[SamGu-NRX]

Why

Second reviewer-confirmed defect from the PR #88 review (reproduced at 393b5de by calling rs.py directly, still live on eval/scorecard-runner at 71b5518): in evaluation/suites/recipes/rs.py, _stated_rest returned the check metadata's rest_minutes before ever reading the step text. The cook reads the instruction text, not the metadata, so:

• metadata rest_minutes: 3 with text "Rest 1 minute" returned 3.0 — the whole-cut check cleared on a rest the instructions never state;
• metadata rest_minutes: 3 with no rest stated in the text also returned 3.0 — the cook was never told to rest the food, and the required-rest check passed anyway.

What changed

The stated rest is now bound to the text the cook reads — the same cook-visibility principle fix 1 (71b5518) applied to the internal reading:

• _stated_rest(step) reads only the step text (_REST_TEXT): "Rest 1 minute" → 1.0; no stated rest → None. The metadata never substitutes.
• When the text has no rest where the class requires one, the required-rest violation fires (recipe fails).
• When the check metadata also names a rest that DISAGREES with the text, the linter adds a violation naming both values — neither value is silently preferred; when they agree (or the metadata is silent) nothing changes.

Boundary: a rest_minutes value agreeing with the text passes unchanged; a check with no rest_minutes metadata and a valid text rest passes unchanged (the 400-recipe control corpus pattern).

Red/green proof

Both reviewer cases are planted as D-series plants D1f/D1g in the suite C planted-failure battery.

Red (844bf1c, rs.py untouched — the defect is live):

stated = 3.0
stated = 3.0
plant D1f whole-cut rest clears on rest_minutes metadata over the text's 1-minute rest: MISSED (food_safety_linter_pass=pass)
plant D1g whole-cut rest clears on rest_minutes metadata with no rest in the text: MISSED (food_safety_linter_pass=pass)
caught 17/19 planted failures; controls 2/2 held        (suite.py --self-test, exit 1)
run.py --self-test: SELF-TEST FAILED 35/36              (exit 2)

Green (2eedb0a, the fix):

stated = 1.0
stated = None
plant D1f ...: caught (food_safety_linter_pass=fail)   <- below-minimum + disagreement naming both values
plant D1g ...: caught (food_safety_linter_pass=fail)   <- must state the required 3-minute rest
caught 19/19 planted failures; controls 2/2 held; part-A safety-lint battery: 4/4 plants, controls held   (suite.py --self-test, exit 0)
self-test ok: 36/36 runner cases + suites behaved as required   (run.py --self-test, exit 0)
all 8 plants caught: the runner refused every forgery            (planted_failures.py --expect-green, exit 0)
allergen-facts verdict self-tests: 7 passed, 0 failed, 0 skipped (test_allergen_facts.py, exit 0)
selftest_c: 4/4 match the fixed contract                         (selftest_c.py, exit 0)

Exact outputs: evaluation/tests/evidence-red-rest-metadata.txt and evaluation/tests/evidence-green-rest-metadata.txt.

No currently-passing control regressed: D1c (whole-cut without the required 3-minute rest) stays caught; the part-A controls (whole-cut text 145F + 3-min rest; poultry text 165F) and the 400-recipe clean corpus stay clean.

Tradeoffs

• The stated rest now comes from the text only. A recipe whose text states a valid rest (>= required) but carries a disagreeing rest_minutes metadata now FAILS until the two agree. That is deliberate: the fix bar is "never silently prefer either" — a metadata value the cook never sees must not quietly bless a different rest, even a safe one.
• _stated_rest lost its chk parameter: with metadata no longer feeding the stated rest, keeping the parameter would misstate the function's inputs. The reviewer's direct-call repro rs._stated_rest(step, chk) becomes rs._stated_rest(step).
• The disagreement flag requires both values to be present; metadata claiming a rest while the text states none is reported by the existing must-state violation (naming the requirement), not a second disagreement line.

Handoff

Done:

• D1f/D1g planted red (844bf1c), fix landed (2eedb0a), full battery green at branch head 2eedb0a; evidence committed under evaluation/tests/.
• New counts: suite C battery 19/19 plants + 2/2 controls + part-A battery 4/4 plants and controls held; runner 36/36; runner planted failures 8/8; allergen-facts 7/7; selftest_c 4/4 — every command exit 0.

Left:

• Parent-thread join: folding this branch into the parent PR — the parent thread owns PR Evaluation scorecard harness: runner, suite contract, registry (fl-eval-harness-f1, m1) #88, the join and eval/scorecard-runner.
• Human/reviewer review of this draft PR.

How to continue:

• Re-verify at the branch head with the five battery commands (all exit 0): python3 evaluation/run.py --self-test; python3 evaluation/tests/planted_failures.py --expect-green; python3 evaluation/suites/recipes/suite.py --self-test; python3 evaluation/suites/recipes/test_allergen_facts.py; python3 evaluation/suites/recipes/selftest_c.py.
• The branch's commits are additive on 71b5518; if eval/scorecard-runner moves before the join, re-run the battery at the merge head (never rewrite this branch's history).

🔗 Obvious Project · 🧵 Obvious Thread · 🔀 Obvious PR Artifact

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}