test(spec): Expand 180-point falsification tests to 28 total

Add comprehensive spec verification tests:

Section A (Model Loading):
- A1: Model loads successfully
- A2: Weights are initialized (not all zeros)
- A4: Metadata matches Qwen2-0.5B-Instruct config
- A6: Weights not random Gaussian (bounded output)

Section C (Forward Pass):
- C4: Causal mask stability (no NaN/Inf across sequence lengths)

Section D (Generation):
- D5: Throughput baseline measurement

Section E (Visual Control):
- E1: Top-k logit extraction with softmax
- E3: Token rate measurement

Section G (Code Quality):
- G1: Coverage helper exercising multiple code paths

Section H (Lifecycle):
- H6: Model inspectable (config attributes)

Additional:
- Tensor operations correctness
- Numerical stability edge cases

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: noahgift