docs(per-role-context): Opus cold-verify — 7 contradictions resolved, 8-item verify-list run

[artyhoo]

What

Opus cold-verify of the 2026-07-26/27 per-role-context research material (PRs #1176-#1180). Resolves all 7 contradictions with primary-source evidence, runs the 8-item verify-list, corrects prior claims by ID (no prior deliverable is edited — research-patches are append-only), and ships a ready-to-dispatch kickoff for the two research tasks this seat could not run.

No design, no shape picked, no fork resolved — that is reserved for the later /arch session.

Three results that change the decision space

1. C10's causal claim is REFUTED by its own falsifier. The uniform digest is not the 2026-05-09 anti-drift machinery. .claude/session-bootstrap.md landed 2026-05-09 (458cd0b52d) as a read-first file with no auto-delivery; the injection hook landed 2026-05-11 (5c0d32ec72, closing "Wave 6 D-1", incident unmentioned); the SubagentStart uniform digest — the exact behaviour C10 defends — landed 2026-06-01 (c2f83e7bc1), 23 days later from orchestrator-gate work. .claude/rules/doc-authority-hierarchy.md:35 explicitly denies the "token-distance forgetting" mechanism that would justify uniform injection, and docs/meta-factory/research-patches/2026-05-16-goal-clarity-dialogue.md:37-41 records the digest amplifying drift. The main stated blocker against per-role shaping has no evidentiary base.

2. C11's always-on measurement is corrected for host. claudeMdExcludes does work on host CC — verified by live self-observation: exactly the 7 non-excluded no-paths: rules were injected, none of the 4 excluded. Host rule load = 90,699 B, not 131,408 B. The aif container measured a genuinely different environment — exactly what the token-audit kickoff's BINDING per-environment-attribution rule predicted. Official docs confirm .claude/rules/** exclusion is supported; this repo's four entries are repo-relative literals where the docs specify absolute globs.

3. A parked BUILD trigger has already fired. docs/meta-factory/research-patches/2026-06-13-inject-layer-extension-rphase.md:79-82 set the DEFER re-trigger at "≥6 marked rules"; today grep -lE '^paths:|^<!-- globs:|<!-- inject:' .claude/rules/*.md | wc -l → 16 of 26 (^paths: alone → 15), against the R-phase's own baseline of 3 of 12 at :75.

Files

• docs/meta-factory/research-patches/2026-07-31-per-role-context-opus-cold-verify.md (new, 433 lines) — the review.
• .claude/orchestrator-prompts/per-role-context-cold-verify/kickoff.md (new) — ready-to-dispatch: (T1) the 8-dimension cold-review that has never run (task 4e73e54e returned a blocker-report because its inputs were absent from the container branch — they are on staging now); (T2) a sweep of the aif-handoff runtime's own agent definitions, the layer nobody has surveyed. Carries a host-verify contract; bash scripts/host-verify.sh --list per-role-context-cold-verify exits 0.
• docs/superpowers/specs/2026-07-27-per-role-context-bundle-for-opus.md (+10) — a STATUS pointer only, so the /arch session does not consume refuted claims unflagged. Body byte-unmodified.

Not dispatched: the aif runtime was down at review time (Docker socket absent; /runtime-profiles and /tasks both HTTP 000). Read-only diagnosis only, per aif-doctor.

§1.7 Forward-check applied

• .claude/rules/recommendation-laziness-discipline.md §3 — no ADOPT/BUILD/REJECT/DEFER verdict is issued and no candidate shape is picked. §4.3 reports that a pre-existing trigger's condition is met — a measurement, explicitly disclaimed in-file as "not a recommendation to build".
• .claude/rules/phase-research-coverage.md §1.7 — every resolution in §1 carries command output or file:line, not prose. The load-bearing new claim at §1 docs: Phase 7.5 — plan formalization (deterministic-v1 stance + v2 backlog + self-diagnostics design) #7 states its own falsifier ("rewrite the four entries as absolute globs; if container behaviour is unchanged, the cause is environmental after all").
• .claude/rules/attention-is-not-a-mechanism.md §1 — §4.2 names the 40,709-byte claudeMdExcludes saving as #hope-as-gate (silent over-injection, zero asserting channels) and surfaces it rather than proposing a gate; proposing one would be design, reserved for /arch.
• .claude/rules/ai-laziness-traps.md §2 — T7: the review resolved contradiction feat: audit fixes Phase 1+2+3 (R2/R11 drift, manifest SSOT, depcruise integration) #1 against the brief's leading hypothesis and downgraded rows 1/5 from "contradiction" to layer confusion. T13: the ADOPTED superpowers substrate was audited locally rather than trusted — DeepWiki, the upstream-authority channel, was found wrong (re-review-prompt.md exists in v6.2.0, 106 lines, per RELEASE-NOTES.md:10). T14: verify-list feat(meta-factory): Phase 5+6 — L2 Research Agent + L3 Synthesizer Path A (deterministic v1) #5's clean result is reported as "confirmed at narrowed scope" plus the one refuting line (.claude/hooks/warn-subagent-report.sh:70-75), not as "clean".
• .claude/rules/doc-authority-hierarchy.md §2-§3 — Authoritative-for / NOT-authoritative-for header present on both new files.
• .claude/rules/destination-environment-verification.md §1 — the kickoff declares a real host-verify contract (substance checks on the two markdown deliverables), not an opt-out; the accuracy half is explicitly routed to Task 1 dimension F rather than faked as a shell gate.
• CLAUDE.md Artifact Ownership Contract — the two GLM patches and three aif task outputs are corrected by ID and left byte-unmodified.

§1.7 Backward-check applied

Class of this change = a cold-verification record that corrects a prior research patch's claims by ID without editing it. Sibling surfaces enumerated (ls docs/meta-factory/research-patches/2026-0[5-7]-* + grep for supersede/REFUTED language) — five of the seven verdicted surfaces are untouched by this diff:

• docs/meta-factory/research-patches/2026-07-27-per-role-context-addendum-fresh-2026.md — SWEPT-CLEAN: it is the immediate precedent for this exact pattern (its §0 corrects the 2026-07-26 patch by claim-ID). This patch extends the chain and supersedes it only on C11's host attribution, stated explicitly rather than silently.
• docs/meta-factory/research-patches/2026-05-16-goal-clarity-dialogue.md:37-41 — GAP CLOSED, not created: it already recorded the digest amplifying drift, and C10 was authored 2026-07-26 in contradiction of it without citing it. Not superseded — the older record was right and had gone unread.
• docs/meta-factory/research-patches/2026-06-13-inject-layer-extension-rphase.md:79-82 + .claude/orchestrator-prompts/inject-layer-extension/done.md:5 — GAP FOUND: a DEFER whose re-trigger has fired. Recorded, not unilaterally reopened — that is the owning umbrella's call.
• docs/meta-factory/prior-art-evaluations.md:122,173 (chore(deps): bump js-yaml and markdownlint-cli2 #50, docs(research): niche-strategy + growth roadmap proposal-of-record #101) — SWEPT, both stale-but-not-wrong: chore(deps): bump js-yaml and markdownlint-cli2 #50 still cites only the background skill-context probe; docs(research): niche-strategy + growth roadmap proposal-of-record #101's "reconcile read-trigger vs edit-trigger" is still open, and per-role markers land on that seam. Append-only SSOT, reviewer-read-only — cited, untouched.
• .claude/orchestrator-prompts/session-start-token-audit/kickoff.md:22-26 (branch, not staging) — CORROBORATED, not contradicted: its claudeMdExcludes claim is the side that survives, and its own per-environment-attribution rule is what predicted the container divergence. Operator-owned; not edited.
• .claude/rules/zcode-parity-doctrine.md §2 row 19 — GAP FOUND: .claude/hooks/warn-subagent-report.sh:69-75 carries the Explore noise-guard but plugin/hooks/warn-subagent-report-zcode has zero agent_type occurrences — the repo's single role branch is missing from its portable twin. Out of scope per the CLAUDE.md PR-strategy rule; surfaced as observation only.
• Sibling 2026-07-* patches (12 files) — all carry §1.7 markers; this patch conforms. SWEPT-CLEAN.

How to verify

bash scripts/host-verify.sh --list per-role-context-cold-verify   # exits 0
grep -lE '^paths:|^<!-- globs:|<!-- inject:' .claude/rules/*.md | wc -l   # 16 — §4.3's fired trigger
git log --format='%ad %h %s' --date=short --diff-filter=A -- .claude/hooks/inject-subagent-digest.sh  # 2026-06-01, §4.1's chronology

Fidelity verdict

FIDELITY: skipped — host-authored Opus review PR, not aif factory output: there is no container diff and no stage kickoff substrate for agents/fidelity-auditor.md to audit a WHAT against. The deliverable is a research record plus a dispatch kickoff, both verifiable directly from the diff; the kickoff's own acceptance contract is declared in-file and bash scripts/host-verify.sh --list per-role-context-cold-verify exits 0.