier {
       const items: PdfObject[] = [];
 
       for (const item of srcObj) {
-        items.push(await this.copyObject(item));
+        items.push(this.copyObject(item));
       }
 
       const copiedArr = new PdfArray(items);
@@ -226,7 +228,7 @@ export class ObjectCopier {
   /**
    * Copy a dict reference, handling circular references.
    */
-  private async copyDictRef(key: string, srcDict: PdfDict): Promise<PdfRef> {
+  private copyDictRef(key: string, srcDict: PdfDict): PdfRef {
     // Clone the dict shell first
     const cloned = srcDict.clone();
 
@@ -235,7 +237,7 @@ export class ObjectCopier {
     this.refMap.set(key, destRef);
 
     // Now copy all values (which may reference back to us)
-    await this.copyDictValues(cloned);
+    this.copyDictValues(cloned);
 
     return destRef;
   }
@@ -243,7 +245,7 @@ export class ObjectCopier {
   /**
    * Copy a stream reference, handling circular references and encryption.
    */
-  private async copyStreamRef(key: string, srcStream: PdfStream): Promise<PdfRef> {
+  private copyStreamRef(key: string, srcStream: PdfStream): PdfRef {
     const sourceWasEncrypted = this.source.isEncrypted;
 
     // Clone the stream's dictionary
@@ -310,7 +312,7 @@ export class ObjectCopier {
     // Now copy dict values (which may reference back to us)
     // Note: we modify the already-registered stream's dict entries
     for (const [entryKey, value] of clonedDict) {
-      const copied = await this.copyObject(value);
+      const copied = this.copyObject(value);
 
       copiedStream.set(entryKey.value, copied);
     }
@@ -321,7 +323,7 @@ export class ObjectCopier {
   /**
    * Copy a dictionary, remapping all reference values.
    */
-  private async copyDict(dict: PdfDict): Promise<PdfDict> {
+  private copyDict(dict: PdfDict): PdfDict {
     const cloned = dict.clone();
 
     return this.copyDictValues(cloned);
@@ -331,9 +333,9 @@ export class ObjectCopier {
    * Copy all values in a dictionary, remapping references.
    * Modifies the dict in place and returns it.
    */
-  private async copyDictValues(dict: PdfDict): Promise<PdfDict> {
+  private copyDictValues(dict: PdfDict): PdfDict {
     for (const [key, value] of dict) {
-      const copied = await this.copyObject(value);
+      const copied = this.copyObject(value);
 
       dict.set(key.value, copied);
     }
@@ -344,11 +346,11 @@ export class ObjectCopier {
   /**
    * Copy an array, remapping all reference elements.
    */
-  private async copyArray(arr: PdfArray): Promise<PdfArray> {
+  private copyArray(arr: PdfArray): PdfArray {
     const items: PdfObject[] = [];
 
     for (const item of arr) {
-      items.push(await this.copyObject(item));
+      items.push(this.copyObject(item));
     }
 
     return new PdfArray(items);
@@ -360,14 +362,14 @@ export class ObjectCopier {
    * If source wasn't encrypted, copies raw encoded bytes (fastest).
    * If source was encrypted, decodes and re-encodes with same filters.
    */
-  private async copyStream(stream: PdfStream): Promise<PdfStream> {
+  private copyStream(stream: PdfStream): PdfStream {
     const sourceWasEncrypted = this.source.isEncrypted;
 
     // Clone the stream's dictionary
     const clonedDict = stream.clone();
 
     // Copy dict values (remapping refs, but not stream data yet)
-    await this.copyDictValues(clonedDict);
+    this.copyDictValues(clonedDict);
 
     if (!sourceWasEncrypted) {
       // Source wasn't encrypted - copy raw encoded bytes directly

From a17416a9d455a138a90bc3f31f277bc06c858d0f Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 10:56:04 +1100
Subject: [PATCH 3/9] perf(cache): replace internal LRU cache with npm
 lru-cache
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The internal LRU cache did Map.delete()+Map.set() on every get() to
maintain recency ordering. The npm lru-cache package uses a
doubly-linked-list for O(1) operations without Map rehashing.

Benchmarks show significant gains especially on large PDF parsing:
- 2000-page split: 506.6ms → 432.3ms (1.17x incremental)
- Single page from 2000p: 41.0ms → 25.5ms (1.61x incremental)
- Cumulative from baseline: 1.35x–1.60x across split workloads
---
 bun.lock                                      |   5 +-
 package.json                                  |   1 +
 reports/baselines/comparison-npm-lru-cache.md |  39 +++
 reports/baselines/npm-lru-cache.json          | 224 ++++++++++++++++++
 src/helpers/lru-cache.test.ts                 |  18 +-
 src/helpers/lru-cache.ts                      |  91 +------
 src/objects/pdf-name.ts                       |   2 +-
 src/objects/pdf-ref.ts                        |   2 +-
 8 files changed, 283 insertions(+), 99 deletions(-)
 create mode 100644 reports/baselines/comparison-npm-lru-cache.md
 create mode 100644 reports/baselines/npm-lru-cache.json

diff --git a/bun.lock b/bun.lock
index 2d34aa2..23cceae 100644
--- a/bun.lock
+++ b/bun.lock
@@ -8,6 +8,7 @@
         "@noble/ciphers": "^2.1.1",
         "@noble/hashes": "^2.0.1",
         "@scure/base": "^2.0.0",
+        "lru-cache": "^11.2.6",
         "pako": "^2.1.0",
         "pkijs": "^3.3.3",
       },
@@ -492,7 +493,7 @@
 
     "long": ["long@5.3.2", "", {}, "sha512-mNAgZ1GmyNhD7AuqnTG3/VQ26o760+ZYBPKjPvugO8+nLbYfX6TVpJPseBvopbdY+qpZ/lKUnmEc1LeZYS3QAA=="],
 
-    "lru-cache": ["lru-cache@10.4.3", "", {}, "sha512-JNAzZcXrCt42VGLuYz0zfAzDfAvJWW6AfYlDBQyDV5DClI2m5sAmK+OIO7s59XfsRsWHp02jAJrRadPRGTt6SQ=="],
+    "lru-cache": ["lru-cache@11.2.6", "", {}, "sha512-ESL2CrkS/2wTPfuend7Zhkzo2u0daGJ/A2VucJOgQ/C48S/zB8MMeMHSGKYpXhIjbPxfuezITkaBH1wqv00DDQ=="],
 
     "magic-string": ["magic-string@0.30.21", "", { "dependencies": { "@jridgewell/sourcemap-codec": "^1.5.5" } }, "sha512-vd2F4YUyEXKGcLHoq+TEyCjxueSeHnFxyyjNp80yg0XV4vUhnDer/lvvlqM/arB5bXQN5K2/3oinyCRyx8T2CQ=="],
 
@@ -706,6 +707,8 @@
 
     "micromatch/picomatch": ["picomatch@2.3.1", "", {}, "sha512-JU3teHTNjmE2VCGFzuY8EXzCDVwEqB2a8fsIvwaStHhAWJEeVd1o1QD80CU6+ZdEXXSLbSsuLwJjkCBWqRQUVA=="],
 
+    "path-scurry/lru-cache": ["lru-cache@10.4.3", "", {}, "sha512-JNAzZcXrCt42VGLuYz0zfAzDfAvJWW6AfYlDBQyDV5DClI2m5sAmK+OIO7s59XfsRsWHp02jAJrRadPRGTt6SQ=="],
+
     "pdf-lib/pako": ["pako@1.0.11", "", {}, "sha512-4hLB8Py4zZce5s4yd9XzopqwVv/yGNhV1Bl8NTmCq1763HeK2+EwVTv+leGeL13Dnh2wfbqowVPXCIO0z4taYw=="],
 
     "pkijs/@noble/hashes": ["@noble/hashes@1.4.0", "", {}, "sha512-V1JJ1WTRUqHHrOSh597hURcMqVKVGL/ea3kv0gSnEdsEZ0/+VyPghM1lMNGc00z7CIQorSvbKpuJkxvuHbvdbg=="],
diff --git a/package.json b/package.json
index 230f00b..0005466 100644
--- a/package.json
+++ b/package.json
@@ -67,6 +67,7 @@
     "@noble/ciphers": "^2.1.1",
     "@noble/hashes": "^2.0.1",
     "@scure/base": "^2.0.0",
+    "lru-cache": "^11.2.6",
     "pako": "^2.1.0",
     "pkijs": "^3.3.3"
   },
diff --git a/reports/baselines/comparison-npm-lru-cache.md b/reports/baselines/comparison-npm-lru-cache.md
new file mode 100644
index 0000000..6bba72c
--- /dev/null
+++ b/reports/baselines/comparison-npm-lru-cache.md
@@ -0,0 +1,39 @@
+# Benchmark Comparison: npm lru-cache
+
+## Change
+
+Replaced internal LRU cache (Map delete+set on every get) with the npm `lru-cache`
+package (v11, doubly-linked-list, O(1) get/set without Map rehashing).
+
+## Results (vs original baseline)
+
+| Benchmark                                              | Baseline (ms) | Sync Copier (ms) | + npm lru-cache (ms) | Total Speedup |
+| :----------------------------------------------------- | ------------: | ---------------: | -------------------: | ------------: |
+| extractPages (1 page from small PDF)                   |          0.45 |             0.40 |                 0.38 |         1.17x |
+| extractPages (1 page from 100-page PDF)                |          1.78 |             1.74 |                 1.66 |         1.07x |
+| extractPages (1 page from 2000-page PDF)               |         40.86 |            41.04 |                25.50 |         1.60x |
+| split 100-page PDF (0.1MB)                             |         31.64 |            27.35 |                24.59 |         1.29x |
+| split 2000-page PDF (0.9MB)                            |        582.50 |           506.61 |               432.27 |         1.35x |
+| extract first 10 pages from 2000-page PDF              |         42.87 |            43.21 |                27.31 |         1.57x |
+| extract first 100 pages from 2000-page PDF             |         50.92 |            53.87 |                35.21 |         1.45x |
+| extract every 10th page from 2000-page PDF (200 pages) |         58.78 |            56.65 |                40.52 |         1.45x |
+
+## Incremental improvement (sync copier -> + npm lru-cache)
+
+| Benchmark                                              | Sync Copier (ms) | + npm lru-cache (ms) | Incremental Speedup |
+| :----------------------------------------------------- | ---------------: | -------------------: | ------------------: |
+| extractPages (1 page from small PDF)                   |             0.40 |                 0.38 |               1.03x |
+| extractPages (1 page from 100-page PDF)                |             1.74 |                 1.66 |               1.05x |
+| extractPages (1 page from 2000-page PDF)               |            41.04 |                25.50 |               1.61x |
+| split 100-page PDF (0.1MB)                             |            27.35 |                24.59 |               1.11x |
+| split 2000-page PDF (0.9MB)                            |           506.61 |               432.27 |               1.17x |
+| extract first 10 pages from 2000-page PDF              |            43.21 |                27.31 |               1.58x |
+| extract first 100 pages from 2000-page PDF             |            53.87 |                35.21 |               1.53x |
+| extract every 10th page from 2000-page PDF (200 pages) |            56.65 |                40.52 |               1.40x |
+
+## Key Takeaways
+
+- **2000-page split**: 582.5ms -> 432.3ms (1.35x faster total)
+- **100-page split**: 31.6ms -> 24.6ms (1.29x faster total)
+- **Single page from 2000p**: 40.9ms -> 25.5ms (1.60x faster total)
+- **npm lru-cache incremental gain on 2000-page split**: 506.6ms -> 432.3ms (1.17x)
diff --git a/reports/baselines/npm-lru-cache.json b/reports/baselines/npm-lru-cache.json
new file mode 100644
index 0000000..346c580
--- /dev/null
+++ b/reports/baselines/npm-lru-cache.json
@@ -0,0 +1,224 @@
+{
+  "files": [
+    {
+      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
+      "groups": [
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
+          "benchmarks": [
+            {
+              "id": "-1072438084_0_0",
+              "name": "extractPages (1 page from small PDF)",
+              "rank": 1,
+              "rme": 0.6803617904005139,
+              "samples": [],
+              "totalTime": 500.2104360000012,
+              "min": 0.31749999999988177,
+              "max": 0.7807909999999083,
+              "hz": 2598.9061931526694,
+              "period": 0.3847772584615394,
+              "mean": 0.3847772584615394,
+              "variance": 0.002319155302547339,
+              "sd": 0.04815760897872048,
+              "sem": 0.0013356517575117053,
+              "df": 1299,
+              "critical": 1.96,
+              "moe": 0.0026178774447229423,
+              "p75": 0.38512500000001637,
+              "p99": 0.615833000000066,
+              "p995": 0.6623329999999896,
+              "p999": 0.7236249999998563,
+              "sampleCount": 1300,
+              "median": 0.37283300000012787
+            },
+            {
+              "id": "-1072438084_0_1",
+              "name": "extractPages (1 page from 100-page PDF)",
+              "rank": 2,
+              "rme": 0.787155411012552,
+              "samples": [],
+              "totalTime": 500.06587099999865,
+              "min": 1.5184580000000096,
+              "max": 2.1133749999999054,
+              "hz": 601.92070176291,
+              "period": 1.6613484086378694,
+              "mean": 1.6613484086378694,
+              "variance": 0.013399752069859515,
+              "sd": 0.11575729812784814,
+              "sem": 0.006672139742022405,
+              "df": 300,
+              "critical": 1.96,
+              "moe": 0.013077393894363914,
+              "p75": 1.7012500000000728,
+              "p99": 2.0060419999999795,
+              "p995": 2.059125000000222,
+              "p999": 2.1133749999999054,
+              "sampleCount": 301,
+              "median": 1.6227090000002136
+            },
+            {
+              "id": "-1072438084_0_2",
+              "name": "extractPages (1 page from 2000-page PDF)",
+              "rank": 3,
+              "rme": 1.172293974657617,
+              "samples": [],
+              "totalTime": 509.9992109999989,
+              "min": 24.20520899999974,
+              "max": 26.485957999999755,
+              "hz": 39.215746943577216,
+              "period": 25.499960549999948,
+              "mean": 25.499960549999948,
+              "variance": 0.40798446042404196,
+              "sd": 0.6387366127161038,
+              "sem": 0.14282584857511646,
+              "df": 19,
+              "critical": 2.093,
+              "moe": 0.29893450106771874,
+              "p75": 26.051500000000033,
+              "p99": 26.485957999999755,
+              "p995": 26.485957999999755,
+              "p999": 26.485957999999755,
+              "sampleCount": 20,
+              "median": 25.388771000000133
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
+          "benchmarks": [
+            {
+              "id": "-1072438084_1_0",
+              "name": "split 100-page PDF (0.1MB)",
+              "rank": 1,
+              "rme": 1.5707569902861633,
+              "samples": [],
+              "totalTime": 516.3171660000003,
+              "min": 23.525666999999885,
+              "max": 26.712125000000015,
+              "hz": 40.672674439028796,
+              "period": 24.586531714285726,
+              "mean": 24.586531714285726,
+              "variance": 0.7197857225650164,
+              "sd": 0.8484018638387214,
+              "sem": 0.18513646479916948,
+              "df": 20,
+              "critical": 2.086,
+              "moe": 0.3861946655710675,
+              "p75": 24.975667000000158,
+              "p99": 26.712125000000015,
+              "p995": 26.712125000000015,
+              "p999": 26.712125000000015,
+              "sampleCount": 21,
+              "median": 24.294166000000132
+            },
+            {
+              "id": "-1072438084_1_1",
+              "name": "split 2000-page PDF (0.9MB)",
+              "rank": 2,
+              "rme": 0,
+              "samples": [],
+              "totalTime": 432.27162499999986,
+              "min": 432.27162499999986,
+              "max": 432.27162499999986,
+              "hz": 2.313360262774593,
+              "period": 432.27162499999986,
+              "mean": 432.27162499999986,
+              "variance": 0,
+              "sd": 0,
+              "sem": 0,
+              "df": 0,
+              "critical": 12.71,
+              "moe": 0,
+              "p75": 432.27162499999986,
+              "p99": 432.27162499999986,
+              "p995": 432.27162499999986,
+              "p999": 432.27162499999986,
+              "sampleCount": 1,
+              "median": 432.27162499999986
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
+          "benchmarks": [
+            {
+              "id": "-1072438084_2_0",
+              "name": "extract first 10 pages from 2000-page PDF",
+              "rank": 1,
+              "rme": 2.223598050281543,
+              "samples": [],
+              "totalTime": 518.8923730000015,
+              "min": 26.314790999999786,
+              "max": 30.981832999999824,
+              "hz": 36.61645649202854,
+              "period": 27.31012489473692,
+              "mean": 27.31012489473692,
+              "variance": 1.5873088504924244,
+              "sd": 1.2598844591836287,
+              "sem": 0.28903731779572794,
+              "df": 18,
+              "critical": 2.101,
+              "moe": 0.6072674046888245,
+              "p75": 27.280499999999847,
+              "p99": 30.981832999999824,
+              "p995": 30.981832999999824,
+              "p999": 30.981832999999824,
+              "sampleCount": 19,
+              "median": 26.98291700000027
+            },
+            {
+              "id": "-1072438084_2_1",
+              "name": "extract first 100 pages from 2000-page PDF",
+              "rank": 2,
+              "rme": 2.213921234147061,
+              "samples": [],
+              "totalTime": 528.2034549999998,
+              "min": 34.1123749999997,
+              "max": 39.86729100000048,
+              "hz": 28.398148209765125,
+              "period": 35.21356366666666,
+              "mean": 35.21356366666666,
+              "variance": 1.9814401301546256,
+              "sd": 1.4076363628986805,
+              "sem": 0.3634501460681713,
+              "df": 14,
+              "critical": 2.145,
+              "moe": 0.7796005633162274,
+              "p75": 35.0894159999998,
+              "p99": 39.86729100000048,
+              "p995": 39.86729100000048,
+              "p999": 39.86729100000048,
+              "sampleCount": 15,
+              "median": 34.765957999999955
+            },
+            {
+              "id": "-1072438084_2_2",
+              "name": "extract every 10th page from 2000-page PDF (200 pages)",
+              "rank": 3,
+              "rme": 2.0448631675208837,
+              "samples": [],
+              "totalTime": 526.8241219999973,
+              "min": 38.62562499999967,
+              "max": 42.899374999999964,
+              "hz": 24.67616697323527,
+              "period": 40.52493246153825,
+              "mean": 40.52493246153825,
+              "variance": 1.8801911150990929,
+              "sd": 1.3712006108148773,
+              "sem": 0.38030262394158326,
+              "df": 12,
+              "critical": 2.179,
+              "moe": 0.8286794175687099,
+              "p75": 41.67979099999957,
+              "p99": 42.899374999999964,
+              "p995": 42.899374999999964,
+              "p999": 42.899374999999964,
+              "sampleCount": 13,
+              "median": 40.114832999999635
+            }
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/src/helpers/lru-cache.test.ts b/src/helpers/lru-cache.test.ts
index acb3fe8..7385bc4 100644
--- a/src/helpers/lru-cache.test.ts
+++ b/src/helpers/lru-cache.test.ts
@@ -4,7 +4,7 @@ import { LRUCache } from "./lru-cache";
 
 describe("LRUCache", () => {
   it("stores and retrieves values", () => {
-    const cache = new LRUCache<string, number>(10);
+    const cache = new LRUCache<string, number>({ max: 10 });
 
     cache.set("a", 1);
     cache.set("b", 2);
@@ -15,7 +15,7 @@ describe("LRUCache", () => {
   });
 
   it("updates existing values", () => {
-    const cache = new LRUCache<string, number>(10);
+    const cache = new LRUCache<string, number>({ max: 10 });
 
     cache.set("a", 1);
     cache.set("a", 2);
@@ -25,7 +25,7 @@ describe("LRUCache", () => {
   });
 
   it("evicts least recently used when at capacity", () => {
-    const cache = new LRUCache<string, number>(3);
+    const cache = new LRUCache<string, number>({ max: 3 });
 
     cache.set("a", 1);
     cache.set("b", 2);
@@ -45,7 +45,7 @@ describe("LRUCache", () => {
   });
 
   it("get updates recency", () => {
-    const cache = new LRUCache<string, number>(3);
+    const cache = new LRUCache<string, number>({ max: 3 });
 
     cache.set("a", 1);
     cache.set("b", 2);
@@ -64,7 +64,7 @@ describe("LRUCache", () => {
   });
 
   it("has checks existence without updating recency", () => {
-    const cache = new LRUCache<string, number>(3);
+    const cache = new LRUCache<string, number>({ max: 3 });
 
     cache.set("a", 1);
     cache.set("b", 2);
@@ -81,7 +81,7 @@ describe("LRUCache", () => {
   });
 
   it("delete removes entries", () => {
-    const cache = new LRUCache<string, number>(10);
+    const cache = new LRUCache<string, number>({ max: 10 });
 
     cache.set("a", 1);
     cache.set("b", 2);
@@ -93,7 +93,7 @@ describe("LRUCache", () => {
   });
 
   it("clear removes all entries", () => {
-    const cache = new LRUCache<string, number>(10);
+    const cache = new LRUCache<string, number>({ max: 10 });
 
     cache.set("a", 1);
     cache.set("b", 2);
@@ -105,8 +105,8 @@ describe("LRUCache", () => {
     expect(cache.get("a")).toBeUndefined();
   });
 
-  it("works with default size", () => {
-    const cache = new LRUCache<string, number>();
+  it("works with large capacity", () => {
+    const cache = new LRUCache<string, number>({ max: 10000 });
 
     // Should be able to add many items
     for (let i = 0; i < 1000; i++) {
diff --git a/src/helpers/lru-cache.ts b/src/helpers/lru-cache.ts
index cb64792..77ef617 100644
--- a/src/helpers/lru-cache.ts
+++ b/src/helpers/lru-cache.ts
@@ -1,91 +1,8 @@
 /**
- * Simple LRU (Least Recently Used) cache implementation.
+ * LRU cache re-export.
  *
- * Used for interning frequently-used PDF objects (PdfName, PdfRef)
- * while preventing unbounded memory growth.
+ * Wraps the `lru-cache` npm package to provide the same interface used
+ * by PdfName and PdfRef interning caches.
  */
 
-/**
- * A bounded cache that evicts least-recently-used entries when full.
- *
- * @typeParam K - Key type
- * @typeParam V - Value type
- */
-export class LRUCache<K, V> {
-  private readonly maxSize: number;
-  private readonly cache = new Map<K, V>();
-
-  /**
-   * Create a new LRU cache.
-   *
-   * @param maxSize - Maximum number of entries to retain (default: 10000)
-   */
-  constructor(maxSize = 10000) {
-    this.maxSize = maxSize;
-  }
-
-  /**
-   * Get a value from the cache, updating its recency.
-   *
-   * @returns The cached value, or undefined if not present
-   */
-  get(key: K): V | undefined {
-    const value = this.cache.get(key);
-
-    if (value !== undefined) {
-      // Move to end (most recently used)
-      this.cache.delete(key);
-      this.cache.set(key, value);
-    }
-
-    return value;
-  }
-
-  /**
-   * Check if a key exists in the cache (without updating recency).
-   */
-  has(key: K): boolean {
-    return this.cache.has(key);
-  }
-
-  /**
-   * Add or update a value in the cache.
-   *
-   * If the cache is at capacity, the least-recently-used entry is evicted.
-   */
-  set(key: K, value: V): void {
-    // If key exists, delete it first so it becomes the most recent
-    if (this.cache.has(key)) {
-      this.cache.delete(key);
-    } else if (this.cache.size >= this.maxSize) {
-      // Evict the oldest entry (first in Map iteration order)
-      const oldestKey = this.cache.keys().next().value;
-      if (oldestKey !== undefined) {
-        this.cache.delete(oldestKey);
-      }
-    }
-
-    this.cache.set(key, value);
-  }
-
-  /**
-   * Remove a value from the cache.
-   */
-  delete(key: K): boolean {
-    return this.cache.delete(key);
-  }
-
-  /**
-   * Clear all entries from the cache.
-   */
-  clear(): void {
-    this.cache.clear();
-  }
-
-  /**
-   * Get the current number of entries in the cache.
-   */
-  get size(): number {
-    return this.cache.size;
-  }
-}
+export { LRUCache } from "lru-cache";
diff --git a/src/objects/pdf-name.ts b/src/objects/pdf-name.ts
index a8238e9..7ba0dc8 100644
--- a/src/objects/pdf-name.ts
+++ b/src/objects/pdf-name.ts
@@ -57,7 +57,7 @@ export class PdfName implements PdfPrimitive {
     return "name";
   }
 
-  private static cache = new LRUCache<string, PdfName>(DEFAULT_NAME_CACHE_SIZE);
+  private static cache = new LRUCache<string, PdfName>({ max: DEFAULT_NAME_CACHE_SIZE });
 
   /**
    * Pre-cached common names that should never be evicted.
diff --git a/src/objects/pdf-ref.ts b/src/objects/pdf-ref.ts
index 3c4e474..77312c7 100644
--- a/src/objects/pdf-ref.ts
+++ b/src/objects/pdf-ref.ts
@@ -23,7 +23,7 @@ export class PdfRef implements PdfPrimitive {
     return "ref";
   }
 
-  private static cache = new LRUCache<string, PdfRef>(DEFAULT_REF_CACHE_SIZE);
+  private static cache = new LRUCache<string, PdfRef>({ max: DEFAULT_REF_CACHE_SIZE });
 
   private constructor(
     readonly objectNumber: number,

From 05fd0e9ec106ca0accd7a17654b086eafdf30212 Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 11:46:18 +1100
Subject: [PATCH 4/9] perf(writer): cache PdfName bytes, shared hex table, skip
 tiny deflate
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Three changes:
- PdfName.toBytes() caches serialized bytes on the interned instance
  (compute once, writeBytes on every subsequent call). ASCII fast-path
  skips TextEncoder entirely for the 99% of names that are pure ASCII.
- Shared HEX_TABLE in buffer.ts replaces per-byte toString(16) calls
  in both bytesToHex and escapeName.
- Skip deflate for streams under 512 bytes (configurable via
  compressionThreshold). Deflate init zeros a 64KB hash table per call;
  for tiny streams the overhead dwarfs any savings.
- Expose compressStreams and compressionThreshold on SaveOptions.

Cumulative from baseline: 582ms → 245ms (2.38x) on 2000-page split.
---
 reports/baselines/escape-name-opt.json   | 224 +++++++++++++++++++++++
 reports/baselines/skip-tiny-deflate.json | 224 +++++++++++++++++++++++
 src/api/pdf.ts                           |  24 +++
 src/helpers/buffer.ts                    |   9 +-
 src/objects/pdf-name.ts                  |  53 +++++-
 src/writer/pdf-writer.ts                 |  35 +++-
 6 files changed, 554 insertions(+), 15 deletions(-)
 create mode 100644 reports/baselines/escape-name-opt.json
 create mode 100644 reports/baselines/skip-tiny-deflate.json

diff --git a/reports/baselines/escape-name-opt.json b/reports/baselines/escape-name-opt.json
new file mode 100644
index 0000000..55cfe9e
--- /dev/null
+++ b/reports/baselines/escape-name-opt.json
@@ -0,0 +1,224 @@
+{
+  "files": [
+    {
+      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
+      "groups": [
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
+          "benchmarks": [
+            {
+              "id": "-1072438084_0_0",
+              "name": "extractPages (1 page from small PDF)",
+              "rank": 1,
+              "rme": 1.3598162943493441,
+              "samples": [],
+              "totalTime": 500.19283100000087,
+              "min": 0.3138329999999314,
+              "max": 2.999708000000055,
+              "hz": 2802.9190206446556,
+              "period": 0.35677092082739004,
+              "mean": 0.35677092082739004,
+              "variance": 0.008589645712391514,
+              "sd": 0.09268034156384791,
+              "sem": 0.002475218936179106,
+              "df": 1401,
+              "critical": 1.96,
+              "moe": 0.004851429114911048,
+              "p75": 0.3457499999999527,
+              "p99": 0.5737500000000182,
+              "p995": 0.6079170000000431,
+              "p999": 0.9181670000000395,
+              "sampleCount": 1402,
+              "median": 0.333729500000004
+            },
+            {
+              "id": "-1072438084_0_1",
+              "name": "extractPages (1 page from 100-page PDF)",
+              "rank": 2,
+              "rme": 0.5969408576151285,
+              "samples": [],
+              "totalTime": 500.4518780000001,
+              "min": 1.4726660000001175,
+              "max": 2.0149169999999685,
+              "hz": 633.4275360637171,
+              "period": 1.5787125488958993,
+              "mean": 1.5787125488958993,
+              "variance": 0.007328512823299263,
+              "sd": 0.0856067335161158,
+              "sem": 0.0048081531778861415,
+              "df": 316,
+              "critical": 1.96,
+              "moe": 0.009423980228656837,
+              "p75": 1.589292000000114,
+              "p99": 1.8592499999999745,
+              "p995": 1.8713330000000497,
+              "p999": 2.0149169999999685,
+              "sampleCount": 317,
+              "median": 1.55074999999988
+            },
+            {
+              "id": "-1072438084_0_2",
+              "name": "extractPages (1 page from 2000-page PDF)",
+              "rank": 3,
+              "rme": 0.8810702753157642,
+              "samples": [],
+              "totalTime": 503.6995829999996,
+              "min": 24.38437500000009,
+              "max": 26.65845900000022,
+              "hz": 39.70620718183127,
+              "period": 25.184979149999982,
+              "mean": 25.184979149999982,
+              "variance": 0.2247997529768304,
+              "sd": 0.47413052314402876,
+              "sem": 0.10601880799575855,
+              "df": 19,
+              "critical": 2.093,
+              "moe": 0.22189736513512265,
+              "p75": 25.31841600000007,
+              "p99": 26.65845900000022,
+              "p995": 26.65845900000022,
+              "p999": 26.65845900000022,
+              "sampleCount": 20,
+              "median": 25.21672899999976
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
+          "benchmarks": [
+            {
+              "id": "-1072438084_1_0",
+              "name": "split 100-page PDF (0.1MB)",
+              "rank": 1,
+              "rme": 1.7048381748393946,
+              "samples": [],
+              "totalTime": 502.2399970000006,
+              "min": 19.57883300000003,
+              "max": 23.024417000000085,
+              "hz": 47.78591936794706,
+              "period": 20.92666654166669,
+              "mean": 20.92666654166669,
+              "variance": 0.7136030401800851,
+              "sd": 0.8447502827345399,
+              "sem": 0.17243392939762045,
+              "df": 23,
+              "critical": 2.069,
+              "moe": 0.3567657999236767,
+              "p75": 21.305459000000155,
+              "p99": 23.024417000000085,
+              "p995": 23.024417000000085,
+              "p999": 23.024417000000085,
+              "sampleCount": 24,
+              "median": 20.864958000000115
+            },
+            {
+              "id": "-1072438084_1_1",
+              "name": "split 2000-page PDF (0.9MB)",
+              "rank": 2,
+              "rme": 0,
+              "samples": [],
+              "totalTime": 369.2785839999997,
+              "min": 369.2785839999997,
+              "max": 369.2785839999997,
+              "hz": 2.7079826540929353,
+              "period": 369.2785839999997,
+              "mean": 369.2785839999997,
+              "variance": 0,
+              "sd": 0,
+              "sem": 0,
+              "df": 0,
+              "critical": 12.71,
+              "moe": 0,
+              "p75": 369.2785839999997,
+              "p99": 369.2785839999997,
+              "p995": 369.2785839999997,
+              "p999": 369.2785839999997,
+              "sampleCount": 1,
+              "median": 369.2785839999997
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
+          "benchmarks": [
+            {
+              "id": "-1072438084_2_0",
+              "name": "extract first 10 pages from 2000-page PDF",
+              "rank": 1,
+              "rme": 2.2077471741058514,
+              "samples": [],
+              "totalTime": 511.5155429999995,
+              "min": 25.185083999999733,
+              "max": 31.580417000000125,
+              "hz": 37.14452133471146,
+              "period": 26.9218706842105,
+              "mean": 26.9218706842105,
+              "variance": 1.5205847790653222,
+              "sd": 1.2331199370155859,
+              "sem": 0.2828971153008515,
+              "df": 18,
+              "critical": 2.101,
+              "moe": 0.594366839247089,
+              "p75": 27.052708000000166,
+              "p99": 31.580417000000125,
+              "p995": 31.580417000000125,
+              "p999": 31.580417000000125,
+              "sampleCount": 19,
+              "median": 26.722999999999956
+            },
+            {
+              "id": "-1072438084_2_1",
+              "name": "extract first 100 pages from 2000-page PDF",
+              "rank": 2,
+              "rme": 1.8050175681822698,
+              "samples": [],
+              "totalTime": 517.9517489999989,
+              "min": 33.36449999999968,
+              "max": 37.865499999999884,
+              "hz": 28.960226563497965,
+              "period": 34.53011659999993,
+              "mean": 34.53011659999993,
+              "variance": 1.2664720864594134,
+              "sd": 1.1253764198966556,
+              "sem": 0.2905709421649515,
+              "df": 14,
+              "critical": 2.145,
+              "moe": 0.6232746709438209,
+              "p75": 35.07374999999956,
+              "p99": 37.865499999999884,
+              "p995": 37.865499999999884,
+              "p999": 37.865499999999884,
+              "sampleCount": 15,
+              "median": 34.1270829999994
+            },
+            {
+              "id": "-1072438084_2_2",
+              "name": "extract every 10th page from 2000-page PDF (200 pages)",
+              "rank": 3,
+              "rme": 3.7908921701889815,
+              "samples": [],
+              "totalTime": 528.9624159999994,
+              "min": 37.919458000000304,
+              "max": 48.4839579999998,
+              "hz": 24.576415274086347,
+              "period": 40.689416615384566,
+              "mean": 40.689416615384566,
+              "variance": 6.5144038865667575,
+              "sd": 2.552333028146358,
+              "sem": 0.707889815772459,
+              "df": 12,
+              "critical": 2.179,
+              "moe": 1.542491908568188,
+              "p75": 40.70625000000018,
+              "p99": 48.4839579999998,
+              "p995": 48.4839579999998,
+              "p999": 48.4839579999998,
+              "sampleCount": 13,
+              "median": 40.056916999999885
+            }
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/reports/baselines/skip-tiny-deflate.json b/reports/baselines/skip-tiny-deflate.json
new file mode 100644
index 0000000..d11d64e
--- /dev/null
+++ b/reports/baselines/skip-tiny-deflate.json
@@ -0,0 +1,224 @@
+{
+  "files": [
+    {
+      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
+      "groups": [
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
+          "benchmarks": [
+            {
+              "id": "-1072438084_0_0",
+              "name": "extractPages (1 page from small PDF)",
+              "rank": 1,
+              "rme": 0.8892829061609049,
+              "samples": [],
+              "totalTime": 500.31841499999973,
+              "min": 0.2905420000001868,
+              "max": 1.9279579999999896,
+              "hz": 2924.1378213112557,
+              "period": 0.3419811449077237,
+              "mean": 0.3419811449077237,
+              "variance": 0.0035222193288500063,
+              "sd": 0.05934828833968176,
+              "sem": 0.001551622379570276,
+              "df": 1462,
+              "critical": 1.96,
+              "moe": 0.0030411798639577406,
+              "p75": 0.3381249999999909,
+              "p99": 0.5579999999999927,
+              "p995": 0.5912919999998394,
+              "p999": 0.6483749999999873,
+              "sampleCount": 1463,
+              "median": 0.3292500000002292
+            },
+            {
+              "id": "-1072438084_0_1",
+              "name": "extractPages (1 page from 100-page PDF)",
+              "rank": 2,
+              "rme": 0.6942473447539961,
+              "samples": [],
+              "totalTime": 500.1834050000034,
+              "min": 1.3748750000004293,
+              "max": 1.868916999999783,
+              "hz": 669.7543274151564,
+              "period": 1.4930847910447862,
+              "mean": 1.4930847910447862,
+              "variance": 0.009369794423687357,
+              "sd": 0.09679769844209808,
+              "sem": 0.005288623222833765,
+              "df": 334,
+              "critical": 1.96,
+              "moe": 0.010365701516754179,
+              "p75": 1.5309159999997064,
+              "p99": 1.8027499999998327,
+              "p995": 1.8117079999997259,
+              "p999": 1.868916999999783,
+              "sampleCount": 335,
+              "median": 1.4698750000002292
+            },
+            {
+              "id": "-1072438084_0_2",
+              "name": "extractPages (1 page from 2000-page PDF)",
+              "rank": 3,
+              "rme": 1.1888586380169315,
+              "samples": [],
+              "totalTime": 501.3285840000003,
+              "min": 24.17533400000002,
+              "max": 27.005499999999756,
+              "hz": 39.89399495321812,
+              "period": 25.066429200000016,
+              "mean": 25.066429200000016,
+              "variance": 0.4054496384268641,
+              "sd": 0.6367492743826758,
+              "sem": 0.14238146621433284,
+              "df": 19,
+              "critical": 2.093,
+              "moe": 0.2980044087865986,
+              "p75": 25.286291999999776,
+              "p99": 27.005499999999756,
+              "p995": 27.005499999999756,
+              "p999": 27.005499999999756,
+              "sampleCount": 20,
+              "median": 24.951479000000063
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
+          "benchmarks": [
+            {
+              "id": "-1072438084_1_0",
+              "name": "split 100-page PDF (0.1MB)",
+              "rank": 1,
+              "rme": 1.241299549103146,
+              "samples": [],
+              "totalTime": 509.5840829999979,
+              "min": 12.014166999999816,
+              "max": 13.951500000000124,
+              "hz": 78.49538738438218,
+              "period": 12.739602074999947,
+              "mean": 12.739602074999947,
+              "variance": 0.2443938158093934,
+              "sd": 0.4943620290934503,
+              "sem": 0.07816550003188641,
+              "df": 39,
+              "critical": 2.0231,
+              "moe": 0.15813662311450938,
+              "p75": 12.96362500000032,
+              "p99": 13.951500000000124,
+              "p995": 13.951500000000124,
+              "p999": 13.951500000000124,
+              "sampleCount": 40,
+              "median": 12.703645499999766
+            },
+            {
+              "id": "-1072438084_1_1",
+              "name": "split 2000-page PDF (0.9MB)",
+              "rank": 2,
+              "rme": 0,
+              "samples": [],
+              "totalTime": 244.68245800000022,
+              "min": 244.68245800000022,
+              "max": 244.68245800000022,
+              "hz": 4.086929680917294,
+              "period": 244.68245800000022,
+              "mean": 244.68245800000022,
+              "variance": 0,
+              "sd": 0,
+              "sem": 0,
+              "df": 0,
+              "critical": 12.71,
+              "moe": 0,
+              "p75": 244.68245800000022,
+              "p99": 244.68245800000022,
+              "p995": 244.68245800000022,
+              "p999": 244.68245800000022,
+              "sampleCount": 1,
+              "median": 244.68245800000022
+            }
+          ]
+        },
+        {
+          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
+          "benchmarks": [
+            {
+              "id": "-1072438084_2_0",
+              "name": "extract first 10 pages from 2000-page PDF",
+              "rank": 1,
+              "rme": 0.9557725912034225,
+              "samples": [],
+              "totalTime": 504.57903800000076,
+              "min": 24.360125000000153,
+              "max": 25.940999999999804,
+              "hz": 39.637001329413074,
+              "period": 25.228951900000038,
+              "mean": 25.228951900000038,
+              "variance": 0.2654600064528459,
+              "sd": 0.5152281110856102,
+              "sem": 0.11520850803062373,
+              "df": 19,
+              "critical": 2.093,
+              "moe": 0.24113140730809546,
+              "p75": 25.62262499999997,
+              "p99": 25.940999999999804,
+              "p995": 25.940999999999804,
+              "p999": 25.940999999999804,
+              "sampleCount": 20,
+              "median": 25.240332999999737
+            },
+            {
+              "id": "-1072438084_2_1",
+              "name": "extract first 100 pages from 2000-page PDF",
+              "rank": 2,
+              "rme": 0.8129752528372926,
+              "samples": [],
+              "totalTime": 509.13470899999993,
+              "min": 25.983500000000276,
+              "max": 28.018167000000176,
+              "hz": 37.31821787856149,
+              "period": 26.796563631578945,
+              "mean": 26.796563631578945,
+              "variance": 0.20427459388908475,
+              "sd": 0.45196746994566406,
+              "sem": 0.10368844880320557,
+              "df": 18,
+              "critical": 2.101,
+              "moe": 0.2178494309355349,
+              "p75": 27.132666999999856,
+              "p99": 28.018167000000176,
+              "p995": 28.018167000000176,
+              "p999": 28.018167000000176,
+              "sampleCount": 19,
+              "median": 26.71374999999989
+            },
+            {
+              "id": "-1072438084_2_2",
+              "name": "extract every 10th page from 2000-page PDF (200 pages)",
+              "rank": 3,
+              "rme": 1.7260235531528858,
+              "samples": [],
+              "totalTime": 526.418208000001,
+              "min": 28.0747080000001,
+              "max": 31.5014170000004,
+              "hz": 34.19334613896935,
+              "period": 29.245456000000054,
+              "mean": 29.245456000000054,
+              "variance": 1.0301911737761424,
+              "sd": 1.014983336698757,
+              "sem": 0.23923386672368,
+              "df": 17,
+              "critical": 2.11,
+              "moe": 0.5047834587869647,
+              "p75": 29.7938750000003,
+              "p99": 31.5014170000004,
+              "p995": 31.5014170000004,
+              "p999": 31.5014170000004,
+              "sampleCount": 18,
+              "median": 28.999896000000717
+            }
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/src/api/pdf.ts b/src/api/pdf.ts
index d34fc69..b2198ee 100644
--- a/src/api/pdf.ts
+++ b/src/api/pdf.ts
@@ -125,6 +125,26 @@ export interface SaveOptions {
    * @default false
    */
   subsetFonts?: boolean;
+
+  /**
+   * Compress uncompressed streams with FlateDecode (default: true).
+   *
+   * When enabled, streams without a /Filter entry will be compressed
+   * before writing. Streams that already have filters (including image
+   * formats like DCTDecode/JPXDecode) are left unchanged.
+   */
+  compressStreams?: boolean;
+
+  /**
+   * Minimum stream size in bytes to attempt compression (default: 512).
+   *
+   * Streams smaller than this threshold are written uncompressed.
+   * Deflate initialization has a fixed overhead that dominates for small
+   * payloads, and tiny streams rarely achieve meaningful compression.
+   *
+   * Set to 0 to compress all streams regardless of size.
+   */
+  compressionThreshold?: number;
 }
 
 /**
@@ -3141,6 +3161,8 @@ export class PDF {
         id: fileId,
         useXRefStream,
         securityHandler,
+        compressStreams: options.compressStreams,
+        compressionThreshold: options.compressionThreshold,
       });
 
       // Reset pending security state after successful save
@@ -3158,6 +3180,8 @@ export class PDF {
       id: fileId,
       useXRefStream,
       securityHandler,
+      compressStreams: options.compressStreams,
+      compressionThreshold: options.compressionThreshold,
     });
 
     // Reset pending security state after successful save
diff --git a/src/helpers/buffer.ts b/src/helpers/buffer.ts
index 2aab629..d5bdded 100644
--- a/src/helpers/buffer.ts
+++ b/src/helpers/buffer.ts
@@ -43,6 +43,13 @@ export function concatBytes(arrays: Uint8Array[]): Uint8Array {
   return result;
 }
 
+/** Pre-computed hex lookup: byte value → "XX" uppercase string. */
+export const HEX_TABLE: string[] = new Array(256);
+
+for (let i = 0; i < 256; i++) {
+  HEX_TABLE[i] = i.toString(16).toUpperCase().padStart(2, "0");
+}
+
 /**
  * Convert bytes to uppercase hex string.
  *
@@ -58,7 +65,7 @@ export function bytesToHex(bytes: Uint8Array): string {
   let hex = "";
 
   for (const byte of bytes) {
-    hex += byte.toString(16).toUpperCase().padStart(2, "0");
+    hex += HEX_TABLE[byte];
   }
 
   return hex;
diff --git a/src/objects/pdf-name.ts b/src/objects/pdf-name.ts
index 7ba0dc8..d52561c 100644
--- a/src/objects/pdf-name.ts
+++ b/src/objects/pdf-name.ts
@@ -1,3 +1,4 @@
+import { HEX_TABLE } from "#src/helpers/buffer";
 import { CHAR_HASH, DELIMITERS, WHITESPACE } from "#src/helpers/chars";
 import { LRUCache } from "#src/helpers/lru-cache";
 import type { ByteWriter } from "#src/io/byte-writer";
@@ -9,11 +10,25 @@ import type { PdfPrimitive } from "./pdf-primitive";
 // Plus anything outside printable ASCII (33-126)
 const NAME_NEEDS_ESCAPE = new Set([...WHITESPACE, ...DELIMITERS, CHAR_HASH]);
 
+/** Module-level encoder — avoids constructing one per escapeName call. */
+const textEncoder = new TextEncoder();
+
 /**
- * Default cache size for PdfName interning.
- * Can be overridden via PdfName.setCacheSize().
+ * Check whether a name is pure "safe" ASCII — every char is printable ASCII
+ * (33–126) and not in the escape set. If so, no escaping is needed and we
+ * can skip the TextEncoder entirely.
  */
-const DEFAULT_NAME_CACHE_SIZE = 10000;
+function isSimpleAsciiName(name: string): boolean {
+  for (let i = 0; i < name.length; i++) {
+    const c = name.charCodeAt(i);
+
+    if (c < 33 || c > 126 || NAME_NEEDS_ESCAPE.has(c)) {
+      return false;
+    }
+  }
+
+  return true;
+}
 
 /**
  * Escape a PDF name for serialization.
@@ -24,15 +39,18 @@ const DEFAULT_NAME_CACHE_SIZE = 10000;
  * - The # character itself
  */
 function escapeName(name: string): string {
-  const encoder = new TextEncoder();
-  const bytes = encoder.encode(name);
+  // Fast path: pure safe ASCII — no encoding or escaping needed
+  if (isSimpleAsciiName(name)) {
+    return name;
+  }
+
+  const bytes = textEncoder.encode(name);
 
   let result = "";
 
   for (const byte of bytes) {
     if (byte < 33 || byte > 126 || NAME_NEEDS_ESCAPE.has(byte)) {
-      // Use hex escape
-      result += `#${byte.toString(16).toUpperCase().padStart(2, "0")}`;
+      result += `#${HEX_TABLE[byte]}`;
     } else {
       result += String.fromCharCode(byte);
     }
@@ -41,6 +59,12 @@ function escapeName(name: string): string {
   return result;
 }
 
+/**
+ * Default cache size for PdfName interning.
+ * Can be overridden via PdfName.setCacheSize().
+ */
+const DEFAULT_NAME_CACHE_SIZE = 10000;
+
 /**
  * PDF name object (interned).
  *
@@ -80,6 +104,9 @@ export class PdfName implements PdfPrimitive {
   static readonly Filter = PdfName.createPermanent("Filter");
   static readonly FlateDecode = PdfName.createPermanent("FlateDecode");
 
+  /** Cached serialized form (e.g. "/Type"). Computed lazily on first toBytes(). */
+  private cachedBytes: Uint8Array | null = null;
+
   private constructor(readonly value: string) {}
 
   /**
@@ -124,7 +151,17 @@ export class PdfName implements PdfPrimitive {
   }
 
   toBytes(writer: ByteWriter): void {
-    writer.writeAscii(`/${escapeName(this.value)}`);
+    let bytes = this.cachedBytes;
+
+    if (bytes === null) {
+      const escaped = escapeName(this.value);
+
+      bytes = textEncoder.encode(`/${escaped}`);
+
+      this.cachedBytes = bytes;
+    }
+
+    writer.writeBytes(bytes);
   }
 
   /**
diff --git a/src/writer/pdf-writer.ts b/src/writer/pdf-writer.ts
index 27469df..371a073 100644
--- a/src/writer/pdf-writer.ts
+++ b/src/writer/pdf-writer.ts
@@ -54,6 +54,18 @@ export interface WriteOptions {
    */
   compressStreams?: boolean;
 
+  /**
+   * Minimum stream size in bytes to attempt compression (default: 512).
+   *
+   * Streams smaller than this threshold are written uncompressed.
+   * Deflate initialization has a fixed cost (~0.023ms for pako's 64KB
+   * hash table) that dominates for small payloads, and tiny streams
+   * rarely achieve meaningful compression.
+   *
+   * Set to 0 to compress all streams regardless of size.
+   */
+  compressionThreshold?: number;
+
   /**
    * Security handler for encrypting content.
    *
@@ -106,7 +118,13 @@ function writeIndirectObject(writer: ByteWriter, ref: PdfRef, obj: PdfObject): v
  * Streams that already have filters are returned unchanged - this includes
  * image formats (DCTDecode, JPXDecode, etc.) that are already compressed.
  */
-function prepareObjectForWrite(obj: PdfObject, compress: boolean): PdfObject {
+const DEFAULT_COMPRESSION_THRESHOLD = 512;
+
+function prepareObjectForWrite(
+  obj: PdfObject,
+  compress: boolean,
+  compressionThreshold: number,
+): PdfObject {
   // Only process streams
   if (!(obj instanceof PdfStream)) {
     return obj;
@@ -122,8 +140,11 @@ function prepareObjectForWrite(obj: PdfObject, compress: boolean): PdfObject {
     return obj;
   }
 
-  // Empty streams don't need compression
-  if (obj.data.length === 0) {
+  // Pako's deflate initialization zeros a 64KB hash table on every call
+  // (~0.023ms). For streams below the threshold the compression savings
+  // are negligible relative to the init cost, especially when writing
+  // many PDFs (e.g. splitting 2000 pages).
+  if (obj.data.length < compressionThreshold) {
     return obj;
   }
 
@@ -322,6 +343,7 @@ function collectReachableRefs(
 export function writeComplete(registry: ObjectRegistry, options: WriteOptions): WriteResult {
   const writer = new ByteWriter();
   const compress = options.compressStreams ?? true;
+  const threshold = options.compressionThreshold ?? DEFAULT_COMPRESSION_THRESHOLD;
 
   // Version
   const version = options.version ?? "1.7";
@@ -346,7 +368,7 @@ export function writeComplete(registry: ObjectRegistry, options: WriteOptions):
       continue; // Skip orphan objects
     }
     // Prepare object (compress streams if needed)
-    let prepared = prepareObjectForWrite(obj, compress);
+    let prepared = prepareObjectForWrite(obj, compress, threshold);
 
     // Apply encryption if security handler is provided
     // Skip encrypting the /Encrypt dictionary itself
@@ -467,6 +489,7 @@ export function writeIncremental(
   }
 
   const compress = options.compressStreams ?? true;
+  const threshold = options.compressionThreshold ?? DEFAULT_COMPRESSION_THRESHOLD;
 
   // Initialize ByteWriter with original bytes
   const writer = new ByteWriter(options.originalBytes);
@@ -483,7 +506,7 @@ export function writeIncremental(
 
   // Write modified objects
   for (const [ref, obj] of changes.modified) {
-    let prepared = prepareObjectForWrite(obj, compress);
+    let prepared = prepareObjectForWrite(obj, compress, threshold);
 
     // Apply encryption if security handler is provided
     // Skip encrypting the /Encrypt dictionary itself
@@ -505,7 +528,7 @@ export function writeIncremental(
 
   // Write new objects
   for (const [ref, obj] of changes.created) {
-    let prepared = prepareObjectForWrite(obj, compress);
+    let prepared = prepareObjectForWrite(obj, compress, threshold);
 
     // Apply encryption if security handler is provided
     // Skip encrypting the /Encrypt dictionary itself

From 009a6145a6b2c0aa6078f1d418061f59773f0654 Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 13:47:42 +1100
Subject: [PATCH 5/9] ci: add PR benchmark comparison workflow
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Runs splitting benchmarks on both base and PR branches when .ts files
are changed. Posts a comparison table as a sticky PR comment showing
per-benchmark speedup/regression with 🟢/🔴 indicators at ±5% threshold.
---
 .github/workflows/bench-pr.yml |  62 ++++++++++++++
 scripts/bench-compare.ts       | 150 +++++++++++++++++++++++++++++++++
 2 files changed, 212 insertions(+)
 create mode 100644 .github/workflows/bench-pr.yml
 create mode 100644 scripts/bench-compare.ts

diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
new file mode 100644
index 0000000..9321de1
--- /dev/null
+++ b/.github/workflows/bench-pr.yml
@@ -0,0 +1,62 @@
+name: PR Benchmarks
+
+on:
+  pull_request:
+    branches: [main]
+    paths:
+      - "src/**/*.ts"
+      - "benchmarks/**/*.ts"
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+  pull-requests: write
+
+jobs:
+  bench:
+    name: Benchmark Comparison
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout PR
+        uses: actions/checkout@v4
+        with:
+          path: pr
+
+      - name: Checkout base
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.base.sha }}
+          path: base
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Install dependencies (base)
+        run: bun install --frozen-lockfile
+        working-directory: base
+
+      - name: Install dependencies (PR)
+        run: bun install --frozen-lockfile
+        working-directory: pr
+
+      - name: Run benchmarks (base)
+        run: bun run bench -- --outputJson ../base-results.json benchmarks/splitting.bench.ts
+        working-directory: base
+        continue-on-error: true
+
+      - name: Run benchmarks (PR)
+        run: bun run bench -- --outputJson ../pr-results.json benchmarks/splitting.bench.ts
+        working-directory: pr
+
+      - name: Generate comparison
+        run: bun run pr/scripts/bench-compare.ts ../base-results.json ../pr-results.json bench-comment.md
+        working-directory: pr
+
+      - name: Post or update PR comment
+        uses: marocchino/sticky-pull-request-comment@v2
+        with:
+          header: benchmark-results
+          path: pr/bench-comment.md
diff --git a/scripts/bench-compare.ts b/scripts/bench-compare.ts
new file mode 100644
index 0000000..2be007e
--- /dev/null
+++ b/scripts/bench-compare.ts
@@ -0,0 +1,150 @@
+/**
+ * Compare two benchmark JSON files and produce a markdown summary.
+ *
+ * Usage:
+ *   bun run scripts/bench-compare.ts <base.json> <pr.json> <output.md>
+ *
+ * If <base.json> does not exist, outputs PR-only results.
+ */
+
+import { existsSync, readFileSync } from "node:fs";
+
+interface Bench {
+  name: string;
+  mean: number;
+  hz: number;
+  p99: number;
+  rme: number;
+  sampleCount: number;
+}
+
+interface Group {
+  fullName: string;
+  benchmarks: Bench[];
+}
+
+interface File {
+  filepath: string;
+  groups: Group[];
+}
+
+interface Output {
+  files: File[];
+}
+
+function buildMap(data: Output): Map<string, Bench> {
+  const map = new Map<string, Bench>();
+
+  for (const file of data.files) {
+    for (const group of file.groups) {
+      for (const b of group.benchmarks) {
+        map.set(b.name, b);
+      }
+    }
+  }
+
+  return map;
+}
+
+function formatMs(ms: number): string {
+  if (ms >= 1000) {
+    return `${(ms / 1000).toFixed(2)}s`;
+  }
+
+  if (ms >= 1) {
+    return `${ms.toFixed(2)}ms`;
+  }
+
+  return `${(ms * 1000).toFixed(0)}μs`;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+
+const [basePath, prPath, outputPath] = process.argv.slice(2);
+
+if (!prPath || !outputPath) {
+  console.error("Usage: bun run scripts/bench-compare.ts <base.json> <pr.json> <output.md>");
+  process.exit(1);
+}
+
+const pr: Output = JSON.parse(readFileSync(prPath, "utf-8"));
+const prMap = buildMap(pr);
+const hasBase = existsSync(basePath);
+
+const lines: string[] = [];
+lines.push("## Benchmark Results");
+lines.push("");
+
+if (hasBase) {
+  const base: Output = JSON.parse(readFileSync(basePath, "utf-8"));
+  const baseMap = buildMap(base);
+
+  lines.push("| Benchmark | Base | PR | Change |");
+  lines.push("|:---|---:|---:|---:|");
+
+  let hasRegression = false;
+
+  for (const [name, b] of baseMap) {
+    const p = prMap.get(name);
+
+    if (!p) {
+      continue;
+    }
+
+    const ratio = b.mean / p.mean;
+    let change: string;
+    let indicator = "";
+
+    if (ratio > 1.05) {
+      change = `${ratio.toFixed(2)}x faster`;
+      indicator = " 🟢";
+    } else if (ratio < 0.95) {
+      change = `${(1 / ratio).toFixed(2)}x slower`;
+      indicator = " 🔴";
+      hasRegression = true;
+    } else {
+      change = "~same";
+    }
+
+    lines.push(`| ${name} | ${formatMs(b.mean)} | ${formatMs(p.mean)} | ${change}${indicator} |`);
+  }
+
+  // Show benchmarks only in PR (new benchmarks)
+  for (const [name, p] of prMap) {
+    if (!baseMap.has(name)) {
+      lines.push(`| ${name} | — | ${formatMs(p.mean)} | *new* |`);
+    }
+  }
+
+  lines.push("");
+
+  if (hasRegression) {
+    lines.push("> ⚠️ **Performance regression detected.** Please review the changes above.");
+  } else {
+    lines.push("> ✅ No performance regressions detected.");
+  }
+} else {
+  // No base results — just show PR numbers
+  lines.push("*No base benchmarks available for comparison (new benchmark suite?).*");
+  lines.push("");
+  lines.push("| Benchmark | Mean | p99 | Samples |");
+  lines.push("|:---|---:|---:|---:|");
+
+  for (const [name, p] of prMap) {
+    lines.push(`| ${name} | ${formatMs(p.mean)} | ${formatMs(p.p99)} | ${p.sampleCount} |`);
+  }
+}
+
+lines.push("");
+lines.push(
+  `<details><summary>Environment</summary>\n\n` +
+    `- Runner: \`ubuntu-latest\`\n` +
+    `- Runtime: Bun ${process.versions.bun}\n` +
+    `- Benchmark: \`benchmarks/splitting.bench.ts\`\n\n` +
+    `*Results are machine-dependent. Thresholds: >5% faster 🟢, >5% slower 🔴.*\n` +
+    `</details>`,
+);
+
+const body = lines.join("\n");
+await Bun.write(outputPath, body);
+console.log(body);

From ca39f3ec49dc096b5da76ffeecafa681d49d010f Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 13:54:59 +1100
Subject: [PATCH 6/9] chore: tidy baselines remove local report

---
 reports/baselines/baseline.json               | 224 -----------------
 reports/baselines/comparison-npm-lru-cache.md |  39 ---
 reports/baselines/comparison-sync-copier.md   |  28 ---
 reports/baselines/escape-name-opt.json        | 224 -----------------
 reports/baselines/npm-lru-cache.json          | 224 -----------------
 reports/baselines/skip-tiny-deflate.json      | 224 -----------------
 reports/baselines/sync-copier.json            | 224 -----------------
 reports/benchmarks.md                         | 233 +++++++++---------
 8 files changed, 121 insertions(+), 1299 deletions(-)
 delete mode 100644 reports/baselines/baseline.json
 delete mode 100644 reports/baselines/comparison-npm-lru-cache.md
 delete mode 100644 reports/baselines/comparison-sync-copier.md
 delete mode 100644 reports/baselines/escape-name-opt.json
 delete mode 100644 reports/baselines/npm-lru-cache.json
 delete mode 100644 reports/baselines/skip-tiny-deflate.json
 delete mode 100644 reports/baselines/sync-copier.json

diff --git a/reports/baselines/baseline.json b/reports/baselines/baseline.json
deleted file mode 100644
index bae2d4c..0000000
--- a/reports/baselines/baseline.json
+++ /dev/null
@@ -1,224 +0,0 @@
-{
-  "files": [
-    {
-      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
-      "groups": [
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
-          "benchmarks": [
-            {
-              "id": "-1072438084_0_0",
-              "name": "extractPages (1 page from small PDF)",
-              "rank": 1,
-              "rme": 1.598586232947195,
-              "samples": [],
-              "totalTime": 500.1328439999987,
-              "min": 0.3573750000000473,
-              "max": 2.4449999999999363,
-              "hz": 2215.411391778147,
-              "period": 0.45138343321299523,
-              "mean": 0.45138343321299523,
-              "variance": 0.015017269876550877,
-              "sd": 0.12254497083336745,
-              "sem": 0.0036815068475241517,
-              "df": 1107,
-              "critical": 1.96,
-              "moe": 0.007215753421147338,
-              "p75": 0.4394999999999527,
-              "p99": 0.8506250000000364,
-              "p995": 0.9532910000000356,
-              "p999": 1.7487909999999829,
-              "sampleCount": 1108,
-              "median": 0.4120829999999387
-            },
-            {
-              "id": "-1072438084_0_1",
-              "name": "extractPages (1 page from 100-page PDF)",
-              "rank": 2,
-              "rme": 1.14326877952979,
-              "samples": [],
-              "totalTime": 500.9256660000017,
-              "min": 1.502375000000029,
-              "max": 2.347500000000082,
-              "hz": 560.9614740722809,
-              "period": 1.782653615658369,
-              "mean": 1.782653615658369,
-              "variance": 0.03038259004746126,
-              "sd": 0.17430602412843127,
-              "sem": 0.010398225630092402,
-              "df": 280,
-              "critical": 1.96,
-              "moe": 0.02038052223498111,
-              "p75": 1.8196669999999813,
-              "p99": 2.319291000000021,
-              "p995": 2.3412920000000668,
-              "p999": 2.347500000000082,
-              "sampleCount": 281,
-              "median": 1.7196670000000722
-            },
-            {
-              "id": "-1072438084_0_2",
-              "name": "extractPages (1 page from 2000-page PDF)",
-              "rank": 3,
-              "rme": 2.6124445200935758,
-              "samples": [],
-              "totalTime": 531.2322489999992,
-              "min": 37.85783300000003,
-              "max": 44.21758299999965,
-              "hz": 24.471405914214404,
-              "period": 40.864019153846094,
-              "mean": 40.864019153846094,
-              "variance": 3.120363595900889,
-              "sd": 1.7664550930892324,
-              "sem": 0.4899264933798263,
-              "df": 12,
-              "critical": 2.179,
-              "moe": 1.0675498290746415,
-              "p75": 42.14979199999971,
-              "p99": 44.21758299999965,
-              "p995": 44.21758299999965,
-              "p999": 44.21758299999965,
-              "sampleCount": 13,
-              "median": 40.22212499999978
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
-          "benchmarks": [
-            {
-              "id": "-1072438084_1_0",
-              "name": "split 100-page PDF (0.1MB)",
-              "rank": 1,
-              "rme": 1.8856591767197517,
-              "samples": [],
-              "totalTime": 506.2218300000013,
-              "min": 30.098917000000256,
-              "max": 34.55699999999979,
-              "hz": 31.60669700870063,
-              "period": 31.638864375000082,
-              "mean": 31.638864375000082,
-              "variance": 1.2540684871776795,
-              "sd": 1.119851993424881,
-              "sem": 0.27996299835622024,
-              "df": 15,
-              "critical": 2.131,
-              "moe": 0.5966011494971053,
-              "p75": 32.05224999999973,
-              "p99": 34.55699999999979,
-              "p995": 34.55699999999979,
-              "p999": 34.55699999999979,
-              "sampleCount": 16,
-              "median": 31.37264550000009
-            },
-            {
-              "id": "-1072438084_1_1",
-              "name": "split 2000-page PDF (0.9MB)",
-              "rank": 2,
-              "rme": 0,
-              "samples": [],
-              "totalTime": 582.5043749999995,
-              "min": 582.5043749999995,
-              "max": 582.5043749999995,
-              "hz": 1.716725303565318,
-              "period": 582.5043749999995,
-              "mean": 582.5043749999995,
-              "variance": 0,
-              "sd": 0,
-              "sem": 0,
-              "df": 0,
-              "critical": 12.71,
-              "moe": 0,
-              "p75": 582.5043749999995,
-              "p99": 582.5043749999995,
-              "p995": 582.5043749999995,
-              "p999": 582.5043749999995,
-              "sampleCount": 1,
-              "median": 582.5043749999995
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
-          "benchmarks": [
-            {
-              "id": "-1072438084_2_0",
-              "name": "extract first 10 pages from 2000-page PDF",
-              "rank": 1,
-              "rme": 3.094676818514817,
-              "samples": [],
-              "totalTime": 514.4472479999995,
-              "min": 39.471832999999606,
-              "max": 46.71299999999974,
-              "hz": 23.326006790107296,
-              "period": 42.87060399999996,
-              "mean": 42.87060399999996,
-              "variance": 4.360044973374979,
-              "sd": 2.0880720709245115,
-              "sem": 0.602774486117803,
-              "df": 11,
-              "critical": 2.201,
-              "moe": 1.3267066439452846,
-              "p75": 43.80012500000066,
-              "p99": 46.71299999999974,
-              "p995": 46.71299999999974,
-              "p999": 46.71299999999974,
-              "sampleCount": 12,
-              "median": 42.56633349999993
-            },
-            {
-              "id": "-1072438084_2_1",
-              "name": "extract first 100 pages from 2000-page PDF",
-              "rank": 2,
-              "rme": 3.2071455847378436,
-              "samples": [],
-              "totalTime": 509.16504200000236,
-              "min": 48.13025000000016,
-              "max": 54.86754199999996,
-              "hz": 19.639997201535987,
-              "period": 50.91650420000023,
-              "mean": 50.91650420000023,
-              "variance": 5.211578755405502,
-              "sd": 2.282888248558282,
-              "sem": 0.7219126509076774,
-              "df": 9,
-              "critical": 2.262,
-              "moe": 1.6329664163531663,
-              "p75": 52.17775000000074,
-              "p99": 54.86754199999996,
-              "p995": 54.86754199999996,
-              "p999": 54.86754199999996,
-              "sampleCount": 10,
-              "median": 51.348021000000244
-            },
-            {
-              "id": "-1072438084_2_2",
-              "name": "extract every 10th page from 2000-page PDF (200 pages)",
-              "rank": 3,
-              "rme": 4.394613298511672,
-              "samples": [],
-              "totalTime": 529.0494579999995,
-              "min": 54.37287500000002,
-              "max": 64.11266699999942,
-              "hz": 17.011642038200534,
-              "period": 58.78327311111106,
-              "mean": 58.78327311111106,
-              "variance": 11.294649634264848,
-              "sd": 3.360751349663472,
-              "sem": 1.120250449887824,
-              "df": 8,
-              "critical": 2.306,
-              "moe": 2.5832975374413225,
-              "p75": 60.755959000000075,
-              "p99": 64.11266699999942,
-              "p995": 64.11266699999942,
-              "p999": 64.11266699999942,
-              "sampleCount": 9,
-              "median": 57.71983300000011
-            }
-          ]
-        }
-      ]
-    }
-  ]
-}
diff --git a/reports/baselines/comparison-npm-lru-cache.md b/reports/baselines/comparison-npm-lru-cache.md
deleted file mode 100644
index 6bba72c..0000000
--- a/reports/baselines/comparison-npm-lru-cache.md
+++ /dev/null
@@ -1,39 +0,0 @@
-# Benchmark Comparison: npm lru-cache
-
-## Change
-
-Replaced internal LRU cache (Map delete+set on every get) with the npm `lru-cache`
-package (v11, doubly-linked-list, O(1) get/set without Map rehashing).
-
-## Results (vs original baseline)
-
-| Benchmark                                              | Baseline (ms) | Sync Copier (ms) | + npm lru-cache (ms) | Total Speedup |
-| :----------------------------------------------------- | ------------: | ---------------: | -------------------: | ------------: |
-| extractPages (1 page from small PDF)                   |          0.45 |             0.40 |                 0.38 |         1.17x |
-| extractPages (1 page from 100-page PDF)                |          1.78 |             1.74 |                 1.66 |         1.07x |
-| extractPages (1 page from 2000-page PDF)               |         40.86 |            41.04 |                25.50 |         1.60x |
-| split 100-page PDF (0.1MB)                             |         31.64 |            27.35 |                24.59 |         1.29x |
-| split 2000-page PDF (0.9MB)                            |        582.50 |           506.61 |               432.27 |         1.35x |
-| extract first 10 pages from 2000-page PDF              |         42.87 |            43.21 |                27.31 |         1.57x |
-| extract first 100 pages from 2000-page PDF             |         50.92 |            53.87 |                35.21 |         1.45x |
-| extract every 10th page from 2000-page PDF (200 pages) |         58.78 |            56.65 |                40.52 |         1.45x |
-
-## Incremental improvement (sync copier -> + npm lru-cache)
-
-| Benchmark                                              | Sync Copier (ms) | + npm lru-cache (ms) | Incremental Speedup |
-| :----------------------------------------------------- | ---------------: | -------------------: | ------------------: |
-| extractPages (1 page from small PDF)                   |             0.40 |                 0.38 |               1.03x |
-| extractPages (1 page from 100-page PDF)                |             1.74 |                 1.66 |               1.05x |
-| extractPages (1 page from 2000-page PDF)               |            41.04 |                25.50 |               1.61x |
-| split 100-page PDF (0.1MB)                             |            27.35 |                24.59 |               1.11x |
-| split 2000-page PDF (0.9MB)                            |           506.61 |               432.27 |               1.17x |
-| extract first 10 pages from 2000-page PDF              |            43.21 |                27.31 |               1.58x |
-| extract first 100 pages from 2000-page PDF             |            53.87 |                35.21 |               1.53x |
-| extract every 10th page from 2000-page PDF (200 pages) |            56.65 |                40.52 |               1.40x |
-
-## Key Takeaways
-
-- **2000-page split**: 582.5ms -> 432.3ms (1.35x faster total)
-- **100-page split**: 31.6ms -> 24.6ms (1.29x faster total)
-- **Single page from 2000p**: 40.9ms -> 25.5ms (1.60x faster total)
-- **npm lru-cache incremental gain on 2000-page split**: 506.6ms -> 432.3ms (1.17x)
diff --git a/reports/baselines/comparison-sync-copier.md b/reports/baselines/comparison-sync-copier.md
deleted file mode 100644
index 8752df2..0000000
--- a/reports/baselines/comparison-sync-copier.md
+++ /dev/null
@@ -1,28 +0,0 @@
-# Benchmark Comparison: Sync ObjectCopier
-
-## Change
-
-Removed async/await from all ObjectCopier methods. The copier does zero I/O —
-every method was async but never awaited anything asynchronous. Removing the
-async overhead eliminates microtask scheduling on every recursive call in the
-deep-copy graph walk.
-
-## Results
-
-| Benchmark                                              | Baseline (ms) | Sync (ms) | Speedup |
-| :----------------------------------------------------- | ------------: | --------: | ------: |
-| extractPages (1 page from small PDF)                   |          0.45 |      0.40 |   1.14x |
-| extractPages (1 page from 100-page PDF)                |          1.78 |      1.74 |   1.03x |
-| extractPages (1 page from 2000-page PDF)               |         40.86 |     41.04 |   1.00x |
-| split 100-page PDF (0.1MB)                             |         31.64 |     27.35 |   1.16x |
-| split 2000-page PDF (0.9MB)                            |        582.50 |    506.61 |   1.15x |
-| extract first 10 pages from 2000-page PDF              |         42.87 |     43.21 |   0.99x |
-| extract first 100 pages from 2000-page PDF             |         50.92 |     53.87 |   0.95x |
-| extract every 10th page from 2000-page PDF (200 pages) |         58.78 |     56.65 |   1.04x |
-
-## Key Takeaways
-
-- **100-page split**: 31.6ms -> 27.3ms (1.16x faster)
-- **2000-page split**: 582.5ms -> 506.6ms (1.15x faster)
-- Single-page extraction from small PDFs: ~14% faster (0.45ms -> 0.40ms)
-- Batch extraction noise is within margin of error for single runs
diff --git a/reports/baselines/escape-name-opt.json b/reports/baselines/escape-name-opt.json
deleted file mode 100644
index 55cfe9e..0000000
--- a/reports/baselines/escape-name-opt.json
+++ /dev/null
@@ -1,224 +0,0 @@
-{
-  "files": [
-    {
-      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
-      "groups": [
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
-          "benchmarks": [
-            {
-              "id": "-1072438084_0_0",
-              "name": "extractPages (1 page from small PDF)",
-              "rank": 1,
-              "rme": 1.3598162943493441,
-              "samples": [],
-              "totalTime": 500.19283100000087,
-              "min": 0.3138329999999314,
-              "max": 2.999708000000055,
-              "hz": 2802.9190206446556,
-              "period": 0.35677092082739004,
-              "mean": 0.35677092082739004,
-              "variance": 0.008589645712391514,
-              "sd": 0.09268034156384791,
-              "sem": 0.002475218936179106,
-              "df": 1401,
-              "critical": 1.96,
-              "moe": 0.004851429114911048,
-              "p75": 0.3457499999999527,
-              "p99": 0.5737500000000182,
-              "p995": 0.6079170000000431,
-              "p999": 0.9181670000000395,
-              "sampleCount": 1402,
-              "median": 0.333729500000004
-            },
-            {
-              "id": "-1072438084_0_1",
-              "name": "extractPages (1 page from 100-page PDF)",
-              "rank": 2,
-              "rme": 0.5969408576151285,
-              "samples": [],
-              "totalTime": 500.4518780000001,
-              "min": 1.4726660000001175,
-              "max": 2.0149169999999685,
-              "hz": 633.4275360637171,
-              "period": 1.5787125488958993,
-              "mean": 1.5787125488958993,
-              "variance": 0.007328512823299263,
-              "sd": 0.0856067335161158,
-              "sem": 0.0048081531778861415,
-              "df": 316,
-              "critical": 1.96,
-              "moe": 0.009423980228656837,
-              "p75": 1.589292000000114,
-              "p99": 1.8592499999999745,
-              "p995": 1.8713330000000497,
-              "p999": 2.0149169999999685,
-              "sampleCount": 317,
-              "median": 1.55074999999988
-            },
-            {
-              "id": "-1072438084_0_2",
-              "name": "extractPages (1 page from 2000-page PDF)",
-              "rank": 3,
-              "rme": 0.8810702753157642,
-              "samples": [],
-              "totalTime": 503.6995829999996,
-              "min": 24.38437500000009,
-              "max": 26.65845900000022,
-              "hz": 39.70620718183127,
-              "period": 25.184979149999982,
-              "mean": 25.184979149999982,
-              "variance": 0.2247997529768304,
-              "sd": 0.47413052314402876,
-              "sem": 0.10601880799575855,
-              "df": 19,
-              "critical": 2.093,
-              "moe": 0.22189736513512265,
-              "p75": 25.31841600000007,
-              "p99": 26.65845900000022,
-              "p995": 26.65845900000022,
-              "p999": 26.65845900000022,
-              "sampleCount": 20,
-              "median": 25.21672899999976
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
-          "benchmarks": [
-            {
-              "id": "-1072438084_1_0",
-              "name": "split 100-page PDF (0.1MB)",
-              "rank": 1,
-              "rme": 1.7048381748393946,
-              "samples": [],
-              "totalTime": 502.2399970000006,
-              "min": 19.57883300000003,
-              "max": 23.024417000000085,
-              "hz": 47.78591936794706,
-              "period": 20.92666654166669,
-              "mean": 20.92666654166669,
-              "variance": 0.7136030401800851,
-              "sd": 0.8447502827345399,
-              "sem": 0.17243392939762045,
-              "df": 23,
-              "critical": 2.069,
-              "moe": 0.3567657999236767,
-              "p75": 21.305459000000155,
-              "p99": 23.024417000000085,
-              "p995": 23.024417000000085,
-              "p999": 23.024417000000085,
-              "sampleCount": 24,
-              "median": 20.864958000000115
-            },
-            {
-              "id": "-1072438084_1_1",
-              "name": "split 2000-page PDF (0.9MB)",
-              "rank": 2,
-              "rme": 0,
-              "samples": [],
-              "totalTime": 369.2785839999997,
-              "min": 369.2785839999997,
-              "max": 369.2785839999997,
-              "hz": 2.7079826540929353,
-              "period": 369.2785839999997,
-              "mean": 369.2785839999997,
-              "variance": 0,
-              "sd": 0,
-              "sem": 0,
-              "df": 0,
-              "critical": 12.71,
-              "moe": 0,
-              "p75": 369.2785839999997,
-              "p99": 369.2785839999997,
-              "p995": 369.2785839999997,
-              "p999": 369.2785839999997,
-              "sampleCount": 1,
-              "median": 369.2785839999997
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
-          "benchmarks": [
-            {
-              "id": "-1072438084_2_0",
-              "name": "extract first 10 pages from 2000-page PDF",
-              "rank": 1,
-              "rme": 2.2077471741058514,
-              "samples": [],
-              "totalTime": 511.5155429999995,
-              "min": 25.185083999999733,
-              "max": 31.580417000000125,
-              "hz": 37.14452133471146,
-              "period": 26.9218706842105,
-              "mean": 26.9218706842105,
-              "variance": 1.5205847790653222,
-              "sd": 1.2331199370155859,
-              "sem": 0.2828971153008515,
-              "df": 18,
-              "critical": 2.101,
-              "moe": 0.594366839247089,
-              "p75": 27.052708000000166,
-              "p99": 31.580417000000125,
-              "p995": 31.580417000000125,
-              "p999": 31.580417000000125,
-              "sampleCount": 19,
-              "median": 26.722999999999956
-            },
-            {
-              "id": "-1072438084_2_1",
-              "name": "extract first 100 pages from 2000-page PDF",
-              "rank": 2,
-              "rme": 1.8050175681822698,
-              "samples": [],
-              "totalTime": 517.9517489999989,
-              "min": 33.36449999999968,
-              "max": 37.865499999999884,
-              "hz": 28.960226563497965,
-              "period": 34.53011659999993,
-              "mean": 34.53011659999993,
-              "variance": 1.2664720864594134,
-              "sd": 1.1253764198966556,
-              "sem": 0.2905709421649515,
-              "df": 14,
-              "critical": 2.145,
-              "moe": 0.6232746709438209,
-              "p75": 35.07374999999956,
-              "p99": 37.865499999999884,
-              "p995": 37.865499999999884,
-              "p999": 37.865499999999884,
-              "sampleCount": 15,
-              "median": 34.1270829999994
-            },
-            {
-              "id": "-1072438084_2_2",
-              "name": "extract every 10th page from 2000-page PDF (200 pages)",
-              "rank": 3,
-              "rme": 3.7908921701889815,
-              "samples": [],
-              "totalTime": 528.9624159999994,
-              "min": 37.919458000000304,
-              "max": 48.4839579999998,
-              "hz": 24.576415274086347,
-              "period": 40.689416615384566,
-              "mean": 40.689416615384566,
-              "variance": 6.5144038865667575,
-              "sd": 2.552333028146358,
-              "sem": 0.707889815772459,
-              "df": 12,
-              "critical": 2.179,
-              "moe": 1.542491908568188,
-              "p75": 40.70625000000018,
-              "p99": 48.4839579999998,
-              "p995": 48.4839579999998,
-              "p999": 48.4839579999998,
-              "sampleCount": 13,
-              "median": 40.056916999999885
-            }
-          ]
-        }
-      ]
-    }
-  ]
-}
diff --git a/reports/baselines/npm-lru-cache.json b/reports/baselines/npm-lru-cache.json
deleted file mode 100644
index 346c580..0000000
--- a/reports/baselines/npm-lru-cache.json
+++ /dev/null
@@ -1,224 +0,0 @@
-{
-  "files": [
-    {
-      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
-      "groups": [
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
-          "benchmarks": [
-            {
-              "id": "-1072438084_0_0",
-              "name": "extractPages (1 page from small PDF)",
-              "rank": 1,
-              "rme": 0.6803617904005139,
-              "samples": [],
-              "totalTime": 500.2104360000012,
-              "min": 0.31749999999988177,
-              "max": 0.7807909999999083,
-              "hz": 2598.9061931526694,
-              "period": 0.3847772584615394,
-              "mean": 0.3847772584615394,
-              "variance": 0.002319155302547339,
-              "sd": 0.04815760897872048,
-              "sem": 0.0013356517575117053,
-              "df": 1299,
-              "critical": 1.96,
-              "moe": 0.0026178774447229423,
-              "p75": 0.38512500000001637,
-              "p99": 0.615833000000066,
-              "p995": 0.6623329999999896,
-              "p999": 0.7236249999998563,
-              "sampleCount": 1300,
-              "median": 0.37283300000012787
-            },
-            {
-              "id": "-1072438084_0_1",
-              "name": "extractPages (1 page from 100-page PDF)",
-              "rank": 2,
-              "rme": 0.787155411012552,
-              "samples": [],
-              "totalTime": 500.06587099999865,
-              "min": 1.5184580000000096,
-              "max": 2.1133749999999054,
-              "hz": 601.92070176291,
-              "period": 1.6613484086378694,
-              "mean": 1.6613484086378694,
-              "variance": 0.013399752069859515,
-              "sd": 0.11575729812784814,
-              "sem": 0.006672139742022405,
-              "df": 300,
-              "critical": 1.96,
-              "moe": 0.013077393894363914,
-              "p75": 1.7012500000000728,
-              "p99": 2.0060419999999795,
-              "p995": 2.059125000000222,
-              "p999": 2.1133749999999054,
-              "sampleCount": 301,
-              "median": 1.6227090000002136
-            },
-            {
-              "id": "-1072438084_0_2",
-              "name": "extractPages (1 page from 2000-page PDF)",
-              "rank": 3,
-              "rme": 1.172293974657617,
-              "samples": [],
-              "totalTime": 509.9992109999989,
-              "min": 24.20520899999974,
-              "max": 26.485957999999755,
-              "hz": 39.215746943577216,
-              "period": 25.499960549999948,
-              "mean": 25.499960549999948,
-              "variance": 0.40798446042404196,
-              "sd": 0.6387366127161038,
-              "sem": 0.14282584857511646,
-              "df": 19,
-              "critical": 2.093,
-              "moe": 0.29893450106771874,
-              "p75": 26.051500000000033,
-              "p99": 26.485957999999755,
-              "p995": 26.485957999999755,
-              "p999": 26.485957999999755,
-              "sampleCount": 20,
-              "median": 25.388771000000133
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
-          "benchmarks": [
-            {
-              "id": "-1072438084_1_0",
-              "name": "split 100-page PDF (0.1MB)",
-              "rank": 1,
-              "rme": 1.5707569902861633,
-              "samples": [],
-              "totalTime": 516.3171660000003,
-              "min": 23.525666999999885,
-              "max": 26.712125000000015,
-              "hz": 40.672674439028796,
-              "period": 24.586531714285726,
-              "mean": 24.586531714285726,
-              "variance": 0.7197857225650164,
-              "sd": 0.8484018638387214,
-              "sem": 0.18513646479916948,
-              "df": 20,
-              "critical": 2.086,
-              "moe": 0.3861946655710675,
-              "p75": 24.975667000000158,
-              "p99": 26.712125000000015,
-              "p995": 26.712125000000015,
-              "p999": 26.712125000000015,
-              "sampleCount": 21,
-              "median": 24.294166000000132
-            },
-            {
-              "id": "-1072438084_1_1",
-              "name": "split 2000-page PDF (0.9MB)",
-              "rank": 2,
-              "rme": 0,
-              "samples": [],
-              "totalTime": 432.27162499999986,
-              "min": 432.27162499999986,
-              "max": 432.27162499999986,
-              "hz": 2.313360262774593,
-              "period": 432.27162499999986,
-              "mean": 432.27162499999986,
-              "variance": 0,
-              "sd": 0,
-              "sem": 0,
-              "df": 0,
-              "critical": 12.71,
-              "moe": 0,
-              "p75": 432.27162499999986,
-              "p99": 432.27162499999986,
-              "p995": 432.27162499999986,
-              "p999": 432.27162499999986,
-              "sampleCount": 1,
-              "median": 432.27162499999986
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
-          "benchmarks": [
-            {
-              "id": "-1072438084_2_0",
-              "name": "extract first 10 pages from 2000-page PDF",
-              "rank": 1,
-              "rme": 2.223598050281543,
-              "samples": [],
-              "totalTime": 518.8923730000015,
-              "min": 26.314790999999786,
-              "max": 30.981832999999824,
-              "hz": 36.61645649202854,
-              "period": 27.31012489473692,
-              "mean": 27.31012489473692,
-              "variance": 1.5873088504924244,
-              "sd": 1.2598844591836287,
-              "sem": 0.28903731779572794,
-              "df": 18,
-              "critical": 2.101,
-              "moe": 0.6072674046888245,
-              "p75": 27.280499999999847,
-              "p99": 30.981832999999824,
-              "p995": 30.981832999999824,
-              "p999": 30.981832999999824,
-              "sampleCount": 19,
-              "median": 26.98291700000027
-            },
-            {
-              "id": "-1072438084_2_1",
-              "name": "extract first 100 pages from 2000-page PDF",
-              "rank": 2,
-              "rme": 2.213921234147061,
-              "samples": [],
-              "totalTime": 528.2034549999998,
-              "min": 34.1123749999997,
-              "max": 39.86729100000048,
-              "hz": 28.398148209765125,
-              "period": 35.21356366666666,
-              "mean": 35.21356366666666,
-              "variance": 1.9814401301546256,
-              "sd": 1.4076363628986805,
-              "sem": 0.3634501460681713,
-              "df": 14,
-              "critical": 2.145,
-              "moe": 0.7796005633162274,
-              "p75": 35.0894159999998,
-              "p99": 39.86729100000048,
-              "p995": 39.86729100000048,
-              "p999": 39.86729100000048,
-              "sampleCount": 15,
-              "median": 34.765957999999955
-            },
-            {
-              "id": "-1072438084_2_2",
-              "name": "extract every 10th page from 2000-page PDF (200 pages)",
-              "rank": 3,
-              "rme": 2.0448631675208837,
-              "samples": [],
-              "totalTime": 526.8241219999973,
-              "min": 38.62562499999967,
-              "max": 42.899374999999964,
-              "hz": 24.67616697323527,
-              "period": 40.52493246153825,
-              "mean": 40.52493246153825,
-              "variance": 1.8801911150990929,
-              "sd": 1.3712006108148773,
-              "sem": 0.38030262394158326,
-              "df": 12,
-              "critical": 2.179,
-              "moe": 0.8286794175687099,
-              "p75": 41.67979099999957,
-              "p99": 42.899374999999964,
-              "p995": 42.899374999999964,
-              "p999": 42.899374999999964,
-              "sampleCount": 13,
-              "median": 40.114832999999635
-            }
-          ]
-        }
-      ]
-    }
-  ]
-}
diff --git a/reports/baselines/skip-tiny-deflate.json b/reports/baselines/skip-tiny-deflate.json
deleted file mode 100644
index d11d64e..0000000
--- a/reports/baselines/skip-tiny-deflate.json
+++ /dev/null
@@ -1,224 +0,0 @@
-{
-  "files": [
-    {
-      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
-      "groups": [
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
-          "benchmarks": [
-            {
-              "id": "-1072438084_0_0",
-              "name": "extractPages (1 page from small PDF)",
-              "rank": 1,
-              "rme": 0.8892829061609049,
-              "samples": [],
-              "totalTime": 500.31841499999973,
-              "min": 0.2905420000001868,
-              "max": 1.9279579999999896,
-              "hz": 2924.1378213112557,
-              "period": 0.3419811449077237,
-              "mean": 0.3419811449077237,
-              "variance": 0.0035222193288500063,
-              "sd": 0.05934828833968176,
-              "sem": 0.001551622379570276,
-              "df": 1462,
-              "critical": 1.96,
-              "moe": 0.0030411798639577406,
-              "p75": 0.3381249999999909,
-              "p99": 0.5579999999999927,
-              "p995": 0.5912919999998394,
-              "p999": 0.6483749999999873,
-              "sampleCount": 1463,
-              "median": 0.3292500000002292
-            },
-            {
-              "id": "-1072438084_0_1",
-              "name": "extractPages (1 page from 100-page PDF)",
-              "rank": 2,
-              "rme": 0.6942473447539961,
-              "samples": [],
-              "totalTime": 500.1834050000034,
-              "min": 1.3748750000004293,
-              "max": 1.868916999999783,
-              "hz": 669.7543274151564,
-              "period": 1.4930847910447862,
-              "mean": 1.4930847910447862,
-              "variance": 0.009369794423687357,
-              "sd": 0.09679769844209808,
-              "sem": 0.005288623222833765,
-              "df": 334,
-              "critical": 1.96,
-              "moe": 0.010365701516754179,
-              "p75": 1.5309159999997064,
-              "p99": 1.8027499999998327,
-              "p995": 1.8117079999997259,
-              "p999": 1.868916999999783,
-              "sampleCount": 335,
-              "median": 1.4698750000002292
-            },
-            {
-              "id": "-1072438084_0_2",
-              "name": "extractPages (1 page from 2000-page PDF)",
-              "rank": 3,
-              "rme": 1.1888586380169315,
-              "samples": [],
-              "totalTime": 501.3285840000003,
-              "min": 24.17533400000002,
-              "max": 27.005499999999756,
-              "hz": 39.89399495321812,
-              "period": 25.066429200000016,
-              "mean": 25.066429200000016,
-              "variance": 0.4054496384268641,
-              "sd": 0.6367492743826758,
-              "sem": 0.14238146621433284,
-              "df": 19,
-              "critical": 2.093,
-              "moe": 0.2980044087865986,
-              "p75": 25.286291999999776,
-              "p99": 27.005499999999756,
-              "p995": 27.005499999999756,
-              "p999": 27.005499999999756,
-              "sampleCount": 20,
-              "median": 24.951479000000063
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
-          "benchmarks": [
-            {
-              "id": "-1072438084_1_0",
-              "name": "split 100-page PDF (0.1MB)",
-              "rank": 1,
-              "rme": 1.241299549103146,
-              "samples": [],
-              "totalTime": 509.5840829999979,
-              "min": 12.014166999999816,
-              "max": 13.951500000000124,
-              "hz": 78.49538738438218,
-              "period": 12.739602074999947,
-              "mean": 12.739602074999947,
-              "variance": 0.2443938158093934,
-              "sd": 0.4943620290934503,
-              "sem": 0.07816550003188641,
-              "df": 39,
-              "critical": 2.0231,
-              "moe": 0.15813662311450938,
-              "p75": 12.96362500000032,
-              "p99": 13.951500000000124,
-              "p995": 13.951500000000124,
-              "p999": 13.951500000000124,
-              "sampleCount": 40,
-              "median": 12.703645499999766
-            },
-            {
-              "id": "-1072438084_1_1",
-              "name": "split 2000-page PDF (0.9MB)",
-              "rank": 2,
-              "rme": 0,
-              "samples": [],
-              "totalTime": 244.68245800000022,
-              "min": 244.68245800000022,
-              "max": 244.68245800000022,
-              "hz": 4.086929680917294,
-              "period": 244.68245800000022,
-              "mean": 244.68245800000022,
-              "variance": 0,
-              "sd": 0,
-              "sem": 0,
-              "df": 0,
-              "critical": 12.71,
-              "moe": 0,
-              "p75": 244.68245800000022,
-              "p99": 244.68245800000022,
-              "p995": 244.68245800000022,
-              "p999": 244.68245800000022,
-              "sampleCount": 1,
-              "median": 244.68245800000022
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
-          "benchmarks": [
-            {
-              "id": "-1072438084_2_0",
-              "name": "extract first 10 pages from 2000-page PDF",
-              "rank": 1,
-              "rme": 0.9557725912034225,
-              "samples": [],
-              "totalTime": 504.57903800000076,
-              "min": 24.360125000000153,
-              "max": 25.940999999999804,
-              "hz": 39.637001329413074,
-              "period": 25.228951900000038,
-              "mean": 25.228951900000038,
-              "variance": 0.2654600064528459,
-              "sd": 0.5152281110856102,
-              "sem": 0.11520850803062373,
-              "df": 19,
-              "critical": 2.093,
-              "moe": 0.24113140730809546,
-              "p75": 25.62262499999997,
-              "p99": 25.940999999999804,
-              "p995": 25.940999999999804,
-              "p999": 25.940999999999804,
-              "sampleCount": 20,
-              "median": 25.240332999999737
-            },
-            {
-              "id": "-1072438084_2_1",
-              "name": "extract first 100 pages from 2000-page PDF",
-              "rank": 2,
-              "rme": 0.8129752528372926,
-              "samples": [],
-              "totalTime": 509.13470899999993,
-              "min": 25.983500000000276,
-              "max": 28.018167000000176,
-              "hz": 37.31821787856149,
-              "period": 26.796563631578945,
-              "mean": 26.796563631578945,
-              "variance": 0.20427459388908475,
-              "sd": 0.45196746994566406,
-              "sem": 0.10368844880320557,
-              "df": 18,
-              "critical": 2.101,
-              "moe": 0.2178494309355349,
-              "p75": 27.132666999999856,
-              "p99": 28.018167000000176,
-              "p995": 28.018167000000176,
-              "p999": 28.018167000000176,
-              "sampleCount": 19,
-              "median": 26.71374999999989
-            },
-            {
-              "id": "-1072438084_2_2",
-              "name": "extract every 10th page from 2000-page PDF (200 pages)",
-              "rank": 3,
-              "rme": 1.7260235531528858,
-              "samples": [],
-              "totalTime": 526.418208000001,
-              "min": 28.0747080000001,
-              "max": 31.5014170000004,
-              "hz": 34.19334613896935,
-              "period": 29.245456000000054,
-              "mean": 29.245456000000054,
-              "variance": 1.0301911737761424,
-              "sd": 1.014983336698757,
-              "sem": 0.23923386672368,
-              "df": 17,
-              "critical": 2.11,
-              "moe": 0.5047834587869647,
-              "p75": 29.7938750000003,
-              "p99": 31.5014170000004,
-              "p995": 31.5014170000004,
-              "p999": 31.5014170000004,
-              "sampleCount": 18,
-              "median": 28.999896000000717
-            }
-          ]
-        }
-      ]
-    }
-  ]
-}
diff --git a/reports/baselines/sync-copier.json b/reports/baselines/sync-copier.json
deleted file mode 100644
index 7b6bd2f..0000000
--- a/reports/baselines/sync-copier.json
+++ /dev/null
@@ -1,224 +0,0 @@
-{
-  "files": [
-    {
-      "filepath": "/Users/lucas/dev/libpdf/benchmarks/splitting.bench.ts",
-      "groups": [
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Extract single page",
-          "benchmarks": [
-            {
-              "id": "-1072438084_0_0",
-              "name": "extractPages (1 page from small PDF)",
-              "rank": 1,
-              "rme": 0.8797965308418093,
-              "samples": [],
-              "totalTime": 500.17104999999685,
-              "min": 0.3400409999999283,
-              "max": 0.8508749999998599,
-              "hz": 2517.1388867868463,
-              "period": 0.39727644956314284,
-              "mean": 0.39727644956314284,
-              "variance": 0.004003720204126598,
-              "sd": 0.0632749571641625,
-              "sem": 0.0017832777658714494,
-              "df": 1258,
-              "critical": 1.96,
-              "moe": 0.0034952244211080407,
-              "p75": 0.3927090000001954,
-              "p99": 0.74350000000004,
-              "p995": 0.7761249999998654,
-              "p999": 0.8387500000000045,
-              "sampleCount": 1259,
-              "median": 0.38120900000012625
-            },
-            {
-              "id": "-1072438084_0_1",
-              "name": "extractPages (1 page from 100-page PDF)",
-              "rank": 2,
-              "rme": 0.9765406302412275,
-              "samples": [],
-              "totalTime": 500.3271310000009,
-              "min": 1.584292000000005,
-              "max": 2.372624999999971,
-              "hz": 575.6233914886368,
-              "period": 1.737246982638892,
-              "mean": 1.737246982638892,
-              "variance": 0.021576654723250537,
-              "sd": 0.14688994085113705,
-              "sem": 0.008655572771994156,
-              "df": 287,
-              "critical": 1.96,
-              "moe": 0.016964922633108545,
-              "p75": 1.756249999999909,
-              "p99": 2.2702919999999267,
-              "p995": 2.3456670000000486,
-              "p999": 2.372624999999971,
-              "sampleCount": 288,
-              "median": 1.6839165000000094
-            },
-            {
-              "id": "-1072438084_0_2",
-              "name": "extractPages (1 page from 2000-page PDF)",
-              "rank": 3,
-              "rme": 2.1980090898940152,
-              "samples": [],
-              "totalTime": 533.5146669999999,
-              "min": 38.59916599999997,
-              "max": 43.44508300000007,
-              "hz": 24.366715301568274,
-              "period": 41.039589769230766,
-              "mean": 41.039589769230766,
-              "variance": 2.227890706742653,
-              "sd": 1.4926120416044664,
-              "sem": 0.4139760961830687,
-              "df": 12,
-              "critical": 2.179,
-              "moe": 0.9020539135829065,
-              "p75": 42.321292000000085,
-              "p99": 43.44508300000007,
-              "p995": 43.44508300000007,
-              "p999": 43.44508300000007,
-              "sampleCount": 13,
-              "median": 40.792207999999846
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Split into single-page PDFs",
-          "benchmarks": [
-            {
-              "id": "-1072438084_1_0",
-              "name": "split 100-page PDF (0.1MB)",
-              "rank": 1,
-              "rme": 1.9707618954593826,
-              "samples": [],
-              "totalTime": 519.5589599999998,
-              "min": 25.109875000000102,
-              "max": 29.522958000000017,
-              "hz": 36.569478081948596,
-              "period": 27.34520842105262,
-              "mean": 27.34520842105262,
-              "variance": 1.2500641058887851,
-              "sd": 1.11806265740735,
-              "sem": 0.25650116506237763,
-              "df": 18,
-              "critical": 2.101,
-              "moe": 0.5389089477960554,
-              "p75": 28.18224999999984,
-              "p99": 29.522958000000017,
-              "p995": 29.522958000000017,
-              "p999": 29.522958000000017,
-              "sampleCount": 19,
-              "median": 27.39729099999977
-            },
-            {
-              "id": "-1072438084_1_1",
-              "name": "split 2000-page PDF (0.9MB)",
-              "rank": 2,
-              "rme": 0,
-              "samples": [],
-              "totalTime": 506.605583,
-              "min": 506.605583,
-              "max": 506.605583,
-              "hz": 1.9739221863253724,
-              "period": 506.605583,
-              "mean": 506.605583,
-              "variance": 0,
-              "sd": 0,
-              "sem": 0,
-              "df": 0,
-              "critical": 12.71,
-              "moe": 0,
-              "p75": 506.605583,
-              "p99": 506.605583,
-              "p995": 506.605583,
-              "p999": 506.605583,
-              "sampleCount": 1,
-              "median": 506.605583
-            }
-          ]
-        },
-        {
-          "fullName": "benchmarks/splitting.bench.ts > Batch page extraction",
-          "benchmarks": [
-            {
-              "id": "-1072438084_2_0",
-              "name": "extract first 10 pages from 2000-page PDF",
-              "rank": 1,
-              "rme": 2.7205867827906096,
-              "samples": [],
-              "totalTime": 518.4931240000024,
-              "min": 40.28308300000026,
-              "max": 45.694250000000466,
-              "hz": 23.143990623104088,
-              "period": 43.20776033333353,
-              "mean": 43.20776033333353,
-              "variance": 3.422865540761208,
-              "sd": 1.850098792162518,
-              "sem": 0.5340775178412157,
-              "df": 11,
-              "critical": 2.201,
-              "moe": 1.175504616768516,
-              "p75": 44.60529199999928,
-              "p99": 45.694250000000466,
-              "p995": 45.694250000000466,
-              "p999": 45.694250000000466,
-              "sampleCount": 12,
-              "median": 43.13712450000003
-            },
-            {
-              "id": "-1072438084_2_1",
-              "name": "extract first 100 pages from 2000-page PDF",
-              "rank": 2,
-              "rme": 6.90077055615189,
-              "samples": [],
-              "totalTime": 538.6501679999983,
-              "min": 48.26045899999917,
-              "max": 65.11799999999948,
-              "hz": 18.56492505540262,
-              "period": 53.86501679999983,
-              "mean": 53.86501679999983,
-              "variance": 27.003718600062566,
-              "sd": 5.196510232844978,
-              "sem": 1.6432808220162056,
-              "df": 9,
-              "critical": 2.262,
-              "moe": 3.717101219400657,
-              "p75": 54.15062500000022,
-              "p99": 65.11799999999948,
-              "p995": 65.11799999999948,
-              "p999": 65.11799999999948,
-              "sampleCount": 10,
-              "median": 53.544791499999974
-            },
-            {
-              "id": "-1072438084_2_2",
-              "name": "extract every 10th page from 2000-page PDF (200 pages)",
-              "rank": 3,
-              "rme": 3.996294175601932,
-              "samples": [],
-              "totalTime": 509.82821200000035,
-              "min": 52.99541700000009,
-              "max": 62.211042000000816,
-              "hz": 17.65300504790424,
-              "period": 56.64757911111115,
-              "mean": 56.64757911111115,
-              "variance": 8.673642398208653,
-              "sd": 2.9451048195622263,
-              "sem": 0.9817016065207421,
-              "df": 8,
-              "critical": 2.306,
-              "moe": 2.2638039046368315,
-              "p75": 58.50329199999942,
-              "p99": 62.211042000000816,
-              "p995": 62.211042000000816,
-              "p999": 62.211042000000816,
-              "sampleCount": 9,
-              "median": 56.19641700000011
-            }
-          ]
-        }
-      ]
-    }
-  ]
-}
diff --git a/reports/benchmarks.md b/reports/benchmarks.md
index b0b0010..8a563e5 100644
--- a/reports/benchmarks.md
+++ b/reports/benchmarks.md
@@ -1,8 +1,8 @@
 # Benchmark Report
 
-> Generated on 2026-02-16 at 12:50:00 UTC
+> Generated on 2026-02-16 at 21:19:11 UTC
 >
-> System: darwin | Apple M4 Pro (12 cores) | 24GB RAM | Bun 1.3.5
+> System: linux | AMD EPYC 7763 64-Core Processor (4 cores) | 16GB RAM | Bun 1.3.9
 
 ---
 
@@ -22,91 +22,100 @@
 
 | Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
 | :-------- | ------: | ------: | ------: | -----: | ------: |
-| libpdf    |   895.0 |  1.12ms |  1.59ms | ±1.07% |     448 |
-| pdf-lib   |    36.7 | 27.21ms | 29.03ms | ±1.90% |      19 |
+| libpdf    |   362.6 |  2.76ms |  3.88ms | ±1.62% |     182 |
+| pdf-lib   |    25.0 | 39.95ms | 44.72ms | ±3.76% |      13 |
 
-- **libpdf** is 24.35x faster than pdf-lib
+- **libpdf** is 14.48x faster than pdf-lib
 
 ### Create blank PDF
 
-| Benchmark | ops/sec | Mean |   p99 |    RME | Samples |
-| :-------- | ------: | ---: | ----: | -----: | ------: |
-| libpdf    |   38.6K | 26us |  45us | ±0.84% |  19,283 |
-| pdf-lib   |   10.3K | 97us | 461us | ±1.82% |   5,172 |
+| Benchmark | ops/sec |  Mean |    p99 |    RME | Samples |
+| :-------- | ------: | ----: | -----: | -----: | ------: |
+| libpdf    |   10.7K |  94us |  184us | ±2.96% |   5,333 |
+| pdf-lib   |    2.3K | 437us | 1.65ms | ±2.65% |   1,143 |
 
-- **libpdf** is 3.73x faster than pdf-lib
+- **libpdf** is 4.67x faster than pdf-lib
 
 ### Add 10 pages
 
-| Benchmark | ops/sec |  Mean |   p99 |    RME | Samples |
-| :-------- | ------: | ----: | ----: | -----: | ------: |
-| libpdf    |   19.1K |  52us |  87us | ±0.91% |   9,562 |
-| pdf-lib   |    6.3K | 158us | 770us | ±2.55% |   3,173 |
+| Benchmark | ops/sec |  Mean |    p99 |    RME | Samples |
+| :-------- | ------: | ----: | -----: | -----: | ------: |
+| libpdf    |    5.7K | 174us |  653us | ±1.83% |   2,867 |
+| pdf-lib   |    1.9K | 540us | 2.13ms | ±3.10% |     926 |
 
-- **libpdf** is 3.01x faster than pdf-lib
+- **libpdf** is 3.10x faster than pdf-lib
 
 ### Draw 50 rectangles
 
 | Benchmark | ops/sec |   Mean |    p99 |    RME | Samples |
 | :-------- | ------: | -----: | -----: | -----: | ------: |
-| pdf-lib   |    2.2K |  458us | 1.80ms | ±3.80% |   1,093 |
-| libpdf    |   627.4 | 1.59ms | 2.30ms | ±1.44% |     314 |
+| pdf-lib   |   474.2 | 2.11ms | 7.63ms | ±9.71% |     238 |
+| libpdf    |   156.7 | 6.38ms | 8.63ms | ±2.56% |      79 |
 
-- **pdf-lib** is 3.48x faster than libpdf
+- **pdf-lib** is 3.03x faster than libpdf
 
 ### Load and save PDF
 
-| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
-| :-------- | ------: | ------: | ------: | -----: | ------: |
-| libpdf    |   909.3 |  1.10ms |  1.48ms | ±0.88% |     456 |
-| pdf-lib   |    22.0 | 45.45ms | 58.21ms | ±6.77% |      11 |
+| Benchmark | ops/sec |    Mean |      p99 |    RME | Samples |
+| :-------- | ------: | ------: | -------: | -----: | ------: |
+| libpdf    |   339.6 |  2.94ms |   5.18ms | ±3.06% |     171 |
+| pdf-lib   |    10.8 | 92.96ms | 109.93ms | ±5.74% |      10 |
 
-- **libpdf** is 41.33x faster than pdf-lib
+- **libpdf** is 31.57x faster than pdf-lib
 
 ### Load, modify, and save PDF
 
 | Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
 | :-------- | ------: | ------: | ------: | -----: | ------: |
-| libpdf    |    37.9 | 26.38ms | 33.50ms | ±5.76% |      20 |
-| pdf-lib   |    23.1 | 43.25ms | 44.89ms | ±1.52% |      12 |
+| libpdf    |    13.5 | 73.86ms | 88.68ms | ±8.49% |      10 |
+| pdf-lib   |    11.5 | 87.10ms | 93.25ms | ±3.21% |      10 |
 
-- **libpdf** is 1.64x faster than pdf-lib
+- **libpdf** is 1.18x faster than pdf-lib
 
 ### Extract single page from 100-page PDF
 
-| Benchmark | ops/sec |   Mean |    p99 |    RME | Samples |
-| :-------- | ------: | -----: | -----: | -----: | ------: |
-| libpdf    |   503.4 | 1.99ms | 3.10ms | ±1.55% |     252 |
-| pdf-lib   |   155.9 | 6.41ms | 7.44ms | ±1.53% |      79 |
+| Benchmark | ops/sec |   Mean |     p99 |    RME | Samples |
+| :-------- | ------: | -----: | ------: | -----: | ------: |
+| libpdf    |   183.8 | 5.44ms |  8.10ms | ±2.45% |      92 |
+| pdf-lib   |   104.3 | 9.58ms | 11.64ms | ±1.78% |      53 |
 
-- **libpdf** is 3.23x faster than pdf-lib
+- **libpdf** is 1.76x faster than pdf-lib
 
 ### Split 100-page PDF into single-page PDFs
 
-| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
-| :-------- | ------: | ------: | ------: | -----: | ------: |
-| libpdf    |    35.3 | 28.29ms | 29.79ms | ±1.35% |      18 |
-| pdf-lib   |    35.0 | 28.58ms | 32.30ms | ±3.07% |      18 |
+| Benchmark | ops/sec |    Mean |      p99 |    RME | Samples |
+| :-------- | ------: | ------: | -------: | -----: | ------: |
+| pdf-lib   |    11.1 | 90.04ms |  95.21ms | ±5.07% |       6 |
+| libpdf    |    10.7 | 93.65ms | 104.73ms | ±9.17% |       6 |
 
-- **libpdf** is 1.01x faster than pdf-lib
+- **pdf-lib** is 1.04x faster than libpdf
+
+### Split 2000-page PDF into single-page PDFs (0.9MB)
+
+| Benchmark | ops/sec |  Mean |   p99 |    RME | Samples |
+| :-------- | ------: | ----: | ----: | -----: | ------: |
+| libpdf    |   0.627 | 1.60s | 1.60s | ±0.00% |       1 |
+| pdf-lib   |   0.585 | 1.71s | 1.71s | ±0.00% |       1 |
+
+- **libpdf** is 1.07x faster than pdf-lib
 
 ### Copy 10 pages between documents
 
-| Benchmark | ops/sec |   Mean |     p99 |    RME | Samples |
-| :-------- | ------: | -----: | ------: | -----: | ------: |
-| libpdf    |   334.0 | 2.99ms |  3.55ms | ±1.12% |     168 |
-| pdf-lib   |   103.7 | 9.64ms | 14.95ms | ±3.70% |      52 |
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |   103.7 |  9.64ms | 14.56ms | ±3.04% |      52 |
+| pdf-lib   |    78.9 | 12.67ms | 14.16ms | ±1.65% |      40 |
 
-- **libpdf** is 3.22x faster than pdf-lib
+- **libpdf** is 1.31x faster than pdf-lib
 
 ### Merge 2 x 100-page PDFs
 
 | Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
 | :-------- | ------: | ------: | ------: | -----: | ------: |
-| libpdf    |    47.1 | 21.23ms | 24.31ms | ±2.08% |      24 |
-| pdf-lib   |    22.9 | 43.64ms | 47.48ms | ±2.23% |      12 |
+| pdf-lib   |    17.9 | 55.88ms | 58.47ms | ±1.91% |       9 |
+| libpdf    |    12.6 | 79.05ms | 79.54ms | ±0.45% |       7 |
 
-- **libpdf** is 2.06x faster than pdf-lib
+- **pdf-lib** is 1.41x faster than libpdf
 
 ## Copying
 
@@ -114,88 +123,88 @@
 
 | Benchmark                       | ops/sec |    Mean |     p99 |    RME | Samples |
 | :------------------------------ | ------: | ------: | ------: | -----: | ------: |
-| copy 1 page                     |    2.3K |   429us |   735us | ±1.26% |   1,166 |
-| copy 10 pages from 100-page PDF |   344.1 |  2.91ms |  3.57ms | ±1.12% |     173 |
-| copy all 100 pages              |    92.3 | 10.84ms | 13.86ms | ±1.96% |      47 |
+| copy 1 page                     |   731.3 |  1.37ms |  2.63ms | ±3.01% |     366 |
+| copy 10 pages from 100-page PDF |   113.0 |  8.85ms | 12.36ms | ±2.59% |      57 |
+| copy all 100 pages              |    25.7 | 38.97ms | 41.36ms | ±1.41% |      13 |
 
-- **copy 1 page** is 6.78x faster than copy 10 pages from 100-page PDF
-- **copy 1 page** is 25.27x faster than copy all 100 pages
+- **copy 1 page** is 6.47x faster than copy 10 pages from 100-page PDF
+- **copy 1 page** is 28.50x faster than copy all 100 pages
 
 ### Duplicate pages within same document
 
-| Benchmark                                 | ops/sec |  Mean |   p99 |    RME | Samples |
-| :---------------------------------------- | ------: | ----: | ----: | -----: | ------: |
-| duplicate all pages (double the document) |    2.2K | 461us | 798us | ±0.89% |   1,086 |
-| duplicate page 0                          |    2.2K | 464us | 758us | ±0.77% |   1,078 |
+| Benchmark                                 | ops/sec |   Mean |    p99 |    RME | Samples |
+| :---------------------------------------- | ------: | -----: | -----: | -----: | ------: |
+| duplicate page 0                          |   779.3 | 1.28ms | 2.49ms | ±2.00% |     390 |
+| duplicate all pages (double the document) |   778.8 | 1.28ms | 2.48ms | ±2.14% |     390 |
 
-- **duplicate all pages (double the document)** is 1.01x faster than duplicate page 0
+- **duplicate page 0** is 1.00x faster than duplicate all pages (double the document)
 
 ### Merge PDFs
 
 | Benchmark               | ops/sec |    Mean |     p99 |    RME | Samples |
 | :---------------------- | ------: | ------: | ------: | -----: | ------: |
-| merge 2 small PDFs      |    1.4K |   712us |  1.12ms | ±1.00% |     702 |
-| merge 10 small PDFs     |   254.5 |  3.93ms |  5.73ms | ±2.08% |     128 |
-| merge 2 x 100-page PDFs |    48.6 | 20.58ms | 26.32ms | ±3.33% |      25 |
+| merge 2 small PDFs      |   511.1 |  1.96ms |  3.22ms | ±1.91% |     256 |
+| merge 10 small PDFs     |    93.4 | 10.71ms | 14.59ms | ±2.26% |      47 |
+| merge 2 x 100-page PDFs |    13.0 | 77.00ms | 82.43ms | ±3.35% |       7 |
 
-- **merge 2 small PDFs** is 5.51x faster than merge 10 small PDFs
-- **merge 2 small PDFs** is 28.89x faster than merge 2 x 100-page PDFs
+- **merge 2 small PDFs** is 5.47x faster than merge 10 small PDFs
+- **merge 2 small PDFs** is 39.35x faster than merge 2 x 100-page PDFs
 
 ## Drawing
 
-| Benchmark                           | ops/sec |   Mean |    p99 |    RME | Samples |
-| :---------------------------------- | ------: | -----: | -----: | -----: | ------: |
-| draw 100 lines                      |   399.8 | 2.50ms | 2.74ms | ±0.51% |     200 |
-| draw 100 rectangles                 |   360.3 | 2.78ms | 3.55ms | ±1.22% |     181 |
-| draw 100 circles                    |   279.7 | 3.58ms | 4.42ms | ±1.33% |     140 |
-| draw 100 text lines (standard font) |   259.5 | 3.85ms | 4.29ms | ±0.61% |     130 |
-| create 10 pages with mixed content  |   193.0 | 5.18ms | 6.40ms | ±1.35% |      97 |
+| Benchmark                           | ops/sec |    Mean |     p99 |    RME | Samples |
+| :---------------------------------- | ------: | ------: | ------: | -----: | ------: |
+| draw 100 lines                      |    91.8 | 10.89ms | 13.01ms | ±1.41% |      46 |
+| draw 100 rectangles                 |    79.3 | 12.61ms | 16.38ms | ±3.74% |      40 |
+| draw 100 circles                    |    69.0 | 14.50ms | 18.01ms | ±2.25% |      35 |
+| draw 100 text lines (standard font) |    64.9 | 15.41ms | 20.87ms | ±3.54% |      33 |
+| create 10 pages with mixed content  |    46.8 | 21.39ms | 22.42ms | ±1.20% |      24 |
 
-- **draw 100 lines** is 1.11x faster than draw 100 rectangles
-- **draw 100 lines** is 1.43x faster than draw 100 circles
-- **draw 100 lines** is 1.54x faster than draw 100 text lines (standard font)
-- **draw 100 lines** is 2.07x faster than create 10 pages with mixed content
+- **draw 100 lines** is 1.16x faster than draw 100 rectangles
+- **draw 100 lines** is 1.33x faster than draw 100 circles
+- **draw 100 lines** is 1.41x faster than draw 100 text lines (standard font)
+- **draw 100 lines** is 1.96x faster than create 10 pages with mixed content
 
 ## Forms
 
-| Benchmark         | ops/sec |   Mean |    p99 |    RME | Samples |
-| :---------------- | ------: | -----: | -----: | -----: | ------: |
-| read field values |   702.6 | 1.42ms | 1.85ms | ±0.97% |     352 |
-| get form fields   |   677.3 | 1.48ms | 2.36ms | ±1.45% |     339 |
-| flatten form      |   198.8 | 5.03ms | 5.84ms | ±1.28% |     100 |
-| fill text fields  |   155.1 | 6.45ms | 7.36ms | ±1.25% |      78 |
+| Benchmark         | ops/sec |    Mean |     p99 |    RME | Samples |
+| :---------------- | ------: | ------: | ------: | -----: | ------: |
+| read field values |   286.0 |  3.50ms |  6.12ms | ±2.24% |     143 |
+| get form fields   |   247.2 |  4.05ms |  9.66ms | ±5.08% |     124 |
+| flatten form      |    74.3 | 13.46ms | 17.19ms | ±2.79% |      38 |
+| fill text fields  |    55.3 | 18.07ms | 24.87ms | ±3.84% |      28 |
 
-- **read field values** is 1.04x faster than get form fields
-- **read field values** is 3.53x faster than flatten form
-- **read field values** is 4.53x faster than fill text fields
+- **read field values** is 1.16x faster than get form fields
+- **read field values** is 3.85x faster than flatten form
+- **read field values** is 5.17x faster than fill text fields
 
 ## Loading
 
 | Benchmark              | ops/sec |   Mean |    p99 |    RME | Samples |
 | :--------------------- | ------: | -----: | -----: | -----: | ------: |
-| load small PDF (888B)  |   38.7K |   26us |   37us | ±1.10% |  19,336 |
-| load medium PDF (19KB) |   23.8K |   42us |   54us | ±0.84% |  11,904 |
-| load form PDF (116KB)  |    1.6K |  639us | 1.07ms | ±0.96% |     782 |
-| load heavy PDF (9.9MB) |   909.7 | 1.10ms | 1.46ms | ±0.81% |     455 |
+| load small PDF (888B)  |   14.7K |   68us |  139us | ±1.61% |   7,359 |
+| load medium PDF (19KB) |    9.7K |  103us |  197us | ±1.32% |   4,852 |
+| load form PDF (116KB)  |   661.3 | 1.51ms | 2.71ms | ±2.06% |     331 |
+| load heavy PDF (9.9MB) |   417.6 | 2.39ms | 3.61ms | ±1.55% |     209 |
 
-- **load small PDF (888B)** is 1.62x faster than load medium PDF (19KB)
-- **load small PDF (888B)** is 24.73x faster than load form PDF (116KB)
-- **load small PDF (888B)** is 42.51x faster than load heavy PDF (9.9MB)
+- **load small PDF (888B)** is 1.52x faster than load medium PDF (19KB)
+- **load small PDF (888B)** is 22.25x faster than load form PDF (116KB)
+- **load small PDF (888B)** is 35.24x faster than load heavy PDF (9.9MB)
 
 ## Saving
 
 | Benchmark                          | ops/sec |   Mean |    p99 |    RME | Samples |
 | :--------------------------------- | ------: | -----: | -----: | -----: | ------: |
-| save unmodified (19KB)             |   22.2K |   45us |   63us | ±0.78% |  11,103 |
-| incremental save (19KB)            |    6.9K |  144us |  373us | ±1.14% |   3,461 |
-| save with modifications (19KB)     |    2.4K |  422us |  799us | ±0.97% |   1,185 |
-| save heavy PDF (9.9MB)             |   850.9 | 1.18ms | 1.58ms | ±0.91% |     426 |
-| incremental save heavy PDF (9.9MB) |   494.7 | 2.02ms | 2.37ms | ±0.74% |     248 |
+| save unmodified (19KB)             |    7.9K |  126us |  321us | ±1.64% |   3,956 |
+| incremental save (19KB)            |    1.9K |  527us | 1.06ms | ±1.51% |     948 |
+| save with modifications (19KB)     |   762.8 | 1.31ms | 2.72ms | ±2.37% |     382 |
+| save heavy PDF (9.9MB)             |   413.5 | 2.42ms | 3.13ms | ±1.12% |     207 |
+| incremental save heavy PDF (9.9MB) |   137.6 | 7.26ms | 7.58ms | ±0.44% |      69 |
 
-- **save unmodified (19KB)** is 3.21x faster than incremental save (19KB)
-- **save unmodified (19KB)** is 9.37x faster than save with modifications (19KB)
-- **save unmodified (19KB)** is 26.10x faster than save heavy PDF (9.9MB)
-- **save unmodified (19KB)** is 44.89x faster than incremental save heavy PDF (9.9MB)
+- **save unmodified (19KB)** is 4.17x faster than incremental save (19KB)
+- **save unmodified (19KB)** is 10.37x faster than save with modifications (19KB)
+- **save unmodified (19KB)** is 19.13x faster than save heavy PDF (9.9MB)
+- **save unmodified (19KB)** is 57.48x faster than incremental save heavy PDF (9.9MB)
 
 ## Splitting
 
@@ -203,32 +212,32 @@
 
 | Benchmark                                | ops/sec |    Mean |     p99 |    RME | Samples |
 | :--------------------------------------- | ------: | ------: | ------: | -----: | ------: |
-| extractPages (1 page from small PDF)     |    2.2K |   452us |   931us | ±1.63% |   1,106 |
-| extractPages (1 page from 100-page PDF)  |   536.3 |  1.86ms |  3.04ms | ±1.54% |     269 |
-| extractPages (1 page from 2000-page PDF) |    24.7 | 40.43ms | 42.46ms | ±2.46% |      13 |
+| extractPages (1 page from small PDF)     |   753.5 |  1.33ms |  2.47ms | ±2.83% |     377 |
+| extractPages (1 page from 100-page PDF)  |   197.6 |  5.06ms |  9.10ms | ±3.08% |      99 |
+| extractPages (1 page from 2000-page PDF) |    13.1 | 76.08ms | 78.10ms | ±1.22% |      10 |
 
-- **extractPages (1 page from small PDF)** is 4.12x faster than extractPages (1 page from 100-page PDF)
-- **extractPages (1 page from small PDF)** is 89.36x faster than extractPages (1 page from 2000-page PDF)
+- **extractPages (1 page from small PDF)** is 3.81x faster than extractPages (1 page from 100-page PDF)
+- **extractPages (1 page from small PDF)** is 57.32x faster than extractPages (1 page from 2000-page PDF)
 
 ### Split into single-page PDFs
 
-| Benchmark                   | ops/sec |     Mean |      p99 |    RME | Samples |
-| :-------------------------- | ------: | -------: | -------: | -----: | ------: |
-| split 100-page PDF (0.1MB)  |    32.2 |  31.02ms |  35.44ms | ±2.72% |      17 |
-| split 2000-page PDF (0.9MB) |     1.8 | 550.66ms | 550.66ms | ±0.00% |       1 |
+| Benchmark                   | ops/sec |    Mean |      p99 |    RME | Samples |
+| :-------------------------- | ------: | ------: | -------: | -----: | ------: |
+| split 100-page PDF (0.1MB)  |    10.9 | 91.83ms | 102.50ms | ±6.62% |       6 |
+| split 2000-page PDF (0.9MB) |   0.650 |   1.54s |    1.54s | ±0.00% |       1 |
 
-- **split 100-page PDF (0.1MB)** is 17.75x faster than split 2000-page PDF (0.9MB)
+- **split 100-page PDF (0.1MB)** is 16.76x faster than split 2000-page PDF (0.9MB)
 
 ### Batch page extraction
 
-| Benchmark                                              | ops/sec |    Mean |     p99 |    RME | Samples |
-| :----------------------------------------------------- | ------: | ------: | ------: | -----: | ------: |
-| extract first 10 pages from 2000-page PDF              |    23.3 | 42.90ms | 46.80ms | ±3.59% |      12 |
-| extract first 100 pages from 2000-page PDF             |    20.1 | 49.79ms | 52.72ms | ±2.98% |      11 |
-| extract every 10th page from 2000-page PDF (200 pages) |    18.6 | 53.74ms | 59.30ms | ±3.19% |      10 |
+| Benchmark                                              | ops/sec |     Mean |      p99 |    RME | Samples |
+| :----------------------------------------------------- | ------: | -------: | -------: | -----: | ------: |
+| extract first 10 pages from 2000-page PDF              |    12.2 |  81.95ms |  91.73ms | ±4.95% |       7 |
+| extract first 100 pages from 2000-page PDF             |     9.2 | 109.19ms | 110.20ms | ±1.20% |       5 |
+| extract every 10th page from 2000-page PDF (200 pages) |     8.0 | 125.67ms | 127.43ms | ±1.82% |       4 |
 
-- **extract first 10 pages from 2000-page PDF** is 1.16x faster than extract first 100 pages from 2000-page PDF
-- **extract first 10 pages from 2000-page PDF** is 1.25x faster than extract every 10th page from 2000-page PDF (200 pages)
+- **extract first 10 pages from 2000-page PDF** is 1.33x faster than extract first 100 pages from 2000-page PDF
+- **extract first 10 pages from 2000-page PDF** is 1.53x faster than extract every 10th page from 2000-page PDF (200 pages)
 
 ---
 

From d2b074190ef10ca0eac4868912d2a058cc8f393d Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 14:03:33 +1100
Subject: [PATCH 7/9] ci: simplify PR benchmark workflow
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Just run benchmarks and post results as a PR comment.
No base comparison — check manually if needed.
---
 .github/workflows/bench-pr.yml |  36 ++------
 scripts/bench-comment.ts       |  95 +++++++++++++++++++++
 scripts/bench-compare.ts       | 150 ---------------------------------
 3 files changed, 103 insertions(+), 178 deletions(-)
 create mode 100644 scripts/bench-comment.ts
 delete mode 100644 scripts/bench-compare.ts

diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
index 9321de1..d82a1d7 100644
--- a/.github/workflows/bench-pr.yml
+++ b/.github/workflows/bench-pr.yml
@@ -17,46 +17,26 @@ permissions:
 
 jobs:
   bench:
-    name: Benchmark Comparison
+    name: Run Benchmarks
     runs-on: ubuntu-latest
     steps:
-      - name: Checkout PR
+      - name: Checkout
         uses: actions/checkout@v4
-        with:
-          path: pr
-
-      - name: Checkout base
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event.pull_request.base.sha }}
-          path: base
 
       - name: Setup Bun
         uses: oven-sh/setup-bun@v2
 
-      - name: Install dependencies (base)
+      - name: Install dependencies
         run: bun install --frozen-lockfile
-        working-directory: base
-
-      - name: Install dependencies (PR)
-        run: bun install --frozen-lockfile
-        working-directory: pr
-
-      - name: Run benchmarks (base)
-        run: bun run bench -- --outputJson ../base-results.json benchmarks/splitting.bench.ts
-        working-directory: base
-        continue-on-error: true
 
-      - name: Run benchmarks (PR)
-        run: bun run bench -- --outputJson ../pr-results.json benchmarks/splitting.bench.ts
-        working-directory: pr
+      - name: Run benchmarks
+        run: bun run bench -- --outputJson bench-results.json benchmarks/splitting.bench.ts
 
-      - name: Generate comparison
-        run: bun run pr/scripts/bench-compare.ts ../base-results.json ../pr-results.json bench-comment.md
-        working-directory: pr
+      - name: Generate comment
+        run: bun run scripts/bench-comment.ts bench-results.json bench-comment.md
 
       - name: Post or update PR comment
         uses: marocchino/sticky-pull-request-comment@v2
         with:
           header: benchmark-results
-          path: pr/bench-comment.md
+          path: bench-comment.md
diff --git a/scripts/bench-comment.ts b/scripts/bench-comment.ts
new file mode 100644
index 0000000..8cc354d
--- /dev/null
+++ b/scripts/bench-comment.ts
@@ -0,0 +1,95 @@
+/**
+ * Format benchmark JSON results as a markdown comment for PRs.
+ *
+ * Usage:
+ *   bun run scripts/bench-comment.ts <results.json> <output.md>
+ */
+
+import { readFileSync } from "node:fs";
+
+interface Bench {
+  name: string;
+  mean: number;
+  hz: number;
+  p99: number;
+  rme: number;
+  sampleCount: number;
+}
+
+interface Group {
+  fullName: string;
+  benchmarks: Bench[];
+}
+
+interface File {
+  filepath: string;
+  groups: Group[];
+}
+
+interface Output {
+  files: File[];
+}
+
+function formatMs(ms: number): string {
+  if (ms >= 1000) {
+    return `${(ms / 1000).toFixed(2)}s`;
+  }
+
+  if (ms >= 1) {
+    return `${ms.toFixed(2)}ms`;
+  }
+
+  return `${(ms * 1000).toFixed(0)}μs`;
+}
+
+function formatRme(rme: number): string {
+  return `±${rme.toFixed(1)}%`;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+
+const [inputPath, outputPath] = process.argv.slice(2);
+
+if (!inputPath || !outputPath) {
+  console.error("Usage: bun run scripts/bench-comment.ts <results.json> <output.md>");
+  process.exit(1);
+}
+
+const data: Output = JSON.parse(readFileSync(inputPath, "utf-8"));
+
+const lines: string[] = [];
+lines.push("## Benchmark Results");
+lines.push("");
+
+for (const file of data.files) {
+  for (const group of file.groups) {
+    const groupName = group.fullName.includes(" > ")
+      ? group.fullName.split(" > ").slice(1).join(" > ")
+      : group.fullName;
+
+    lines.push(`**${groupName}**`);
+    lines.push("");
+    lines.push("| Benchmark | Mean | p99 | RME | Samples |");
+    lines.push("|:---|---:|---:|---:|---:|");
+
+    for (const b of group.benchmarks) {
+      lines.push(
+        `| ${b.name} | ${formatMs(b.mean)} | ${formatMs(b.p99)} | ${formatRme(b.rme)} | ${b.sampleCount} |`,
+      );
+    }
+
+    lines.push("");
+  }
+}
+
+lines.push(
+  `<details><summary>Environment</summary>\n\n` +
+    `- Runner: \`ubuntu-latest\`\n` +
+    `- Runtime: Bun ${process.versions.bun}\n\n` +
+    `*Results are machine-dependent.*\n` +
+    `</details>`,
+);
+
+const body = lines.join("\n");
+await Bun.write(outputPath, body);
+console.log(body);
diff --git a/scripts/bench-compare.ts b/scripts/bench-compare.ts
deleted file mode 100644
index 2be007e..0000000
--- a/scripts/bench-compare.ts
+++ /dev/null
@@ -1,150 +0,0 @@
-/**
- * Compare two benchmark JSON files and produce a markdown summary.
- *
- * Usage:
- *   bun run scripts/bench-compare.ts <base.json> <pr.json> <output.md>
- *
- * If <base.json> does not exist, outputs PR-only results.
- */
-
-import { existsSync, readFileSync } from "node:fs";
-
-interface Bench {
-  name: string;
-  mean: number;
-  hz: number;
-  p99: number;
-  rme: number;
-  sampleCount: number;
-}
-
-interface Group {
-  fullName: string;
-  benchmarks: Bench[];
-}
-
-interface File {
-  filepath: string;
-  groups: Group[];
-}
-
-interface Output {
-  files: File[];
-}
-
-function buildMap(data: Output): Map<string, Bench> {
-  const map = new Map<string, Bench>();
-
-  for (const file of data.files) {
-    for (const group of file.groups) {
-      for (const b of group.benchmarks) {
-        map.set(b.name, b);
-      }
-    }
-  }
-
-  return map;
-}
-
-function formatMs(ms: number): string {
-  if (ms >= 1000) {
-    return `${(ms / 1000).toFixed(2)}s`;
-  }
-
-  if (ms >= 1) {
-    return `${ms.toFixed(2)}ms`;
-  }
-
-  return `${(ms * 1000).toFixed(0)}μs`;
-}
-
-// ─────────────────────────────────────────────────────────────────────────────
-
-const [basePath, prPath, outputPath] = process.argv.slice(2);
-
-if (!prPath || !outputPath) {
-  console.error("Usage: bun run scripts/bench-compare.ts <base.json> <pr.json> <output.md>");
-  process.exit(1);
-}
-
-const pr: Output = JSON.parse(readFileSync(prPath, "utf-8"));
-const prMap = buildMap(pr);
-const hasBase = existsSync(basePath);
-
-const lines: string[] = [];
-lines.push("## Benchmark Results");
-lines.push("");
-
-if (hasBase) {
-  const base: Output = JSON.parse(readFileSync(basePath, "utf-8"));
-  const baseMap = buildMap(base);
-
-  lines.push("| Benchmark | Base | PR | Change |");
-  lines.push("|:---|---:|---:|---:|");
-
-  let hasRegression = false;
-
-  for (const [name, b] of baseMap) {
-    const p = prMap.get(name);
-
-    if (!p) {
-      continue;
-    }
-
-    const ratio = b.mean / p.mean;
-    let change: string;
-    let indicator = "";
-
-    if (ratio > 1.05) {
-      change = `${ratio.toFixed(2)}x faster`;
-      indicator = " 🟢";
-    } else if (ratio < 0.95) {
-      change = `${(1 / ratio).toFixed(2)}x slower`;
-      indicator = " 🔴";
-      hasRegression = true;
-    } else {
-      change = "~same";
-    }
-
-    lines.push(`| ${name} | ${formatMs(b.mean)} | ${formatMs(p.mean)} | ${change}${indicator} |`);
-  }
-
-  // Show benchmarks only in PR (new benchmarks)
-  for (const [name, p] of prMap) {
-    if (!baseMap.has(name)) {
-      lines.push(`| ${name} | — | ${formatMs(p.mean)} | *new* |`);
-    }
-  }
-
-  lines.push("");
-
-  if (hasRegression) {
-    lines.push("> ⚠️ **Performance regression detected.** Please review the changes above.");
-  } else {
-    lines.push("> ✅ No performance regressions detected.");
-  }
-} else {
-  // No base results — just show PR numbers
-  lines.push("*No base benchmarks available for comparison (new benchmark suite?).*");
-  lines.push("");
-  lines.push("| Benchmark | Mean | p99 | Samples |");
-  lines.push("|:---|---:|---:|---:|");
-
-  for (const [name, p] of prMap) {
-    lines.push(`| ${name} | ${formatMs(p.mean)} | ${formatMs(p.p99)} | ${p.sampleCount} |`);
-  }
-}
-
-lines.push("");
-lines.push(
-  `<details><summary>Environment</summary>\n\n` +
-    `- Runner: \`ubuntu-latest\`\n` +
-    `- Runtime: Bun ${process.versions.bun}\n` +
-    `- Benchmark: \`benchmarks/splitting.bench.ts\`\n\n` +
-    `*Results are machine-dependent. Thresholds: >5% faster 🟢, >5% slower 🔴.*\n` +
-    `</details>`,
-);
-
-const body = lines.join("\n");
-await Bun.write(outputPath, body);
-console.log(body);

From cbf5f98efbcc66355707d34e1d9508ff3b1ab57b Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 14:06:24 +1100
Subject: [PATCH 8/9] ci: run all benchmarks, use collapsible sections in PR
 comment

---
 .github/workflows/bench-pr.yml |  2 +-
 scripts/bench-comment.ts       | 23 +++++++++++++++++++++++
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
index d82a1d7..d0288d7 100644
--- a/.github/workflows/bench-pr.yml
+++ b/.github/workflows/bench-pr.yml
@@ -30,7 +30,7 @@ jobs:
         run: bun install --frozen-lockfile
 
       - name: Run benchmarks
-        run: bun run bench -- --outputJson bench-results.json benchmarks/splitting.bench.ts
+        run: bun run bench -- --outputJson bench-results.json
 
       - name: Generate comment
         run: bun run scripts/bench-comment.ts bench-results.json bench-comment.md
diff --git a/scripts/bench-comment.ts b/scripts/bench-comment.ts
index 8cc354d..3dd103f 100644
--- a/scripts/bench-comment.ts
+++ b/scripts/bench-comment.ts
@@ -1,6 +1,8 @@
 /**
  * Format benchmark JSON results as a markdown comment for PRs.
  *
+ * Each benchmark file gets its own collapsible section.
+ *
  * Usage:
  *   bun run scripts/bench-comment.ts <results.json> <output.md>
  */
@@ -46,6 +48,18 @@ function formatRme(rme: number): string {
   return `±${rme.toFixed(1)}%`;
 }
 
+function fileLabel(filepath: string): string {
+  const match = filepath.match(/([^/]+)\.bench\.ts$/);
+
+  if (!match) {
+    return filepath;
+  }
+
+  const name = match[1];
+
+  return name.charAt(0).toUpperCase() + name.slice(1);
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 
 const [inputPath, outputPath] = process.argv.slice(2);
@@ -62,6 +76,12 @@ lines.push("## Benchmark Results");
 lines.push("");
 
 for (const file of data.files) {
+  const label = fileLabel(file.filepath);
+
+  lines.push(`<details>`);
+  lines.push(`<summary><strong>${label}</strong></summary>`);
+  lines.push("");
+
   for (const group of file.groups) {
     const groupName = group.fullName.includes(" > ")
       ? group.fullName.split(" > ").slice(1).join(" > ")
@@ -80,6 +100,9 @@ for (const file of data.files) {
 
     lines.push("");
   }
+
+  lines.push(`</details>`);
+  lines.push("");
 }
 
 lines.push(

From 44141fc58649cd79b8e1383135678de0b679b15c Mon Sep 17 00:00:00 2001
From: Lucas Smith <me@lucasjamessmith.me>
Date: Wed, 18 Feb 2026 14:07:00 +1100
Subject: [PATCH 9/9] ci: use actual runner info instead of hardcoded string

---
 .github/workflows/bench-pr.yml | 2 ++
 scripts/bench-comment.ts       | 4 +++-
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
index d0288d7..cd3e0f0 100644
--- a/.github/workflows/bench-pr.yml
+++ b/.github/workflows/bench-pr.yml
@@ -34,6 +34,8 @@ jobs:
 
       - name: Generate comment
         run: bun run scripts/bench-comment.ts bench-results.json bench-comment.md
+        env:
+          BENCH_RUNNER: ${{ runner.os }} (${{ runner.arch }})
 
       - name: Post or update PR comment
         uses: marocchino/sticky-pull-request-comment@v2
diff --git a/scripts/bench-comment.ts b/scripts/bench-comment.ts
index 3dd103f..b6df46b 100644
--- a/scripts/bench-comment.ts
+++ b/scripts/bench-comment.ts
@@ -105,9 +105,11 @@ for (const file of data.files) {
   lines.push("");
 }
 
+const runner = process.env.BENCH_RUNNER ?? "local";
+
 lines.push(
   `<details><summary>Environment</summary>\n\n` +
-    `- Runner: \`ubuntu-latest\`\n` +
+    `- Runner: \`${runner}\`\n` +
     `- Runtime: Bun ${process.versions.bun}\n\n` +
     `*Results are machine-dependent.*\n` +
     `</details>`,