feat(bench): cascade scenario family + de-bias the harness (#100)

[Connorrmcd6]

Fixes the pilot ceiling found in the first smoke test (#113) and adds the realistic "cascade" scenario family that is the bench's intended headline.

Why

The first smoke run (haiku×2, #113) showed a success-rate ceiling: under stale docs the model was 14/14 correct, 0% misled, so the Surface effect was flat zero. Two causes:

1. The system prompt told the model "the source code is the ground truth" — i.e. ignore the docs, exactly the behaviour that nullifies a stale-doc effect.
2. The single-file comprehension framing (drifted code + contradicting doc side-by-side) lets the model just read the one function. That isn't how real context rot bites.

What changed

Harness (small, backward-compatible):

• Neutralized the system prompt — declares no precedence between docs and code.
• Hidden-dependency support — meta.toml hidden_paths lists code/ files that stay present for grading (so surf seals a real divergence and the grader runs against them) but are withheld from the prompt. prompts._render_code skips them; run.py/metrics.py/grade_code.py unchanged.

New "cascade" family — the agent edits a visible function whose correctness depends on a hidden dependency it knows only through that dependency's doc. A stale doc propagates into a wrong cascaded edit; in C3 the surf report's new_code is the only window onto the truth. Graders derive the expected value from the real hidden dependency, so the test stays honest.

Scenario	Lang	Hidden dependency drift
cascade-quota-batcher-code	py	limiter capacity (<→<=)
cascade-retry-budget-code	py	retry attempt cap (3→5)
cascade-access-policy-code	py	allow-list → block-list
cascade-page-size-ts-code	ts	default page size (50→25)

The 7 comprehension scenarios from #99 remain as a secondary sub-suite (still evidence for H2).

Verification

Rust gates (bench is outside the workspace, so unaffected — run for cleanliness):

• cargo fmt --all --check — clean
• cargo clippy --all-targets --all-features -- -D warnings — clean
• cargo test --all — all suites pass (72/10/38/25/34, 0 failed)

Bench:

• uv run python tools/author.py --all — all 11 scenarios seal; each stale hub genuinely diverges.
• uv run python -m surface_bench.run --models mock — full offline pipeline runs every cell, no crash; confirms hidden_paths rendering.
• For each cascade: a simulated correct answer grades ok/not-misled and a simulated stale answer grades not-ok/misled (both python3 and node --test paths, under uv run), with the dependency confirmed absent from the prompt.

Paid smoke (haiku×2, ~$0.50): on all four cascades, C2 (fresh) and C3 (surf report) succeed 2/2 while C0/C1 fail — H1 and H3 demonstrated, sanity gate passed:

	C0	C1	C2	C3
ok (of 2)	0	0	2	2

A nice incidental: an early C3 "miss" turned out to be the harness correctly catching a leaked doc-trust instruction in my own task text — fixed, and it proves the bench is sensitive.

Next

Full pilot at --trials 10 (user-run, billed) once this lands.

Closes #100

🤖 Generated with Claude Code