fix(eval): make eval runners self-contained for deterministic headless scoring

[evemcgivern]

What

Make the eval runners self-contained so headless scoring is deterministic.

The runners injected the valid label set as opaque IDs (cat#1..5) but not their
definitions. The headless scan is sandboxed to the temp work dir, so the model could
not read the skill's baseline-categories.md to learn what those numbers mean. It both:

• thrashed on denied reads → intermittently never emitted the <<<EVAL>>> block (PARTIAL), and
• guessed categories → cat#2-vs-cat#3 flips on the same fixture.

Fix

• Build a cat#N = <definition> legend from baseline-categories.md (read-only, in the
  dev repo — no sandbox loosening, no plugin-path discovery) and inject it into the prompt.
• Tell the model the definitions are provided inline, so it need not read skill files; steer
  it to reason directly on the single fixture (the scope/triage/tool stages are N/A for one
  file and only generate denied-read noise).
• Strengthen "always emit the block."
• Applied symmetrically to claude.sh and codex.sh — divergence between runners is a bug.

Evidence

java/held-out, --runs 3:

• before: mean_precision=0.33(±0.47) recall=0.17 + PARTIAL
• after: mean_precision=1.00(±0.00) recall=1.00(±0.00) + gate PASS

The ±0.00 is the point — the runner is now deterministic.

Scope / safety

• Read-only invariant preserved (write-deny tools unchanged; legend is read at runner-build
  time from the dev repo). bats tests/detect.bats green (147/0), sh -n clean.
• Cross-cutting: this fixes nondeterminism for every language's calibration. Other
  languages' baselines were set with the old flaky runner and may now measure higher — a
  re-baseline sweep is noted as follow-up in Eval runners inject opaque category IDs without definitions — nondeterministic headless scoring #71.

Closes #71

🤖 Generated with Claude Code