make sure program graph has the same semantics as old trace/ module

[darinkishore]

3. trace/ module — entire directory is orphaned

• Graph, Node, NodeType, Executor, TrackedValue, IntoTracked, record_node, record_output, is_tracing, trace() — zero imports outside the trace module
  itself
• This was a pre-V1 runtime execution tracing system (task-local DAG recording). Completely superseded by ProgramGraph for dynamic graph execution and
  tracing crate for observability.

yeah so we gotta fix that lol

[darinkishore]

Codebase audit — trace/ module status

Files:

• trace/mod.rs — module root, re-exports everything
• trace/context.rs — task-local CURRENT_TRACE using tokio::task_local!, trace() wrapper function
• trace/dag.rs — Graph, Node, NodeType, edge types
• trace/executor.rs — Executor that replays a trace graph
• trace/value.rs — TrackedValue, IntoTracked trait

Usage: lib.rs exports pub mod trace and pub use trace::*. But nothing outside the trace module imports from it. The Predict::forward() code has no calls into trace recording.

ProgramGraph from the design reference (docs/specs/modules/design_reference.md) was prototyped and intentionally removed — lib.rs docs say: "The type-erased ProgramGraph, DynModule, StrategyFactory layer was prototyped and intentionally removed."

What "same semantics" means:
The trace module captured a DAG of Predict invocations (inputs/outputs at each node). ProgramGraph was a structural graph (nodes = modules, edges = data flow between fields). They serve different purposes:

• Trace = runtime execution record (what happened during a call)
• ProgramGraph = structural description (how modules are wired together)

Options:

1. Revive trace/ for runtime recording — wire Predict::forward() to call record_node() when tracing is active. Migrate from Prediction/Example types to BamlValue/typed values. This gives you execution traces for debugging and optimizer data collection.
2. Delete trace/ entirely — it's dead code. Runtime observability is handled by tracing crate spans (chore: add tracing #56). If you don't need DAG replay, the tracing spans are sufficient.
3. Merge into the tracing story — use tracing spans for the execution DAG (custom span attributes for inputs/outputs/module names) instead of a separate recording system.

Recommendation: option 3. The tracing crate already supports structured span hierarchies. A custom Layer impl could reconstruct the execution DAG from spans without maintaining a parallel recording system.