fix(eval): runner label injection + measured baselines for rust/shell/yaml/swift (#15)

[evemcgivern]

Summary

Closes out baseline calibration. Two related changes:

1. Runner label injection (the fix). The eval-mode prompt named "a language-specific label" without enumerating the valid ones, so for custom-label languages the model invented synonyms the exact-match grader scored as FP+FN. Runners now run detect.sh eval-categories <lang> and inject the exact valid set.
2. Corrected measured baselines (re-measured with the fixed runner, --runs 2):

Language	before	after
rust	0.67 (unwrap-panic ≠ panic)	1.00 / 1.00
shell	0.67 (SC2086 ≠ quoting)	1.00 / 1.00
yaml	0.67 (norway-problem ≠ coerce)	1.00 / 1.00
swift	PARTIAL (transient)	1.00 / 1.00

All 13 languages now have measured seen baselines at 1.00/1.00.

Deliberately NOT included

• java held-out re-measured 0.25/0.25 with the overfit FLAG still firing — and this is not the label artifact (the same fix made rust/shell/yaml perfect). It's a fixture-quality problem: the agent-authored held-out fixtures (e.g. BugPathTraversal — a new File("/data/"+name) constructor, not a realized read/write) are ambiguous, so a real model reasonably doesn't flag them as labeled. Left at its conservative placeholder; tracked as a follow-up to sharpen those fixtures, then re-measure that split. The overfit FLAG is working as intended until then.

Test plan

• bats tests/detect.bats 147/147; sh -n clean
• label fix verified end-to-end (rust unwrap now emits 3:panic)
• re-measure done on a clean checkout (no branch-switching mid-run)

CODEOWNERS review required (tests/eval/). Part of #15.

🤖 Generated with Claude Code