orb(public-safety): sanitizePublicComment's JSDoc claims "reward" is never skippable — it is

[JSONbored]

⚠️ Definition of Done: this issue must be completed in full, in a single PR. Do not split this
work across multiple PRs, and do not defer any Deliverable below to a follow-up issue. A PR that
satisfies only some of the Deliverables, stubs a required test, or leaves a checkbox
partially-done does NOT resolve this issue and will be closed.

Context

src/queue-intelligence.ts:242-244 documents sanitizePublicComment's escape hatch:

EXPLICIT-PHRASE entries in FORBIDDEN_PUBLIC_COMMENT_WORDS ("trust score", "reward", "scoreability", ...)
are NEVER skippable by this flag -- those name the actual private concept regardless of repo, so they stay
enforced everywhere; only the over-broad bare-word check is ever relaxed

The code does the opposite — :253: const forbidden = options?.allowBareScoreTerm ? ALWAYS_FORBIDDEN_PUBLIC_COMMENT_WORDS : FORBIDDEN_PUBLIC_COMMENT_WORDS;

"reward" and "rewards" are members of AMBIGUOUS_PUBLIC_COMMENT_WORDS (:58-59), not ALWAYS_FORBIDDEN_PUBLIC_COMMENT_WORDS. So sanitizePublicComment("12 TAO reward for this", { allowBareScoreTerm: true }) returns the string unchanged instead of throwing. The flag is live in production: src/services/ai-review.ts:3324 sets it for any repo in LOOPOVER_PUBLIC_SCORE_TERMS_ALLOWED_REPOS. The doc's second claim — "only the over-broad bare-word check is ever relaxed" — is also false: the whole 7-word ambiguous tier is relaxed.

The file already contains the correct, newer description of this behaviour in the #9432 block at :248-252. The older JSDoc immediately above it was never updated, so sanitizePublicComment now carries two contradictory contracts, and the wrong one is the one that reads as a safety guarantee.

Two further factual errors in the same paragraph, both about what BARE_SCORE_TERM_PATTERN = /\bscore\w*\b/i (:115) actually matches:

• :217-218 claims the pattern catches "scoring". It cannot — "scoring" contains no score substring. The file itself states this correctly 100 lines earlier, at :112-114.
• :241-243 cites metagraphed's totalScore / baseTotalScore / credibility field names as the motivating example. /\bscore\w*\b/i.test("totalScore") is false (no word boundary between l and S), and credibility appears in no list.

Requirements

• Rewrite the JSDoc block at src/queue-intelligence.ts:210-247 so every claim is true of the code as written:
  • State that allowBareScoreTerm swaps FORBIDDEN_PUBLIC_COMMENT_WORDS for ALWAYS_FORBIDDEN_PUBLIC_COMMENT_WORDS, i.e. it relaxes the entire AMBIGUOUS_PUBLIC_COMMENT_WORDS tier, and refer to that tier by its constant name rather than listing words inline.
  • Remove the false "reward ... NEVER skippable" claim and the false "only the over-broad bare-word check is ever relaxed" claim.
  • Remove "scoring" from the list of words the bare pattern catches, and remove or correct the totalScore/baseTotalScore/credibility motivating example so it names strings the pattern actually matches.
• Decide the behaviour question the doc raises, in favour of the doc's safety intent, and implement it: move "reward" and "rewards" from AMBIGUOUS_PUBLIC_COMMENT_WORDS to ALWAYS_FORBIDDEN_PUBLIC_COMMENT_WORDS, so a bare reward is rejected on every repo including allowlisted ones. Compound phrases already in the always-forbidden tier are unaffected.
• Add an invariant test pinning the two tiers against the JSDoc's claims, so the doc and the constants cannot drift again: assert ALWAYS_FORBIDDEN_PUBLIC_COMMENT_WORDS contains every word the doc names as never-skippable, and that the two tiers are disjoint.
• No change to BARE_SCORE_TERM_PATTERN, to isPublicScoreTermSafeForRepo, or to the allowlist env var.

⚠️ Required pattern: the #9432 block already at src/queue-intelligence.ts:248-252 is the accurate description — the rewritten JSDoc must agree with it, and the two blocks must not restate the same thing twice. The tier-pinning invariant test should follow the source-pinned drift-guard convention in test/unit/advisory-spend-gate.test.ts:49-71. It does NOT satisfy this issue to fix only the comment and leave reward relaxable on allowlisted repos; to move reward to the always-forbidden tier without correcting the three false statements in the JSDoc; or to add a third word list.

Deliverables

• sanitizePublicComment("12 TAO reward for this", { allowBareScoreTerm: true }) throws — asserted by a named regression test citing the doc/code contradiction.
• sanitizePublicComment("this scores well", { allowBareScoreTerm: true }) still does not throw (the intended relaxation is preserved) — asserted.
• sanitizePublicComment("12 TAO reward for this") (flag off) still throws — asserted, so the fix is not a behaviour swap.
• Invariant test: ALWAYS_FORBIDDEN_PUBLIC_COMMENT_WORDS and AMBIGUOUS_PUBLIC_COMMENT_WORDS are disjoint, and every word the JSDoc names as never-skippable is in the always-forbidden tier.
• The JSDoc at src/queue-intelligence.ts:210-247 no longer contains the string "scoring" as a claimed match, no longer names totalScore/baseTotalScore/credibility as caught by the bare pattern, and no longer claims the ambiguous tier is enforced under the flag.

All Deliverables above are required in a single PR. A PR that satisfies only some of them — for example correcting the comment without moving reward/rewards to the always-forbidden tier — does not resolve this issue.

Test Coverage Requirements

99%+ Codecov patch coverage, branch-counted, on src/**; src/queue-intelligence.ts is inside coverage.include. Both arms of the allowBareScoreTerm ternary at :253 must be exercised against a reward-bearing string and a bare-score string, so all four flag-by-word combinations are covered.

Expected Outcome

A repo on the public-score-terms allowlist can no longer have a bare reward reach a public GitHub comment, and the function's JSDoc describes what the code does rather than a stronger guarantee it never provided. The word-tier invariant test makes the next divergence a CI failure.

Links & Resources

src/queue-intelligence.ts:55-62, :110-118, :210-260; src/services/ai-review.ts:3324; test/unit/advisory-spend-gate.test.ts:49-71.

[linear-code]

JSO-6765