T21: guard §1.7 backward-check against restatement-theatre (cold-sweep agent + enumeration format)

[artyhoo]

Guard the §1.7 backward-check against restatement-theatre under context fatigue

Codifies a durable countermeasure for a verified incident: deep in a long session, a §1.7 Backward-check degrades from an outward sweep (sibling surfaces where the change-class must also hold) into an inward restatement of the PR's own changes — which passes CI and lets a real parallel gap ship.

The incident (verified, not taken on faith)

PR #857 shipped as two commits on staging history:

• ec643bac7 — the §1.7 Backward-check restated the PR: «re-adjudicates the S2 §5 risk-acceptances; item 4 promoted documented→coded; items 1-3 kept documented». It named only artifacts the diff already touched.
• bf1b8b5f3 — after operator challenge, the real sweep found the parallel gap: tier1For's candidateFields (Tier-1 host-derivation) rejected IP-literal / punycode / multi-tenant hosts but not bare TLDs, so homepage:"https://com" derived a trusted com host — the identical single-label hole the PR had just closed on the Tier-2 loadAckFile surface. Fixed with if (!host.includes('.')) continue;.

The theatre backward-check let a real gap reach the PR. The CI gate (discipline-self-check.yml) checks only ≥40 non-whitespace chars + ≥1 file.ext:line citation — both trivially satisfied by a restatement that cites the PR's own diff files.

Why context fatigue produces exactly this (diagnosis)

Deep in a loaded context, working memory is saturated with the PR's inward narrative, so «backward-check» pattern-matches to «recap what I changed» — but the check's entire value is outward (surfaces not in front of you). The check is worst exactly where it is most needed. It composes three catalogued traps: T2 (claiming a sweep vs running it), T14 (clean result on low coverage misreported as clean), T15 (the recursive-self-application the check exists to force is skipped) — reinforced by the fakeable gate that green-lights the least-resistance output (the same evidence-skipping shape as T20 / recommendation-laziness).

Design (evaluated, not defaulted)

• (a) Cold sub-agent — the structural root fix (foregrounded). agents/backward-sweep-auditor.md: a fresh, PR-blind agent given only the change's class — it enumerates every parallel surface and reports GAP/CLEAN. Immune to restatement by construction: it never saw the PR, so it cannot recap it. Plugs into the in-repo AI-agnostic-auditor pattern (agents/memory-codification-auditor.md) + T19 cold-QA; read in-session, zero paid-LLM. Classified operator-only (skip-loop in install.sh + setup.d/20-agents.sh) like the sibling probers — §1.7 is project-internal.
• (b) Enumeration authoring format in phase-research-coverage.md §1.7 + the self-reflection skill: Class = X; Surfaces where class-X occurs: [enumerate ALL]; per surface SWEPT-CLEAN | GAP-FOUND. A backward-check whose surface list is exactly the diff's own files is non-conformant by format.
• (c) Mechanical "cites a non-diff path" detector — REJECTED as a gate, on evidence. The incident's restatement ec643bac7 DID cite a non-diff file (packages/core/principles/30-research-source-trust.test.ts:139, absent from that commit's own git diff --name-only) → a naive non-diff-citation detector false-negatives on the very incident. Completeness of the enumeration is a semantic judgment; per no-paid-llm-in-ci.md it cannot be a CI gate. (Learned from the narrow-b FP-84% lesson — do not ship a noisy/fakeable gate.)

Codified: trap T21 #backward-check-restates-not-sweeps in ai-laziness-traps.md (§2 + §4 anti-pattern + §5 promotion at 3 incidents, MANUAL classification).

§1.7 Forward-check applied

• no-paid-llm-in-ci.md — the only mechanism is the read-in-session cold agent (agents/backward-sweep-auditor.md:14), zero API-billed calls; T21's §5 promotion note explicitly bars a syntactic CI gate (.claude/rules/ai-laziness-traps.md:178).
• doc-authority-hierarchy.md §2-§3 — the new agent carries an Authoritative-for header (agents/backward-sweep-auditor.md:11); both edited rule files retain Class + header. Principle 09 green (64/64 across principles 09/12/21/22).
• build-first-reuse-default.md — ADAPT of the in-repo AI-agnostic-auditor + T19 cold-QA patterns (.claude/rules/ai-laziness-traps.md:144); no new dependency, no capability commit (agent is outside packages/).
• dual-implementation-discipline.md §2(iv)/§3 — operator-only internal agent → CC-native-only is the correct triage; no portable fallback owed (setup.d/20-agents.sh:28).
• language-discipline — all new machinery is English (principle 22 green).

§1.7 Backward-check applied

Class of this change = a discipline self-check enforced by a syntactic proxy (section-presence / min-char-length / citation-count) that a shaped-but-empty (or restated-with-self-citation) body can satisfy without doing the substantive work.

Surfaces where the class occurs (population 12, cold-swept via agents/backward-sweep-auditor.md over .github/workflows + .husky + packages/core/hooks/checks + packages/core/principles) — none is in this PR's diff (git diff --name-only origin/staging...HEAD = 7 files, all rules/skill/agent/installer); this is an outward sweep, not a restatement:

Surface	Verdict (against EMPTY-filler)	Backstop
discipline-self-check.yml:71 §1.7 backward ≥40-char	SWEPT-CLEAN	citation substance arm :123 + sanity paired-negative :134
s17.ts:94 §1.7: trailer length	SWEPT-CLEAN	file:line arm s17.ts:96 + body-prose detector
prior-art.ts:160 trailer ≥20	SWEPT-CLEAN	placeholder + broken-citation arms
audit-self.yml:499 pr-commit-trailers (CI mirror, PREPUSH_ONLY=s17)	SWEPT-CLEAN	same TS-core, no re-implemented proxy
13-phase-research-coverage-s17.test.ts:87 hasS17Substance	SWEPT-CLEAN	bare §1.7 deliberately NOT an arm; C3 paired-negative :226
principles 08 / 15 / 09 / 12 (presence proxies)	SWEPT-CLEAN	each carries anti-tautology / broken-ref / fenced-example negative

GAP-FOUND: none for the EMPTY-filler sub-class. But the restatement-WITH-self-citation sub-class (T21) is un-gated across all 12 surfaces — a body citing its own diff files clears every length AND citation arm. This is why the fix ships a cold semantic agent, not a 13th syntactic arm.

Recursive self-application, honestly: my hand enumeration named ~4 siblings; the cold sweep enumerated 12. Had I shipped only my hand list, this very backward-check would have been an under-complete sweep — T21 catching its own author. The full self-review (population-12 sweep + forward/backward walk + residual) is hosted discoverably at docs/meta-factory/research-patches/2026-07-03-backward-check-restatement-t21.md, per the §1.7 self-reflexive-trigger convention this change edits (not buried in a commit message).

Honest residual (what it still won't catch)

The countermeasure lowers the cost of doing it right (cold agent) and raises the visibility of doing it wrong (format makes an omission legible at review), but does not make faking impossible. A determined fatigued author can still (i) skip invoking the cold agent and (ii) hand-author a plausible-but-incomplete surface list — no CI gate can verify semantic completeness of an enumeration without a paid LLM (excluded by policy), and the cold sweep confirmed that every syntactic arm is defeated by restatement-with-self-citation. So the trap moves from passive restatement (cheap, default) to active fabrication (requires inventing a false-complete list) — a real raise in cost, backstopped only by the same review-time judgment (human or reviewer-agent) that caught the incident. It is not a gate, by design.

Verification

• 275/275 principle tests green (packages/core); markdownlint 0 on all edited files.
• Incident mechanism verified: git show ec643bac7 --name-only | grep -c 30-research-source-trust → 0 (cited-but-not-in-diff, confirming the detector-(c) false-negative).

🤖 Generated with Claude Code