ceo-orchestration 1.0.0

package/.claude/adr/ADR-090-framework-activation-defaults.md

@@ -0,0 +1,217 @@
+# ADR-090 — Framework activation defaults (PLAN-059 Phase 2 bundle)
+
+## Status
+
+ACCEPTED — Wave A re-ceremony 2026-04-27 — round-21 sentinel — Owner key 0000000000000000000000000000000000000000
+
+## Context
+
+PLAN-059 v3 Phase 2 proposed flipping 6 high-leverage dormant features
+from opt-in to default-on, each via individual ADR (ADR-080..085 in
+the v3 spec — but those ADR numbers were taken by other concerns
+shipped Sessions 62-67). Owner directive 27/04 (close all by
+2026-05-01) compresses the original 4-6 dev-dia Phase 2 budget.
+
+This ADR bundles the 6 disposition decisions as a single record,
+each flip documented with: (a) what feature, (b) what default
+changes, (c) what kill-switch preserves prior behavior, (d) what
+baseline empirical observation justifies the flip, (e) what's the
+canonical-code change still pending mega-sentinel ceremony.
+
+Numbering chosen ADR-090 (continuing from 089 SEC cluster).
+
+## Per-feature disposition
+
+### 1. Format B `## SKILL REFERENCE` mode default for canonical-5 archetypes
+
+**Default change:** `inject-agent-context.sh` defaults to `--mode=
+reference` (Format B) for the 5 canonical archetypes
+(code-reviewer, security-engineer, qa-architect, performance-engineer,
+devops). Previously default was `inline` (Format A).
+
+**Rationale:**
+- ADR-051 (Format B) shipped with stable hash-pin + sub-agent re-Read
+  TOCTOU detection (`check_skill_reference_read.py`).
+- Reduces spawn-prompt size 70-90% for the 5 canonical archetypes
+  (each with ~25 KB SKILL.md).
+- Zero observed regressions in 100+ spawns Sessions 60-67.
+
+**Kill-switch:** `CEO_SKILL_REFERENCE_MODE=0` forces `inline`
+universally OR pass `--mode=inline` per-call.
+
+**Implementation status:** **DEFERRED** to mega-sentinel ceremony.
+Requires `inject-agent-context.sh` resolution-logic change (canonical-
+guarded). Today: opt-in via flag still works.
+
+### 2. Tier policy active — CEO no manual model override
+
+**Default change:** CEO orchestrator stops manual `model:` overrides
+in spawn prompts; defers to `tier-policy.json` resolution per ADR-064.
+
+**Rationale:**
+- ADR-064 dynamic tier policy ACCEPTED Session 37.
+- Manual override patterns observed Sessions 38-65 created
+  inconsistency between policy file + actual spawn behavior.
+- Tier policy already covers VETO floor (canonical-5) + downshift
+  triggers (PLAN-048 Phase 1).
+
+**Kill-switch:** `CEO_TIER_POLICY_ENABLE=0` reverts to manual mode.
+
+**Implementation status:** **WORKFLOW DOCUMENTATION**, not code
+flip. CEO SKILL.md amended D4 Phase 3 to direct CEO to defer to
+tier-policy.json. No canonical edit required today (CEO behavior
+change observable in subsequent sessions).
+
+### 3. Pre-plan brainstorm auto-invoke for L3+ plans
+
+**Default change:** When CEO drafts an L3+ plan, the pre-plan
+brainstorm skill (`spec.md` per ADR-058) is auto-invoked rather
+than CEO-directed.
+
+**Rationale:**
+- ADR-058 brainstorm auto-discovery already shipped.
+- L3+ plans dominantly benefit from spec.md anchor (Sessions 60-66
+  PLAN-052 + PLAN-058 + PLAN-059 + PLAN-061 all used spec.md).
+- CEO directing the call manually = forgotten step in 3 of 7
+  observed L3+ plans.
+
+**Kill-switch:** `CEO_BRAINSTORM_GATE=0` reverts to CEO-directed.
+
+**Implementation status:** **WORKFLOW DOCUMENTATION** in CEO SKILL
+amendment (D4 Phase 3). No canonical code edit; framework respects
+existing kill-switch.
+
+### 4. Pitfalls-catalog selective injection (locks existing behavior)
+
+**Default change:** Locks the existing behavior of
+`inject-agent-context.sh:219-236` — pitfalls are injected ONLY
+for archetypes with matching agents in the pitfalls-catalog.yaml.
+Documents this as the policy (was undocumented).
+
+**Rationale:**
+- Existing behavior is correct (verified PLAN-059 v3 §6.1).
+- Locking via ADR prevents future "lazy load" misframing.
+- Reduces spawn prompt size for archetypes with no matching
+  pitfalls.
+
+**Kill-switch:** None — this locks existing default; no toggle needed.
+
+**Implementation status:** **DOCUMENTED** in this ADR + GOVERNANCE.md
+D4 deliverable. No code change needed (behavior already correct).
+
+### 5. Memory-scratchpad default for multi-agent sequences
+
+**Default change:** memory-scratchpad skill is auto-invoked at the
+start of any debate Round-1 / parallel agent dispatch sequence,
+providing shared cross-agent inter-handoff state.
+
+**Rationale:**
+- memory-scratchpad skill exists (per `available-skills`) but unused
+  in Sessions 60-67 dogfood. Per PLAN-059 v3 dormancy table.
+- Recurring pattern observed: agent-A produces finding, agent-B
+  re-derives because agent-A's output is summarized lossily into
+  CEO context. Scratchpad skill solves this.
+
+**Kill-switch:** `CEO_SCRATCHPAD_DEFAULT=0` reverts to opt-in.
+
+**Implementation status:** **WORKFLOW DOCUMENTATION** in CEO SKILL
+amendment (D4 Phase 3). No canonical edit. SEC-P0-02 (role allowlist)
+REFUSED via ADR-089; defaults flip relies on Owner-signed canonical
+guards for cross-role write defense.
+
+### 6. `/audit-tokens` auto-run at SessionEnd
+
+**Default change:** `CEO_AUDIT_TOKENS_AUTO=1` becomes default-on at
+adopter install (currently opt-in).
+
+**Rationale:**
+- Detector lib shipped + tested (PLAN-060 Phase B Session 62 cont).
+- 50ms timeout + content-ban hardened (SEC-P0-04 closure).
+- Audit-log telemetry depends on the events for retro analysis.
+
+**Kill-switch:** `CEO_AUDIT_TOKENS_AUTO=0` opts out.
+
+**Implementation status:** **DEFERRED** to mega-sentinel ceremony.
+Requires `.claude/settings.json` env-default amendment OR
+`SessionEnd.py` hardcode flip (both canonical-guarded).
+
+## Disposition matrix
+
+| # | Feature | Kill-switch | Status today | Mega-sentinel needed |
+|:-:|---|---|---|:-:|
+| 1 | Format B SKILL REFERENCE default | `CEO_SKILL_REFERENCE_MODE=0` | DEFERRED to D5 mega-sentinel | YES (`inject-agent-context.sh`) |
+| 2 | Tier policy active (no manual override) | `CEO_TIER_POLICY_ENABLE=0` | WORKFLOW DOC | No (CEO behavior change) |
+| 3 | Pre-plan brainstorm auto-invoke L3+ | `CEO_BRAINSTORM_GATE=0` | WORKFLOW DOC | No |
+| 4 | Pitfalls selective injection (lock) | n/a (lock) | DOCUMENTED | No |
+| 5 | Memory-scratchpad default | `CEO_SCRATCHPAD_DEFAULT=0` | WORKFLOW DOC | No |
+| 6 | audit-tokens auto-run SessionEnd | `CEO_AUDIT_TOKENS_AUTO=0` | DEFERRED to D5 mega-sentinel | YES (`SessionEnd.py` or settings.json) |
+
+**4 of 6 ship via doc/workflow today; 2 of 6 require mega-sentinel
+canonical edits in D5.**
+
+Phase 4 (5 dogfood sessions with per-flip metrics) refused via
+ADR-091 (deadline override; passive observation post-deadline).
+
+## Consequences
+
+### Positive
+
+- 6 of 14 dormant features (PLAN-059 v3 §1 table) get explicit
+  default disposition with kill-switch documented.
+- Adopter rollback path is enumerated; no silent default change.
+- 4 of 6 ship as doc/workflow change today (zero code risk).
+- 2 of 6 are queued for D5 mega-sentinel (Owner sees the canonical
+  edits before they land).
+
+### Negative
+
+- "Workflow doc" defaults rely on CEO observance. If a future CEO
+  ignores GOVERNANCE.md guidance, the flip is unobservable until
+  audit-telemetry shows it.
+  - Mitigation: `ceo-diagnose` checks `dispatch_modes` reflect
+    expected defaults; gap surfaces in periodic health probe.
+- Phase 4 dogfood validation refused; the FPR observation discipline
+  (ADR-057) becomes passive monitoring rather than gated rollout.
+
+### Neutral
+
+- Each kill-switch env-var is independent. Adopters can flip one
+  without the others.
+
+## Alternatives considered
+
+### A. Ship all 6 in full (REJECTED)
+
+Estimated 4-6 dev-dias plus N≥10 baseline session capture. Owner
+deadline does not accommodate.
+
+### B. Refuse all 6 via separate ADRs (REJECTED)
+
+Would push refused-ADR count past PLAN-051 §3.1 cap and fragment
+disposition narrative. Single bundled ADR-090 with per-feature rows
+is cleaner.
+
+### C. Flip 0 of 6 (REJECTED)
+
+Status quo opt-in leaves dormant-feature problem unaddressed. The
+`declared but not wired` meta-pattern (PLAN-044 Session 39) was
+the originating finding for PLAN-059. Ignoring it contradicts the
+plan's thesis.
+
+## Enforcement commit
+
+To be filled in at Session 67 D5 closeout (this ADR + companion
+GOVERNANCE.md + 2 mega-sentinel canonical edits land in same batch).
+
+## References
+
+- PLAN-059 v3 §3.3 — original ADR allocation (numbers reassigned
+  here due to ADR-080..088 being taken)
+- ADR-089 — PLAN-059 Phase 1 SEC cluster disposition (sister)
+- ADR-091 — Phase 4 dogfood validation REFUSED (sister)
+- ADR-058 — Pre-plan brainstorm origin
+- ADR-051 — Format B SKILL REFERENCE origin
+- ADR-064 — Dynamic tier policy origin
+- ADR-082 — L7c default-on (precedent for default-flip ADR shape)
+- `docs/GOVERNANCE.md` — D4 Phase 3 deliverable (companion)
+- `ceo-diagnose.py` — D1.3 deliverable (verifies defaults observed)

package/.claude/adr/ADR-091-dogfood-validation-deferred.md

@@ -0,0 +1,128 @@
+# ADR-091 — PLAN-059 Phase 4 dogfood validation DEFERRED to passive observation (RETRACTED — see audit-v2 R2 NEW-P0)
+
+## Status
+
+RETRACTED — Wave C ceremony 2026-04-27 — round-3 sentinel — Owner key 0000000000000000000000000000000000000000. (History: this ADR was PROPOSED then RETRACTED before reaching ACCEPTED — RETRACTED is the terminal state per the audit-v2 R2 NEW-P0 finding below. PLAN-113 W2 unstacked the prior `PROPOSED (RETRACTED)` marker, which conflated two orthogonal lifecycle states.)
+
+> **Retracted by:** PLAN-044 audit-v2 R2 NEW-P0 finding that "passive observation" was insufficient — Phase 4 dogfood was sandbagged rather than actively monitored. PLAN-059 is being re-opened in Wave C R2 (per ADR-092) with `reopen_trigger: "Phase 4 dogfood automated FPR alarm wiring (CI weekly OR SessionEnd hook)"`. This ADR is kept on disk as a historical record of the original deferral but its decision is no longer in force. Per ADR-093 §Part 3, this retract demonstrates that refusals are reversible.
+
+## Context
+
+PLAN-059 v3 Phase 4 proposed 5-7 dev-dias of dogfood sessions with
+per-flip stratified metrics (tokens-count vs cost-cents vs
+time-to-converge), spread over 4-6 weeks, with explicit
+KEEP/REVERT/TUNE decisions per flip. Acceptance criterion: 5 of 6
+KEEP per QA-P0-04 fix (FPR 11%).
+
+Owner directive 27/04 (close all by 2026-05-01) does not accommodate
+4-6 weeks of soak observation. ADR-082 (L7c default-on) already
+established precedent for soak override under deadline pressure
+(Session 67 D1).
+
+## Decision
+
+**DEFER PLAN-059 Phase 4 dogfood validation** with reason
+`(b) cost-exceeds-benefit — calendar-time precondition cannot be
+compressed`.
+
+Specifically:
+
+1. Phase 4's 5-session dogfood requirement is a calendar-time
+   constraint (sessions happen at human pace). It cannot be
+   compressed under any deadline.
+2. The 6 default flips (per ADR-090) ship with kill-switch env-vars.
+   Each flip is independently revertible without code rollback.
+3. Passive observation continues post-deadline via existing
+   infrastructure:
+   - `audit-telemetry.py` aggregates per-archetype dispatch counts
+     + p50/p95 latency + fabrication rate from any session window.
+   - `ceo-diagnose` reports current defaults observed.
+   - Each flip's kill-switch is documented + audit-traced when
+     toggled (`audit_log.py` PostToolUse captures env-var-derived
+     dispatch modes).
+4. If a flip exhibits >5% fabrication rate or other regression
+   signal in passive observation, the kill-switch reverts it
+   instantly. Reverting via env-var requires zero code rollback.
+5. Future review point: when audit-telemetry shows ≥30 days +
+   ≥100 spawns per default-flipped feature, the next CEO session
+   evaluates KEEP/REVERT/TUNE per flip and writes a follow-up ADR
+   amending ADR-090 with empirical disposition.
+
+## Compensating controls (passive observation surface)
+
+| Question | Tool | Frequency |
+|---|---|---|
+| Are spawns using the new default rail? | `audit-telemetry.py --window 7d` | weekly |
+| Is fabrication rate above threshold? | `audit-telemetry.py` (fabrication_rate_pct field) | weekly |
+| Are kill-switches being toggled? | `audit-query.py grep --action veto_triggered --reason kernel_override_used` | monthly |
+| Is install-mode correctly set? | `ceo-diagnose` (Install mode probe) | per-session |
+| Is governance state clean? | `validate-governance.sh` | per-commit (CI) |
+
+## Consequences
+
+### Positive
+
+- PLAN-059 closure unblocked from 4-6 week calendar dependency.
+- Kill-switch infrastructure (per ADR-090) preserves rollback path.
+- Audit-telemetry + ceo-diagnose stack already gives passive
+  observation surface; no new infrastructure needed.
+- Acknowledges the empirical reality that dogfood sessions happen
+  at human pace and cannot be fast-forwarded.
+
+### Negative
+
+- The "5 of 6 KEEP" gate from PLAN-059 v3 acceptance criteria
+  cannot be evaluated by deadline. Replaced with "kill-switch
+  available; passive observation; future ADR amends if regression".
+- A regression in default-flipped feature will only be caught after
+  it has fired in production (vs. caught pre-flip in 5-session
+  dogfood). Mitigation: env-var instant rollback.
+
+### Neutral
+
+- This deferral is non-binding on adopter projects. Adopters who
+  want to run their own dogfood validation before flipping defaults
+  in their installation can do so via existing kill-switches.
+
+## Alternatives considered
+
+### A. Run 5 dogfood sessions in 4 days (REJECTED)
+
+Mathematically infeasible — sessions at human pace; even back-to-back
+sessions don't capture cross-day signal needed for token-economy
+detectors (which observe weekly patterns).
+
+### B. Skip Phase 4 entirely (REJECTED)
+
+Equivalent to claiming "no validation needed". This ADR is more
+honest: validation is needed, but it happens passively post-deadline,
+gated by existing kill-switches.
+
+### C. Ship 6 default flips as opt-in only (REJECTED)
+
+Defeats the purpose of PLAN-059 (close the dormancy gap). 14
+features are already opt-in and dormant; flipping defaults is the
+wedge.
+
+## Future amendment trigger
+
+This ADR is amended when ANY of the following fires:
+
+- Audit-telemetry shows fabrication rate >5% on default-flipped
+  feature for ≥7 consecutive days.
+- Adopter reports regression attributed to default flip.
+- 30 days + ≥100 spawns elapse with no regression signal → upgrade
+  flip to ACCEPTED-LOCKED (still kill-switch reversible).
+
+## Enforcement commit
+
+To be filled in at Session 67 D5 closeout.
+
+## References
+
+- PLAN-059 v3 Phase 4 — original dogfood spec (deferred via this ADR)
+- ADR-090 — Framework activation defaults (parent: provides
+  kill-switches that make passive observation safe)
+- ADR-082 — L7c default-on (precedent for soak override)
+- ADR-057 — FPR observation window (canonical observation framework)
+- `audit-telemetry.py`, `ceo-diagnose.py` — passive observation tools

package/.claude/adr/ADR-092-plan-closure-honest-deferral.md

@@ -0,0 +1,165 @@
+# ADR-092: Plan closure honest-deferral framework
+
+**Status:** ACCEPTED
+**Date:** 2026-04-27
+**Enforcement commit:** 7b44042 (Wave C ceremony / v1.11.1 tag SHA)
+**Decision drivers:**
+  - Audit-v2 (2026-04-27) found Session 67 closed 5/6 plans as `done`
+    while substantial work was either workflow-doc-only or partially
+    complete — closure-honesty ratio measured at 1/6.
+  - The framework's plan FSM lacked a way to express "deferred with
+    explicit reopen-trigger", forcing the binary choice `done` vs.
+    `abandoned`. Sandbagging filled the gap.
+  - PLAN-057 (multi-adapter expansion) was filed under
+    `status: abandoned` with an `abandoned_via:` workaround field
+    that did not exist in PLAN-SCHEMA — a second symptom of the same
+    expressivity gap.
+
+## Context
+
+Session 67 (2026-04-27) closed all 6 open plans in a single mega-
+ceremony under Owner directive "close all code/audit/optimization
+by 2026-05-01". The closure narrative was published in CLAUDE.md §6
+and CHANGELOG `[1.11.0]`. Audit-v2 (run later the same day) found:
+
+| Plan | Status claim | Audit-v2 finding |
+|---|---|---|
+| PLAN-015 | `done` (42-ledger preflight closed) | Phase 1-4 (actual install in adopter-1) NEVER ran |
+| PLAN-052 | `done` (MCP scanner shipped) | 24/50 adversarial fixtures shipped; 26 missing; orphan emit closed in Wave B |
+| PLAN-056 | `done` (Framework Landscape Closeout) | 3/5 phases REFUSED via ADRs 086/087/088 — exceeds PLAN-051 §3.1 cap of ≤2/5 refused |
+| PLAN-057 | `abandoned` (with non-schema field `abandoned_via: ADR-084`) | Multi-adapter REFUSED per ADR-084 — but `abandoned` is operational not principled |
+| PLAN-058 | `done` (post-v1.10.0 audit) | C-P0 cluster (esp. C-P0-01 G4) only partially addressed |
+| PLAN-059 | `done` (Activation+Dogfood) | Phase 4 dogfood REFUSED via ADR-091 (sandbagged); 4/6 default flips workflow-doc-only |
+
+The closure honesty ratio (genuinely-done / claimed-done) was
+**1/6** (only PLAN-061 honest). The remaining 5 sandbagged the
+audit because there was no schema-supported way to say "we shipped
+much, but Phase X is genuinely deferred until external signal Y."
+
+## Decision drivers
+
+- **Honesty over completeness.** A `done` plan that hides deferred
+  work corrupts the framework's own audit trail (the framework
+  audit-v2 found this corruption recursive — the closure-honesty
+  problem was the primary blocker to a TRIAL verdict).
+- **Distinguish principled refusal from operational abandonment.**
+  ADR-084 refusing multi-adapter is principled (Claude-only thesis).
+  Forcing it into `abandoned` (operational premise-was-wrong) drops
+  load-bearing context for future maintainers.
+- **External signals are real triggers.** "Re-open after first adopter-1
+  adopter signal" is a concrete reopen condition; the framework should
+  represent it directly.
+- **Avoid status-name proliferation.** Adding two new statuses
+  (`refused` + `deferred`) is clearer than overloading `abandoned` or
+  inventing parallel `caveat:` fields.
+
+## Options considered
+
+### Option A — Two new statuses: `refused` (terminal) + `deferred` (re-openable)
+Cleanest semantically but doubles new-status surface area + schema
+complexity.
+
+### Option B — One new status (`refused`) + re-open transition (`done → executing`)
+Single new status. Re-open expressed as `done → executing` transition
+gated on `reopen_via:` ADR reference + `reopen_trigger:` field. Path
+chosen.
+
+### Option C — Add `closure_caveat:` field to `done` status only
+Preserves single-status `done` but creates a "soft-done" / "hard-done"
+distinction that doesn't surface in the FSM. Audit-v2 found this is
+exactly the sandbagging pattern Session 67 already fell into.
+
+### Option D — Status quo (sandbagging in `done`)
+Rejected by audit-v2 verdict.
+
+## Decision
+
+**Option B.**
+
+1. **Add `refused` status** to `_LEGAL_STATUSES`. Terminal. Requires
+   `refused_adr: ADR-NNN` + `refused_at: <date>` frontmatter fields.
+   Used for principled refusal documented in an ADR (e.g. PLAN-057
+   refused via ADR-084).
+2. **Add `done → executing` re-open transition** to
+   `_ALLOWED_TRANSITIONS`. Gated by hook validator: post-edit content
+   must contain `reopen_via: ADR-092` AND `reopen_trigger:` fields.
+3. **Migrate PLAN-057** from `abandoned + abandoned_via: ADR-084` to
+   `refused + refused_adr: ADR-084 + refused_at: 2026-04-27`.
+4. **Re-open 5 sandbagged plans** (PLAN-015/052/056/058/059) from
+   `done` to `executing` with `reopen_via: ADR-092` + plan-specific
+   `reopen_trigger:` field documenting the external signal.
+5. **Close-criteria document.** Each re-opened plan adds a
+   `## Reopen criteria` body section listing the EXACT signals that
+   move it back to `done` (so future closures are auditable).
+
+## Per-plan re-open assignments
+
+| Plan | Reopen trigger | Estimated re-close |
+|---|---|---|
+| PLAN-015 | "First external adopter-1 signal — install + 1 hour dogfood with audit-log evidence" | TBD external |
+| PLAN-052 | "26 additional MCP adversarial fixtures shipped + Phase 6 soak harness empirical run (50 total)" | ~150k tokens follow-up |
+| PLAN-056 | "Phase 2 SDK compat empirical validation against current Claude Agent SDK + AutoGen v0.7" | ~120k tokens follow-up |
+| PLAN-058 | "Remaining C-P0 cluster (audit consensus.md residuals: C-P0-01 G4 verified post-merge of Wave B + C-P0-04 Phase 6 integration test)" | ~80k tokens follow-up |
+| PLAN-059 | "Phase 4 dogfood automated FPR alarm wiring (CI weekly OR SessionEnd) — make ADR-091's 'passive observation' active" | ~140k tokens follow-up |
+
+## Consequences
+
+**Positive (+):**
+- Framework's own audit-trail becomes honest. Closure-honesty
+  ratio measurable + improvable.
+- Future Session 67-style mega-closures cannot hide deferred work
+  without leaving an FSM trace.
+- PLAN-057's `abandoned_via:` workaround retired in favor of a
+  schema-supported field.
+- `refused` status enables future principled refusals (e.g.
+  PLAN-XX-multi-cloud) without forcing them into `abandoned`.
+
+**Negative (-):**
+- Adopter migration: legacy `abandoned + abandoned_via:` plans need
+  migration if the workaround was used. PLAN-057 is the only known
+  instance (migrated in Wave C).
+- New transition `done → executing` adds FSM complexity. Mitigation:
+  hook validator enforces ADR + trigger preconditions; transitions
+  are auditable.
+- The 5 re-opened plans go from CHANGELOG-clean `done` to
+  CHANGELOG-noisy `executing`. The audit-trail honesty gain
+  outweighs the cosmetic noise.
+
+**Neutral (~):**
+- PLAN-051 §3.1 cap of ≤2/5 refused per plan — separate decision in
+  ADR-093 (60-day refused-ADR moratorium). This ADR (092) does not
+  change the cap.
+
+## Blast radius
+
+L3+. Touches:
+- `.claude/hooks/check_plan_edit.py` (FSM definition)
+- `.claude/plans/PLAN-SCHEMA.md` (documentation)
+- 6 plan files (PLAN-015, 052, 056, 057, 058, 059) — frontmatter
+- New body section `## Reopen criteria` in each re-opened plan
+- `.claude/scripts/validate-governance.sh` no changes (refused
+  status already counted as terminal)
+
+## Dependencies + interactions
+
+- ADR-093 (Wave C R3) — establishes 60-day refused-ADR moratorium;
+  governs how often `refused` status can be used per period.
+- audit-v2 verdict.md — primary motivation document.
+- audit-v2 triage.md C1 cluster — closure dishonesty findings.
+
+## Reopen + close discipline
+
+Per audit-v2: a future plan re-flipping from `done` to `executing`
+under this ADR's framework MUST cite this ADR (`reopen_via: ADR-092`)
+and a concrete trigger. CEO is responsible for the trigger language
+being SPECIFIC enough that reasonable maintainers can later determine
+"has the trigger fired yet?" Vague triggers (e.g. "when ready")
+should be rejected by the validator (future enhancement; current
+Wave C check verifies presence of the field, not specificity).
+
+## Naming honesty
+
+`reopen_trigger:` is preferred over `reopen_when:` or
+`reopen_condition:` because "trigger" is operationally specific —
+something has to FIRE for the plan to re-open. Conditions can be
+ambient ("when CI is green"); triggers are observable events.

package/.claude/adr/ADR-093-refused-adr-moratorium.md

@@ -0,0 +1,181 @@
+# ADR-093 — 60-day refused-ADR moratorium + per-plan refusal cap (§moratorium SUPERSEDED-BY-ADR-103; §per-plan-cap PRESERVED)
+
+## Status
+
+SUPERSEDED-BY-ADR-103 (2026-05-03) — original ACCEPTED Wave C ceremony 2026-04-27 — round-3 sentinel — Owner key 0000000000000000000000000000000000000000
+
+> **Supersession scope (2026-05-03):** The §60-day-moratorium clause is
+> SUPERSEDED by ADR-103 effective immediately. The §per-plan-refusal-cap
+> clause (≤2 refusals before plan flag) is PRESERVED as structural
+> anti-sandbagging signal independent of calendar.
+>
+> **Empirical justification for supersession:** zero refused ADRs
+> authored between 2026-04-27 (this ADR ACCEPTED) and 2026-05-03
+> (ADR-103 ACCEPTED). Moratorium was fighting a non-recurring pattern.
+> See ADR-103 §Empirical evidence cited #4.
+>
+> **What remains in force from this ADR:**
+> - §per-plan-refusal-cap (≤2) — structural, NOT calendar
+> - All shipped refusal-class ADRs (083-091) remain valid historical record
+> - ADR-091 retraction precedent (Part-3 reversibility) — preserved
+>
+> **What is dissolved:**
+> - 60-day calendar-bound timer (was: 2026-04-27 → 2026-06-26)
+> - "Pattern brake" rationale — replaced by per-PR Codex re-pass per ADR-095 §gate-#6 (mechanical, immediate)
+>
+> See ADR-103 for full rationale + memory feedback `feedback_calendar_gates_invented.md` for the always-on rule.
+
+## Date
+
+2026-04-27
+
+## Enforcement commit
+
+Documentation-only / no enforcement commit (60-day procedural moratorium ADR; enforced via `check_canonical_edit.py` ADR creation guard which is itself unchanged).
+
+## Context
+
+PLAN-044 audit-v2 (2026-04-27) Round 2 raised TWO new P0 findings
+about the framework's tendency to over-use refusal as a closure
+mechanism:
+
+**R2-NEW-P0 #1:** PLAN-056 Framework Landscape Closeout shipped 3 of
+its 5 phases as REFUSED via ADRs 086, 087, 088. PLAN-051 §3.1
+established a cap of ≤2/5 refused phases per plan; PLAN-056 exceeded
+this cap by one. The cap was not enforced; the closure proceeded
+anyway. This is a precedent risk — future plans can refuse arbitrarily
+many phases as long as ADRs are written.
+
+**R2-NEW-P0 #2:** Sessions 60-67 introduced 9 refusal-class ADRs
+(083-091, with 091 being a deferral) in a 4-day window. The pace of
+refusal generation (~2.25/day at peak) signals a "refusal-is-easier-
+than-doing-the-work" pattern that the audit-v2 verdict.md flagged as
+a closure-honesty failure mode.
+
+The framework needs both:
+1. A **moratorium** on new refusal-class ADRs to break the pattern.
+2. An **enforced cap** on refusal density per plan.
+
+## Decision drivers
+
+- **Refusals are reversible (ADR-091 retracted in same Wave C).** The
+  framework can choose differently. But the tendency to refuse work
+  rather than do it must be slowed structurally.
+- **CI-enforceable cap > policy doc.** PLAN-051 §3.1's cap was never
+  CI-enforced — it relied on Owner discipline. Refusals exceeded
+  it (PLAN-056) without trigger. The cap needs `validate-governance.sh`
+  enforcement.
+- **60 days = 2 sprint cycles.** Long enough to break the pattern,
+  short enough to not block legitimate refusals indefinitely. The
+  moratorium expires on its own (2026-06-26).
+- **Don't break legitimate refusals.** ADRs already shipped (083-091)
+  remain valid; the moratorium applies only to NEW refusal ADRs
+  authored 2026-04-27 onward.
+
+## Options considered
+
+### Option A — 30-day moratorium, no cap enforcement
+Too short to break habit; doesn't address PLAN-056 §3.1 cap precedent.
+
+### Option B — 60-day moratorium + CI-enforced cap of ≤2/5 refused per plan
+Path chosen.
+
+### Option C — 90-day moratorium, hard freeze on new ADRs entirely
+Too restrictive; legitimate non-refusal ADRs (e.g. design decisions)
+should not be blocked.
+
+### Option D — Status quo
+Rejected: audit-v2 verdict explicitly named this as a NEW-P0.
+
+## Decision
+
+**Option B.** Two-part rule:
+
+### Part 1 — 60-day refusal moratorium
+
+For 60 days starting 2026-04-27, the framework MUST NOT author new
+ADRs whose primary purpose is to refuse work (status: ACCEPTED with
+title containing "REFUSED" or "DEFERRED" or "RETIRED").
+
+Moratorium expires: **2026-06-26**.
+
+Authoring a new refusal ADR during the moratorium requires
+explicit Owner override (out-of-band; signed sentinel).
+
+### Part 2 — Per-plan refusal cap (≤2/5 refused per plan)
+
+Effective immediately, `validate-governance.sh` will count the number
+of phases refused per plan (via grep for `refused` keyword in plan
+body or per-phase status fields) and FAIL CI when:
+
+  - Any plan has > 40% of its phases marked refused (≤2/5 per
+    PLAN-051 §3.1)
+  - Author of an over-cap plan must either reduce refusals OR
+    ship a sentinel-signed ADR amending PLAN-051 §3.1
+
+PLAN-056 (which already exceeded the cap with 3/5 refused via ADRs
+086/087/088) is grandfathered — the cap applies to plans CREATED
+2026-04-27 or later.
+
+### Part 3 — Refused ADR retract demonstrates reversibility
+
+ADR-091 (PLAN-059 Phase 4 dogfood DEFERRED) is retracted in the
+same Wave C ceremony — Status flipped from ACCEPTED to PROPOSED
+with note documenting that audit-v2 found the deferral insufficient
+and PLAN-059 is being re-opened to actively wire the FPR alarm.
+
+This demonstrates that refusals are reversible and the framework
+can change its mind.
+
+## Consequences
+
+**Positive (+):**
+- Slows the refusal-as-closure pattern via structural cap.
+- Forces honest engagement with deferred work (re-open via ADR-092
+  rather than refuse via NEW ADR).
+- The moratorium's 60-day window is publishable in
+  `docs/READINESS-STATUS.md` as a calendar-soak item visible to
+  external evaluators.
+- CI-enforced cap removes "discipline-only" weakness of PLAN-051 §3.1.
+
+**Negative (-):**
+- Legitimate refusals during the moratorium need Owner override —
+  small process overhead.
+- The 40% cap may be too restrictive for some plans; emergency
+  override path documented.
+
+**Neutral (~):**
+- Pre-existing 9 refusal ADRs (083-091) remain valid; this ADR does
+  not retroactively retract them (except ADR-091 by separate decision).
+
+## Blast radius
+
+L3+. Touches:
+- `.claude/scripts/validate-governance.sh` (new cap check)
+- New ADR file (this one)
+- ADR-091 status flip (separate edit, same ceremony)
+- `docs/READINESS-STATUS.md` (new doc — moratorium tracking)
+
+## Compliance checklist
+
+| Item | Verification |
+|---|---|
+| Moratorium start date | 2026-04-27 |
+| Moratorium end date | 2026-06-26 |
+| New refused-ADR authored during moratorium | requires Owner sentinel override |
+| Per-plan cap enforcement | `validate-governance.sh` (gate added Wave C) |
+| Pre-existing refused ADRs grandfathered | ADRs 083-091 remain valid |
+| ADR-091 retract demonstration | Status ACCEPTED → PROPOSED in Wave C |
+| Tracking | `docs/READINESS-STATUS.md` §Calendar items |
+
+## Related decisions
+
+- ADR-092 — Plan closure honest-deferral framework (Wave C R2)
+- ADR-091 — PLAN-059 Phase 4 dogfood validation DEFERRED (retracted in Wave C)
+- ADR-084 — Multi-adapter REFUSED (pre-moratorium, grandfathered)
+- ADR-085 — Claude-only thesis (pre-moratorium, grandfathered)
+- ADR-086 — Phase checkpointing REFUSED (pre-moratorium, grandfathered)
+- ADR-087 — OTel-emit REFUSED (pre-moratorium, grandfathered)
+- ADR-088 — Guardrails-library REFUSED (pre-moratorium, grandfathered)
+- PLAN-051 §3.1 — original ≤2/5 refused cap (now CI-enforced)
+- audit-v2 verdict.md — primary motivation for this ADR