phase-research-coverage — searching discipline

Census id: F22 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (staging FETCH_HEAD, read 2026-09-11; re-anchored from census pin a1337cb301). Every file:line below is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.

What it bans

Closing a negative-existence claim with incomplete search coverage. .claude/rules/phase-research-coverage.md:10 «# Phase research coverage — searching discipline»; the application point: :27 «Apply before closing any «N candidates checked → no production analog» claim. Failing any single item → the negative-existence verdict is provisional, not load-bearing.» The checklist (§1, :29-34): own-stack sweep, category sweep, semantic-distance check, adversarial check on negative-existence claims, «Prompt-list ≠ complete», trigger sweep — plus §1.7 recommendation self-discipline (forward-check + backward-check, :35-39). Class header: :12 «> Class: A — companion principle test shipped at [packages/core/principles/13-phase-research-coverage-s17.test.ts] (#74, 2026-05-17); enforces §1.7 Forward+Backward self-review on research patches via the HISTORICAL_CUTOFF mechanism.»

Never (fires)

A load-bearing «no analog» verdict with a floor-level lookup: .claude/rules/phase-research-coverage.md:33 «5. Prompt-list ≠ complete. Hard Constraint #5 / #10 in the entry-research prompt typically lists ≥3 candidates per area. That is a floor, not a ceiling. Closing at the floor is permitted only if items 1-4 above also hold. If they don't, the lookup must continue past the listed minimum.» And the backward-check restatement shape: :37 «A backward-check whose surface list is exactly the diff's own files … is a restatement, not a sweep — the [#backward-check-restates-not-sweeps] (T21) failure that is worst under context fatigue.»

Always (clean)

Each checklist item run and recorded — e.g. item 4: .claude/rules/phase-research-coverage.md:32 «4. Adversarial check on negative-existence claims. A claim of the form «no production tool implements X» is a strong assertion. Before accepting it, generate at least one counter-prompt that assumes the tool exists and tries to find it»; and the §1.7 authoring format: :37 «Authoring format (enforced at review, mandatory since T21): write the backward-check as Class of this change = <content predicate>. Surfaces where class-X occurs: [enumerate the COMPLETE set with grep/find evidence]. Per surface: SWEPT-CLEAN (evidence file:line) | GAP-FOUND (action).»

Enforcement channels

• Rule-index row: .claude/rules/00-rule-index.md:30 «| phase-research-coverage.md | A | phase research, prior-art lookups, negative-existence claims. | paths:(4) |»
• CC-native paths: only — no edit-time sibling, by design: .claude/rules/phase-research-coverage.md:21 «> Delivery channel — CC-native paths: only; no <!-- globs: --> sibling, by design (T-SEF-A).» — the hook glob grammar cannot express the mid-path-star scope, and the expressible superset «would over-fire on every meta-factory doc edit».
• Companion principle test (CI): packages/core/principles/13-phase-research-coverage-s17.test.ts (#74) — §1.7 Forward+Backward enforcement on research patches (:12).
• Cold-agent delegation for sibling sweeps: agents/backward-sweep-auditor.md («hand it ONLY the change's class … so it enumerates siblings without being able to restate the PR it never saw», :37).
• Not backend-rendered — no FF diagnostic applies (FF7001/FF7002, packages/core/diagnostics/registry.ts:321/:329); honest status: read-time paths + CI principle test + review-time format enforcement.