fix(audit-noise): under/at/on carry the same blanket-exemption weakness that in had, so conversational residue still escapes

[kyle-sexton]

Summary

#3162 narrowed conversational-antecedent's in follower from a blanket exemption to a document-locator predicate, because as we discussed in yesterday's meeting was being silently suppressed. under, at, and on still have the identical weakness — they exempt unconditionally.

Same defect class, deliberately left out of that change so it could get its own corpus measurement.

What escapes today

Input	Current	Should be
As we agreed on Tuesday, the cap is 30s.	not flagged	flagged (T1)
As we decided at the standup, retry three times.	not flagged	flagged (T1)
As we agreed under time pressure, this is temporary.	not flagged	flagged (T1)
As we decided on the ADR's recommendation, …	not flagged	correctly not flagged

The referent in the first three is the conversation, not a written locus a reader can open — which is exactly the distinction audit_noise_follower_is_document_locator already encodes for in.

Why this was not folded into #3162

on in particular is high-traffic prose (based on, on the other hand, depends on, on disk), so narrowing it is a materially different false-positive risk from in. #3162's whole method was measuring each pattern change against the repo's 1136-file tracked-markdown corpus before shipping it; bolting three more followers on without that measurement would have undercut the standard the rest of that change was held to.

Suggested approach

The predicate already exists — audit_noise_follower_is_document_locator in plugins/docs-hygiene/skills/audit-noise/scripts/lib/noise-shapes.sh. The work is deciding, per follower, whether it should defer to that predicate or keep exempting unconditionally, and proving each choice against the corpus:

• Baseline the corpus sweep first (detect.sh --paths-file over git ls-files '*.md' minus **/evals/fixtures/** and CHANGELOG.md).
• Adjudicate every added finding as true or false positive — this change raises recall, so a non-zero delta is expected and each one needs a ruling.
• on may warrant a tighter locator set than in, or may not be worth narrowing at all. That is a finding, not a failure.

Acceptance criteria

• Each of under, at, on either defers to the locator predicate or keeps its blanket exemption, with the choice justified against corpus evidence.
• Positive tests: as we agreed on Tuesday and as we decided at the standup flag.
• Negative tests prove no new false positives on the high-traffic on idioms.
• The existing above/below/earlier/later/previously/elsewhere stop-words keep working unchanged.
• SKILL.md's conversational-antecedent treatment row stays true after the change.

Related

• feat(docs-hygiene): audit-noise gains three residue shapes (0.20.0) #3162 — narrowed in; recorded this as a deferred recall gap in its ## Related.

[kyle-sexton]

🤖 AI-authored triage note:

This was generated by AI during triage.

Verification

Confirmed by direct code inspection against main (702967d). audit_noise_line_has_conversational_antecedent's follower-exemption case statement exempts above | below | earlier | later | previously | elsewhere | under | at | on unconditionally (return 1), while in alone is routed through audit_noise_follower_is_document_locator, a predicate (already implemented, generic across prepositions) that only exempts the phrase when its head is an actual document-locus reference (a path, markdown link, section/spec/ADR/etc. word). under, at, and on retain the old blanket exemption #3162 narrowed only for in — confirming the reported gap exactly, including that the fix pattern (route through the existing locator predicate) already exists in the file.

Agent Brief

Type: Bug
Summary: audit-noise's conversational-antecedent shape narrowed the in follower from a blanket exemption to a document-locator check in a prior fix, but under, at, and on still carry the old unconditional exemption, so conversational residue phrased with those prepositions escapes detection.

Current behavior: When a prohibited conversational-antecedent phrase ("as we discussed …") is followed by under, at, or on, the shape exempts it unconditionally regardless of what follows — so "as we agreed on Tuesday" (a real conversational reference) is never flagged, identically to how "as we agreed in yesterday's meeting" behaved before the in fix. Only in was narrowed to check whether the phrase actually points at a document locus.

Desired behavior: under, at, and on are checked with the same document-locator predicate in already uses: exempt the phrase only when what follows is an actual document-locus reference (a path, a markdown link, or a locator word like section/spec/ADR/convention/etc.); otherwise flag it as a conversational-antecedent finding, matching the table in the issue body (e.g. "as we agreed on Tuesday" flags; "as we decided on the ADR's recommendation" stays exempt).

Key interfaces:

• The follower-exemption case statement in the conversational-antecedent shape's detection function — the under | at | on branch should route through the same document-locator predicate the in branch already calls, rather than returning an unconditional exemption.

Acceptance criteria:

• "As we agreed on Tuesday, the cap is 30s." is flagged as a conversational-antecedent finding.
• "As we decided at the standup, retry three times." is flagged as a conversational-antecedent finding.
• "As we agreed under time pressure, this is temporary." is flagged as a conversational-antecedent finding.
• "As we decided on the ADR's recommendation, …" remains correctly exempt (document-locus reference).
• The existing above | below | earlier | later | previously | elsewhere blanket exemptions are unchanged — they are out of scope for this fix.
• The audit-noise detection test suite passes with new cases covering the three prepositions from the issue's own table.

Out of scope:

• Changing the above/below/earlier/later/previously/elsewhere followers' exemption behavior.
• Any change to the in-follower logic itself, which is already correct.