Delegation detail-fidelity + PM-loop hardening (#541)

* feat(gateway): delegation detail-fidelity — details survive hand-off, both directions

Details thinned out at every delegation hop: a PM child task mapped to no
parent criterion was legal (coverage only surfaced at submit_up, after the
whole wave ran — a 12-subtask docs tree grew through 8 review rounds that
way, one child titled 'docs page and route wrapper' shipping only the
page), and QA could pass work on a gestalt read (a 4-scene video brief
shipped 3 scenes past every gate because the features existed only in
prose). Three chokepoint gates:

- delegate (down): every child must declare covers_parent_criteria
  resolving against the parent's real acceptance criteria — no mapping or
  an unresolvable ref rejects naming every offending child and the valid
  criteria; the success envelope carries parent_ac_coverage
  {covered, uncovered} so a wave-planning PM sees remaining gaps in the
  same turn. Full coverage stays enforced at submit_up (waves stay legal).
- pass_review (up): mandatory criteria_verified — one {criterion,
  evidence} entry per task AC, matched by the findings ledger's
  id-or-exact-text matcher, evidence soup-checked and capped; rejects
  naming the unverified criteria; entries render deterministically into
  qa_notes as '[AC] <criterion> — verified: <evidence>' lines. The old
  count-only ac_verdicts gate is superseded (arg kept for back-compat).
- video briefs (structured detail at origination): an enumerable feature
  list (release highlights, or input_props.highlights carried onto a
  reject re-author) becomes its own scene acceptance criterion, bounded to
  the AC caps; a re-author without highlights carries the
  feedback-addressed criterion instead.

Extracted findings.py's criterion matcher into shared unmatched_criteria /
uncovered_acceptance_criteria instead of duplicating it; criteria_verified
joins the WAF free-text exclusion set like findings/issues.

* fix(gateway): break the block/unblock wedge — four hardening fixes from the live PM loop

A cell task looped fe-pm/main-pm block/unblock for hours (10 cycles, 43
spawns): a transient GitHub API error resolving CI became an unwaivable
blocker finding whose own fix text said no code change was required, the
submit freshness guard then demanded a commit no finding called for,
escalate_up auto-blocked, and main-pm's correct recovery plan 422'd on
the approach length cap, degrading it to a bare unblock. Four fixes:

- pr_pass CI-unresolvable refusal is now explicitly transient-worded:
  retry pr_pass shortly, do NOT pr_fail over a CI-status lookup error —
  a platform blip is not a code finding
- submit freshness guard grants ONE unchanged-head resubmission per
  head sha when the findings ledger has zero open rows (all addressed
  without code changes) — stamped via the resubmit_unchanged_head
  marker so the same head can never loop a second time
- unblock carries a flip breaker: block_flip_count marker, and at the
  third flip a one-shot CEO notification flags the task as structurally
  wedged (unblock itself still succeeds — the breaker signals, it does
  not wedge recovery)
- i_will_plan's approach cap truncates at 800 chars instead of
  rejecting — an over-detailed plan must never cost the PM its turn

---------

Co-authored-by: Renn F <rennf93@users.noreply.github.com>

CLAUDE.md

@@ -418,6 +418,8 @@ Agent backends are pluggable. `roboco/llm/providers/` defines an `AgentProvider`
 
 **Reviewer/PM collision map (always-on).** The collision surface authored at delegate time (`intends_to_touch` / `adds_migration` / `touches_shared`) used to be consumed once by `SequencingService` to wire dependency edges and never shown to a reviewer again. `build_collision_context` (`roboco/services/gateway/choreographer/collision.py`, pure — no DB/IO) now surfaces it for a task under review: same-parent siblings that would collide — overlapping declared file globs, or both adding a migration (the Alembic-head collision needs no file overlap) — rendered with the overlapping globs and, where the caller hands real touched files, a declared-vs-actual drift flag. Capped at 10 siblings / 5 globs. The same builder feeds QA's `claim_review` evidence, the PR-gate's `claim_gate_review` evidence, the PM's `i_will_plan` planning briefing (no drift there — no work yet), and the panel's `GET /api/tasks/{id}/collision-map` route backing a Collision tab on the task detail page.
 
+**Delegation detail-fidelity (always-on, 2026-07-16).** Details no longer thin out at hand-off in either direction. DOWN: `delegate` refuses any child that doesn't declare `covers_parent_criteria` mapping onto the parent's real acceptance criteria (matched by id or exact text; an unresolvable ref is rejected naming the valid criteria, never silently dropped — previously the mapping was optional and coverage surfaced only at `submit_up`'s roll-up gate, after the whole wave had already run); the success envelope carries `parent_ac_coverage` `{covered, uncovered}` so a wave-planning PM sees remaining gaps in the same turn, while multi-wave planning stays legal. UP: `pass_review` requires `criteria_verified` — one `{criterion, evidence}` entry per task acceptance criterion (the findings ledger's id-or-exact-text matcher, soup-checked and length-capped evidence), rejecting with the unverified criteria named; entries render deterministically into `qa_notes` as `[AC] <criterion> — verified: <evidence>` lines, so a gestalt "looks good" pass is structurally impossible. Video briefs stopped being prose-only: an enumerable feature list (release `highlights`, or `input_props.highlights` carried onto a reject re-author) becomes its own acceptance criterion ("Every brief-named feature appears as its own fully readable scene: …", bounded to the AC caps; a re-author without highlights carries "every point in the CEO rejection feedback is visibly addressed"), so the dropped-scene class — a four-feature brief shipping three scenes past every gate — is caught by the QA per-AC stamp instead of the CEO's eyeball.
+
 **PR labeler (always-on).** `derive_pr_labels` (`roboco/foundation/policy/pr_labels.py`, pure) derives the org-structure label vocabulary every fleet PR now carries: `to master` vs `to slave` (today, an assembled root→master PR is the only `to master` case — real slave-branch targeting lands with the slave/master fleet wiring), `root` for an assembled root PR, `MegaTask` for a batch-carrying task, and a layer label (`main-pm` for a Main-PM coordination root, `cell/{team}` for a cell-assembled PR, else `subtask/{team}` for a leaf dev PR). Applied best-effort at all three PR-opening sites in `GitService` so a human triaging the PR queue sees which tree and which org layer a PR belongs to at a glance.
 
 **Feature flags / company-in-a-box.** Env-gated, default-off subsystems toggle from the panel's Settings → Feature Flags card (`panel/src/components/settings/feature-flags-card.tsx`) instead of hand-editing env: web research (`ROBOCO_RESEARCH_ENABLED`), the strategy engine (`ROBOCO_STRATEGY_ENGINE_ENABLED`), pitch provisioning (`ROBOCO_PROVISIONING_*`), external / internal PR review, the agent-runtime toolchain match (`ROBOCO_TOOLCHAIN_MATCH_ENABLED`), the architectural-conventions standard (`ROBOCO_CONVENTIONS_ENABLED`), gateway-health recovery (`ROBOCO_GATEWAY_HEALTH_ENABLED`), multi-repo CI-watch (`ROBOCO_CI_WATCH_ENABLED`), the dependency-update bot (`ROBOCO_DEP_UPDATE_ENABLED`), the gated release manager (`ROBOCO_RELEASE_MANAGER_ENABLED`), the organizational memory loop (`ROBOCO_ORG_MEMORY_ENABLED`), the sandboxed dev DB/Redis (`ROBOCO_SANDBOX_DB_ENABLED`), the RoboCo X account (`ROBOCO_X_ENGINE_ENABLED`), the RoboCo video engine (`ROBOCO_VIDEO_ENGINE_ENABLED`), the board roadmap engine (`ROBOCO_ROADMAP_ENGINE_ENABLED`), Fable-mode (`ROBOCO_FABLE_MODE_ENABLED`), the vault weekly report + KB ingest (`ROBOCO_VAULT_REPORT_ENABLED` / `ROBOCO_VAULT_KB_ENABLED`), the env-sync cascade (`ROBOCO_ENV_SYNC_ENABLED`), the Telegram notifications bridge (`ROBOCO_TELEGRAM_ENABLED`), the possibilities matrix (`ROBOCO_POSSIBILITIES_MATRIX_ENABLED`), the docs-divergence sync (`ROBOCO_DOCS_SYNC_ENABLED`), and the self-heal flags above. Cloud auth (`ROBOCO_CLOUD_AUTH_ENABLED`) is deliberately NOT on this card — like `ROBOCO_DB_NETWORK_ISOLATED`, it's a compose/env-coupled flag a runtime toggle can't safely flip mid-session. A toggle persists in the settings store and takes effect on the next backend restart; an unset flag falls back to its environment / config default.

agents/prompts/_generated/qa.md

@@ -12,7 +12,7 @@
 | `give_me_work` | `give_me_work()` |
 | `i_am_blocked` | `i_am_blocked(task_id: UUID, reason: str, blocker_type: str | None = None, what_needed: str | None = None)` |
 | `i_am_idle` | `i_am_idle()` |
-| `pass_review` | `pass_review(task_id: UUID, notes: str, ac_verdicts: list[str] | None = None)` |
+| `pass_review` | `pass_review(task_id: UUID, notes: str, ac_verdicts: list[str] | None = None, criteria_verified: list[str | Any] | None = None)` |
 | `resume` | `resume(task_id: UUID)` |
 | `unclaim` | `unclaim(task_id: UUID)` |
 

agents/prompts/_generated/verbs.md

@@ -57,7 +57,7 @@ real tools live in their agent_sdk drivers, not role_config.
 | `give_me_work` | `give_me_work()` |
 | `i_am_blocked` | `i_am_blocked(task_id: UUID, reason: str, blocker_type: str | None = None, what_needed: str | None = None)` |
 | `i_am_idle` | `i_am_idle()` |
-| `pass_review` | `pass_review(task_id: UUID, notes: str, ac_verdicts: list[str] | None = None)` |
+| `pass_review` | `pass_review(task_id: UUID, notes: str, ac_verdicts: list[str] | None = None, criteria_verified: list[str | Any] | None = None)` |
 | `resume` | `resume(task_id: UUID)` |
 | `unclaim` | `unclaim(task_id: UUID)` |
 

roboco/api/routes/v1/flow_qa.py

@@ -75,7 +75,7 @@ async def qa_pass(
     choreographer: _ChoreographerDep,
 ) -> dict:
     env = await choreographer.pass_review(
-        x_agent_id, body.task_id, body.notes, body.ac_verdicts
+        x_agent_id, body.task_id, body.notes, body.ac_verdicts, body.criteria_verified
     )
     return envelope_to_response(env, request)
 

roboco/api/schemas/v1/flow.py

@@ -26,6 +26,10 @@ StrList = Annotated[list[str], BeforeValidator(coerce_str_list)]
 _AC_MAX_ITEMS = 7
 _AC_MAX_ITEM_CHARS = 200
 
+# approach's truncate-not-reject ceiling — must match _PM_APPROACH_MIN_LEN's
+# sibling ceiling in choreographer._impl (the gate re-derives its own copy).
+_APPROACH_MAX_CHARS = 800
+
 
 class SubTaskCreate(BaseModel):
     """A PM sub_task — a delegate target AND a progress-checklist item.
@@ -201,9 +205,19 @@ class PassReviewRequest(BaseModel):
     ac_verdicts: StrList | None = Field(
         default=None,
         description=(
-            "One verification entry per acceptance criterion (in criterion "
-            "order) stating how QA verified it. Every criterion must be "
-            "covered before a pass is allowed."
+            "Legacy free-text per-criterion verdicts — still folded into the "
+            "persisted notes but no longer gates the pass (see "
+            "criteria_verified)."
+        ),
+    )
+    criteria_verified: list[dict[str, Any]] | None = Field(
+        default=None,
+        description=(
+            "Mandatory when the task has acceptance criteria — one "
+            "{criterion, evidence} entry per criterion. criterion must match "
+            "an AC by id or exact text; evidence must be concrete (file:line, "
+            "screenshot ref, rendered-frame path, test name). Every criterion "
+            "must be covered or the pass is rejected."
         ),
     )
 
@@ -348,9 +362,11 @@ class IWillPlanRequest(BaseModel):
     # min_length must match choreographer._impl._PM_APPROACH_MIN_LEN. Raised
     # 20→150: a 20-char approach was a one-liner; the approach +
     # sub_tasks are also the progress checklist, so they must be substantive.
-    # max_length caps the bloat defect (approach that ran to thousands of
-    # chars restating the description). Must match the gate ceiling.
-    approach: str = Field(..., min_length=150, max_length=800)
+    # Overflow past 800 chars is truncated, not rejected — a 422 here used to
+    # throw away an otherwise-good plan and degrade the PM to a bare unblock
+    # (live incident: an over-length approach 422'd mid-flow). min_length
+    # stays a hard reject: a thin plan IS a defect.
+    approach: str = Field(..., min_length=150)
     sub_tasks: list[SubTaskCreate] = Field(
         default_factory=list,
         description="List of {title, description} — server assigns id + order",
@@ -359,6 +375,13 @@ class IWillPlanRequest(BaseModel):
     risks: list[RiskCreate] = Field(default_factory=list)
     open_questions: list[OpenQuestionCreate] = Field(default_factory=list)
 
+    @field_validator("approach", mode="before")
+    @classmethod
+    def _truncate_approach(cls, v: object) -> object:
+        if isinstance(v, str) and len(v) > _APPROACH_MAX_CHARS:
+            return v[: _APPROACH_MAX_CHARS - 3] + "..."
+        return v
+
 
 class DelegateRequest(BaseModel):
     """HTTP body for cell_pm + main_pm `delegate` verbs.

roboco/foundation/policy/content/markers.py

@@ -460,3 +460,62 @@ def get_docs_sync_release_version(task: HasMarkers) -> str | None:
 
 def set_docs_sync_release_version(task: HasMarkers, version: str) -> None:
     set_marker(task, DOCS_SYNC_RELEASE_VERSION, version)
+
+
+# --- resubmit-unchanged-head exemption -------------------------------------
+# submit_root/submit_up hard-refuse a re-submit whose assembled PR head is
+# unchanged since the last pr_fail (the 2026-06-27 loop-stopper). That's a
+# structural deadlock when the rejection round needed no code change (e.g. a
+# transient CI-lookup error) — every ledger finding gets addressed/waived but
+# the head sha never moves. This records the head sha that was granted ONE
+# resubmit exemption, so a second attempt at the SAME head still refuses (see
+# choreographer._unchanged_pr_guard).
+
+RESUBMIT_UNCHANGED_HEAD = "resubmit_unchanged_head"
+
+
+def get_resubmit_unchanged_head(task: HasMarkers) -> str | None:
+    val = get_marker(task, RESUBMIT_UNCHANGED_HEAD)
+    return str(val) if val else None
+
+
+def set_resubmit_unchanged_head(task: HasMarkers, head_sha: str) -> None:
+    set_marker(task, RESUBMIT_UNCHANGED_HEAD, head_sha)
+
+
+# --- block/unblock flip breaker --------------------------------------------
+# Counts successful `unblock` calls on a task so a repeating block/unblock
+# cycle (e.g. escalate_up auto-blocks, a PM unblocks, repeat — live incident:
+# 10 flips, 43 spawns with no forward progress) can alert the CEO instead of
+# burning spawns forever. Payload: {"count": int, "notified": bool} — one key
+# so the counter and the notified-once flag stay atomic with each other.
+
+BLOCK_FLIP_COUNT = "block_flip_count"
+
+
+def get_block_flip_count(task: HasMarkers) -> int:
+    val = get_marker(task, BLOCK_FLIP_COUNT)
+    count = val.get("count") if isinstance(val, dict) else None
+    return int(count) if isinstance(count, int) else 0
+
+
+def is_block_flip_notified(task: HasMarkers) -> bool:
+    val = get_marker(task, BLOCK_FLIP_COUNT)
+    return bool(val.get("notified")) if isinstance(val, dict) else False
+
+
+def bump_block_flip_count(task: HasMarkers) -> int:
+    """Increment the flip counter; returns the new count."""
+    count = get_block_flip_count(task) + 1
+    set_marker(
+        task,
+        BLOCK_FLIP_COUNT,
+        {"count": count, "notified": is_block_flip_notified(task)},
+    )
+    return count
+
+
+def mark_block_flip_notified(task: HasMarkers) -> None:
+    set_marker(
+        task, BLOCK_FLIP_COUNT, {"count": get_block_flip_count(task), "notified": True}
+    )

roboco/mcp/flow_server.py

@@ -702,17 +702,29 @@ def claim_review(task_id: str) -> dict[str, Any]:
 
 
 def pass_review(
-    task_id: str, notes: str, ac_verdicts: StrList | None = None
+    task_id: str,
+    notes: str,
+    ac_verdicts: StrList | None = None,
+    criteria_verified: list[dict[str, Any]] | None = None,
 ) -> dict[str, Any]:
     """QA: accept the work. notes >= 80 chars; journal:learning required.
 
-    ac_verdicts: one entry per acceptance criterion (in criterion order) stating
-    how you verified it. Every criterion must be covered — a pass is rejected
-    until all are. If any criterion does not hold, call fail_review instead.
+    criteria_verified: MANDATORY when the task has acceptance criteria — one
+    {criterion, evidence} entry per criterion. criterion must match an AC by
+    id or exact text; evidence must be concrete (file:line, screenshot ref,
+    rendered-frame path, test name), not a gestalt "looks good". Every
+    criterion must be covered, or the pass is rejected naming exactly which
+    are missing/unmatched. If any criterion does not hold, call fail_review
+    instead of passing a partial.
+
+    ac_verdicts: legacy free-text, one entry per acceptance criterion — still
+    folded into the persisted notes but no longer gates the pass.
     """
     payload: dict[str, Any] = {"task_id": task_id, "notes": notes}
     if ac_verdicts is not None:
         payload["ac_verdicts"] = ac_verdicts
+    if criteria_verified is not None:
+        payload["criteria_verified"] = criteria_verified
     return _post(_role_path("pass"), payload)
 
 
@@ -965,11 +977,15 @@ def delegate(
         nature: One of "technical" | "non_technical".
         acceptance_criteria: Non-empty list of verifiable outcome strings.
         estimated_complexity: One of "low" | "medium" | "high". Default "medium".
-        covers_parent_criteria: The parent task's acceptance-criterion ids this
-            subtask is responsible for. Declare these so the gateway can verify
-            EVERY parent criterion is claimed by a subtask and satisfied before
-            the parent rolls up — split the parent's criteria across subtasks so
-            their union covers all of them.
+        covers_parent_criteria: The parent task's acceptance-criterion ids (or
+            exact text) this subtask is responsible for. REQUIRED — non-empty
+            and resolvable — whenever the parent carries acceptance criteria;
+            delegate rejects a child that maps to none of them or to a ref
+            that matches no real criterion, naming the parent's actual
+            criteria in the rejection. A single delegate need not cover every
+            criterion (a wave may leave some for a later delegate — see the
+            response's evidence.parent_ac_coverage), but every child that IS
+            delegated must map to something real.
         intends_to_touch: Collision surface — file paths/globs this subtask
             will modify. REQUIRED for task_type="code": the sibling collision
             DAG can only sequence what is declared.

roboco/security.py

@@ -231,6 +231,7 @@ _WAF_FREETEXT_BODY_FIELDS: set[str] = {
     "consequences",
     "content",
     "context",
+    "criteria_verified",
     "decision",
     "description",
     "details",

roboco/services/gateway/choreographer/_impl.py

@@ -102,6 +102,11 @@ _PM_SUBTASK_DESC_MAX_LEN = 600
 # tasks, not one giant plan.
 _PM_SUBTASKS_MAX = 7
 
+# unblock's flip breaker: alert the CEO once a task has flip-flopped
+# block/unblock this many times (live incident: 10 flips, 43 spawns with no
+# forward progress before anyone noticed).
+_BLOCK_FLIP_NOTIFY_THRESHOLD = 3
+
 
 def _thin_subtask_hint(sub_tasks: list[Any]) -> str | None:
     """Return a hint if any PM sub_task is title-only / thin / over-long.
@@ -5206,8 +5211,9 @@ class Choreographer:
         )
         # Spec gate passed. Run delegate-specific guards the spec doesn't
         # model: tracing (journal:decision), chain validation, enum
-        # coercion + assignee-vs-task_type, and parent-ownership/subtask-cap.
-        guard = await self._delegate_extra_guards(
+        # coercion + assignee-vs-task_type, parent-ownership/subtask-cap, and
+        # (last) decomposition-coverage.
+        guard = await self._delegate_post_spec_guards(
             pm_agent_id, parent_task_id, parent, role_str, inputs
         )
         if guard is not None:
@@ -5221,6 +5227,75 @@ class Choreographer:
             pm_agent_id, parent, inputs, briefing, role_str
         )
 
+    async def _delegate_post_spec_guards(
+        self,
+        pm_agent_id: UUID,
+        parent_task_id: UUID,
+        parent: Any,
+        role_str: str,
+        inputs: DelegateInputs,
+    ) -> Envelope | None:
+        """``_delegate_extra_guards``, then the decomposition-coverage gate.
+
+        Split out only to keep ``delegate()``'s own return count under the
+        complexity budget. The coverage gate deliberately runs LAST, after
+        every cap/chain/tracing/static rejection, so those guards keep their
+        own specific message (existing tests assert on it) and the coverage
+        gate only ever fires on an otherwise-valid delegate.
+        """
+        guard = await self._delegate_extra_guards(
+            pm_agent_id, parent_task_id, parent, role_str, inputs
+        )
+        if guard is not None:
+            return guard
+        return self._delegate_ac_coverage_guard(parent, inputs)
+
+    def _delegate_ac_coverage_guard(
+        self, parent: Any, inputs: DelegateInputs
+    ) -> Envelope | None:
+        """Reject a child that doesn't map to the parent's own criteria.
+
+        A parent with real acceptance criteria (every task has some) requires
+        every child to declare which ones it advances via
+        ``covers_parent_criteria`` — a child mapped to nothing is either scope
+        creep or a sign the parent's ACs are stale, and either way the PM must
+        decide now, not discover the gap late at submit_up's roll-up gate
+        (``_parent_acs_covered_envelope``) after the whole wave already ran.
+        Inert when the parent has no acceptance criteria at all (nothing to
+        map against) — mirrors the safe-by-construction posture of
+        ``uncovered_parent_acceptance_criteria`` et al. Does NOT require full
+        coverage in one call: a wave may deliberately leave criteria for a
+        later delegate (see the success envelope's ``parent_ac_coverage``
+        evidence for that signal).
+        """
+        ac_texts = parent.acceptance_criteria or []
+        if not ac_texts:
+            return None
+        refs = inputs.covers_parent_criteria or []
+        bad = self.task.unknown_ac_refs(parent, refs) if refs else []
+        if refs and not bad:
+            return None
+        listing = "; ".join(ac_texts)
+        if not refs:
+            message = (
+                f"'{inputs.title}' declares no covers_parent_criteria, but the "
+                "parent has acceptance criteria to decompose."
+            )
+        else:
+            message = (
+                f"'{inputs.title}' covers_parent_criteria has unresolvable "
+                f"ref(s): {'; '.join(bad)}"
+            )
+        return Envelope.invalid_state(
+            message=message,
+            remediate=(
+                "Map this subtask to the parent criteria it advances via "
+                "covers_parent_criteria (by id or exact text), or fix the "
+                f"parent's criteria first. Parent criteria: {listing}"
+            ),
+            context_briefing={},
+        )
+
     # Gate Set B subtask cap (pre-gateway implicit, made explicit here).
     # Soft warn at 8, hard block at 13. Cap enforced by ``_subtask_cap_guard``.
     _SUBTASK_HARD_CAP: int = 12
@@ -5977,9 +6052,29 @@ class Choreographer:
             status="created",
             task_id=str(new_task.id),
             next=spec_module._INTENT_VERBS["delegate"].next_hint(new_task),
+            evidence=await self._delegate_ac_coverage_evidence(parent, parent_task_id),
             context_briefing=briefing,
         ).with_introspection(task=new_task, role=role_str)
 
+    async def _delegate_ac_coverage_evidence(
+        self, parent: Any, parent_task_id: UUID
+    ) -> dict[str, Any]:
+        """``evidence.parent_ac_coverage`` for a successful delegate.
+
+        Reuses ``uncovered_parent_acceptance_criteria`` — the exact primitive
+        submit_up's roll-up gate checks — so the PM sees, in the same turn,
+        which parent criteria the wave-so-far still leaves uncovered instead
+        of discovering it later at assembly. Wave-based planning stays legal:
+        this never blocks, it only reports. Empty (``{}``) when the parent
+        carries no acceptance criteria at all.
+        """
+        ac_texts = parent.acceptance_criteria or []
+        if not ac_texts:
+            return {}
+        uncovered = await self.task.uncovered_parent_acceptance_criteria(parent_task_id)
+        covered = [t for t in ac_texts if t not in uncovered]
+        return {"parent_ac_coverage": {"covered": covered, "uncovered": uncovered}}
+
     @staticmethod
     def _team_value(value: Any) -> str:
         return value.value if hasattr(value, "value") else str(value)
@@ -6792,6 +6887,7 @@ class Choreographer:
             )
 
         t = await self.task.unblock_with_restore(pm_agent_id, task_id, restore=restore)
+        await self._maybe_notify_block_flip(task_id, t)
         next_msg = (
             "task restored to its pre-block state — original assignee will resume"
             if restore
@@ -6804,6 +6900,39 @@ class Choreographer:
             context_briefing=await self._briefing_for(pm_agent_id, task_id),
         ).with_introspection(task=t, role=role)
 
+    async def _maybe_notify_block_flip(self, task_id: UUID, t: Any) -> None:
+        """Bump the flip counter; alert the CEO once past the threshold.
+
+        A resolver that keeps unblocking a task that keeps re-blocking
+        (fe-pm escalate_up auto-blocks / main_pm unblocks, live incident: 10
+        flips, 43 spawns) is a structural wedge, not automatic recovery
+        working as intended. Signal it once at the 3rd flip; the unblock
+        itself always still succeeds regardless.
+        """
+        flip_count = markers.bump_block_flip_count(t)
+        if (
+            flip_count >= _BLOCK_FLIP_NOTIFY_THRESHOLD
+            and not markers.is_block_flip_notified(t)
+        ):
+            markers.mark_block_flip_notified(t)
+            await self._notify_ceo_block_flip(task_id, flip_count)
+
+    async def _notify_ceo_block_flip(self, task_id: UUID, flip_count: int) -> None:
+        """Best-effort CEO alert for a repeating block/unblock cycle; never
+        raises — the breaker signals, it does not wedge unblock itself."""
+        from roboco.services.notification import NotificationService
+
+        try:
+            await NotificationService().send_block_flip_notification(
+                task_id=str(task_id), flip_count=flip_count
+            )
+        except Exception:
+            logger.warning(
+                "failed to send CEO block-flip notification",
+                task_id=str(task_id),
+                flip_count=flip_count,
+            )
+
     async def _own_review_hint(self, pm_agent_id: UUID, exclude_task_id: UUID) -> str:
         """Remediate suffix naming the PM's OWN task ready to complete.
 
@@ -7248,25 +7377,40 @@ class Choreographer:
             )
         return None
 
-    async def _submit_root_unchanged_pr_guard(
-        self, t: Any, briefing: dict[str, Any]
+    async def _unchanged_pr_guard(
+        self,
+        t: Any,
+        briefing: dict[str, Any],
+        *,
+        verb: str,
+        pr_scope: str,
+        delegate_target: str,
+        subtask_noun: str,
+        work_noun: str,
     ) -> Envelope | None:
-        """Refuse to re-submit a root PR whose diff is unchanged since the last fail.
+        """Shared unchanged-PR loop-stopper for submit_root (root) / submit_up (cell).
 
         The hard loop-stopper for the 2026-06-27 ``pr_fail`` re-submit loop. A
         prior ``pr_fail`` stamped the assembled PR's head SHA into
         ``notes_structured.pr_review.head_sha`` (see ``_capture_pr_head_sha`` /
-        ``_record_gate_verdict``). On the next ``submit_root`` this looks up the
-        PR's CURRENT head SHA and compares. Equal ⇒ no new cell work landed on
-        the root branch since the fail ⇒ the diff the reviewer would see is
-        byte-identical to the one just rejected ⇒ refuse, so no model can loop
-        itself back into ``awaiting_pr_review``. Different ⇒ the branch advanced
-        ⇒ allow. Returns ``None`` (proceed) on every ambiguous case so the gate
-        FAILS OPEN: no prior ``pr_fail`` verdict, no recorded ``head_sha`` (e.g.
-        a verdict written before this field existed), no ``pr_number``, no
+        ``_record_gate_verdict``). This looks up the PR's CURRENT head SHA and
+        compares. Equal ⇒ no new work has landed on the branch since the fail ⇒
+        the diff the reviewer would see is byte-identical to the one just
+        rejected. Different ⇒ the branch advanced ⇒ allow.
+
+        A same-head resubmit isn't always a loop: when the rejection round's
+        findings needed no code change (e.g. a transient CI-lookup error
+        ledgered as a finding, then waived/addressed), a flat refuse deadlocks
+        the task forever — the head sha can never move because there is
+        nothing to fix. So once every ledger finding is resolved
+        (``_open_finding_ids`` empty), ONE resubmission is exempted per head
+        sha (``markers.resubmit_unchanged_head``); a second attempt at the SAME
+        head still refuses — one-shot, not a standing bypass.
+
+        Returns ``None`` (proceed) on every ambiguous case so the gate FAILS
+        OPEN: no prior ``pr_fail`` verdict, no recorded ``head_sha`` (e.g. a
+        verdict written before this field existed), no ``pr_number``, no
         resolvable project, or a git/closed-PR lookup that returns ``None``.
-        Only the exact-unchanged case is hard-blocked; the rest fall through to
-        the reviewer, who can still ``pr_fail`` if the diff is bad.
         """
         pr_review = (getattr(t, "notes_structured", None) or {}).get("pr_review") or {}
         if pr_review.get("verdict") != "failed":
@@ -7277,24 +7421,55 @@ class Choreographer:
         current = await self._current_pr_head_sha(t)
         if current is None or current != recorded:
             return None
+        exemption_note = ""
+        if not await self._open_finding_ids(t.id):
+            if markers.get_resubmit_unchanged_head(t) == current:
+                exemption_note = (
+                    " (the one-shot exemption for this head was already used)"
+                )
+            else:
+                markers.set_resubmit_unchanged_head(t, current)
+                logger.info(
+                    "unchanged-PR resubmit exemption granted — all findings"
+                    " resolved without a code change",
+                    task_id=str(t.id),
+                    verb=verb,
+                    head=current[:7],
+                )
+                return None
         return Envelope.invalid_state(
             message=(
-                "the assembled root PR is unchanged since the last pr_fail"
-                f" (head {current[:7]}). No new cell work has landed on the"
-                " root branch, so re-submitting would re-open the exact diff"
-                " the reviewer just rejected and loop straight back to"
-                " awaiting_pr_review."
+                f"the assembled {pr_scope} PR is unchanged since the last"
+                f" pr_fail (head {current[:7]}). No new {work_noun} has"
+                f" landed on the {pr_scope} branch, so re-submitting would"
+                " re-open the exact diff the reviewer just rejected and loop"
+                f" straight back to awaiting_pr_review.{exemption_note}"
             ),
             remediate=(
-                "re-delegate the fixes to the owning cell PM(s) via"
-                " delegate(...) and wait for the cell subtasks to complete"
-                " and the root branch to be re-assembled. Do NOT call"
-                " submit_root again until new cell work has advanced the"
-                " root branch HEAD."
+                f"re-delegate the fixes to the {delegate_target} via"
+                f" delegate(...) and wait for the {subtask_noun} to complete"
+                f" and the {pr_scope} branch to be re-assembled. Do NOT call"
+                f" {verb} again until new {work_noun} has advanced the"
+                f" {pr_scope} branch HEAD."
             ),
             context_briefing=briefing,
         )
 
+    async def _submit_root_unchanged_pr_guard(
+        self, t: Any, briefing: dict[str, Any]
+    ) -> Envelope | None:
+        """Refuse to re-submit a root PR whose diff is unchanged since the last
+        fail. See ``_unchanged_pr_guard`` for the shared logic."""
+        return await self._unchanged_pr_guard(
+            t,
+            briefing,
+            verb="submit_root",
+            pr_scope="root",
+            delegate_target="owning cell PM(s)",
+            subtask_noun="cell subtasks",
+            work_noun="cell work",
+        )
+
     async def _current_pr_head_sha(self, t: Any) -> str | None:
         """Best-effort current head SHA of the task's assembled PR (fail-open).
 
@@ -7338,33 +7513,17 @@ class Choreographer:
 
         Refuses re-submit when the cell PR's current head SHA equals the SHA
         ``pr_fail`` recorded in ``notes_structured.pr_review.head_sha`` (no new
-        dev work landed ⇒ byte-identical diff). Ambiguous cases FAIL OPEN via
-        ``_current_pr_head_sha``; only the exact-unchanged case is hard-blocked.
+        dev work landed ⇒ byte-identical diff). See ``_unchanged_pr_guard`` for
+        the shared logic (including the findings-resolved one-shot exemption).
         """
-        pr_review = (getattr(t, "notes_structured", None) or {}).get("pr_review") or {}
-        if pr_review.get("verdict") != "failed":
-            return None
-        recorded = pr_review.get("head_sha")
-        if not recorded:
-            return None
-        current = await self._current_pr_head_sha(t)
-        if current is None or current != recorded:
-            return None
-        return Envelope.invalid_state(
-            message=(
-                "the assembled cell PR is unchanged since the last pr_fail"
-                f" (head {current[:7]}). No new dev work has landed on the"
-                " cell branch, so re-submitting would re-open the exact diff"
-                " the reviewer just rejected and loop straight back to"
-                " awaiting_pr_review."
-            ),
-            remediate=(
-                "re-delegate the fixes to the owning developer(s) via"
-                " delegate(...) and wait for the dev subtasks to complete and"
-                " the cell branch to be re-assembled. Do NOT call submit_up"
-                " again until new dev work has advanced the cell branch HEAD."
-            ),
-            context_briefing=briefing,
+        return await self._unchanged_pr_guard(
+            t,
+            briefing,
+            verb="submit_up",
+            pr_scope="cell",
+            delegate_target="owning developer(s)",
+            subtask_noun="dev subtasks",
+            work_noun="dev work",
         )
 
     async def submit_root(

roboco/services/gateway/choreographer/findings.py

@@ -124,13 +124,42 @@ def findings_count_hint(findings: Sequence[Any]) -> str | None:
     )
 
 
-def unknown_finding_criteria(task: Any, findings: list[Finding]) -> list[str]:
-    """Findings whose ``criterion`` matches neither an AC id nor AC text.
+def unmatched_criteria(task: Any, criteria: list[str]) -> list[str]:
+    """``criteria`` entries matching neither an AC id nor AC text on ``task``.
 
-    Mirrors ``TaskService.unknown_ac_refs`` — a criterion may be supplied by
+    The shared "by id or exact text" match — a criterion may be supplied by
     its stable id (``acceptance_criteria_ids``, migration 036) or its exact
     text, since both representations already circulate (``declare_coverage``
-    accepts either). Findings with no ``criterion`` are unconstrained.
+    accepts either). Reused by ``unknown_finding_criteria`` (a finding's
+    ``criterion``) and QA's ``criteria_verified`` per-AC gate (qa.py) so the
+    match semantics can't drift between the two call sites.
+    """
+    valid_ids = set(getattr(task, "acceptance_criteria_ids", None) or [])
+    valid_texts = set(getattr(task, "acceptance_criteria", None) or [])
+    return [c for c in criteria if c not in valid_ids and c not in valid_texts]
+
+
+def uncovered_acceptance_criteria(task: Any, criteria: list[str]) -> list[str]:
+    """Task ACs (by exact text) not named by any of ``criteria`` (id or text).
+
+    The reverse of ``unmatched_criteria``: walks the task's own criteria
+    instead of the caller-supplied ones, so a per-AC verification gate can
+    name exactly which acceptance criterion still lacks a matching entry.
+    """
+    ac_ids = list(getattr(task, "acceptance_criteria_ids", None) or [])
+    ac_texts = list(getattr(task, "acceptance_criteria", None) or [])
+    provided = set(criteria)
+    uncovered = []
+    for idx, text in enumerate(ac_texts):
+        ac_id = ac_ids[idx] if idx < len(ac_ids) else None
+        if text in provided or (ac_id is not None and ac_id in provided):
+            continue
+        uncovered.append(text)
+    return uncovered
+
+
+def unknown_finding_criteria(task: Any, findings: list[Finding]) -> list[str]:
+    """Findings whose ``criterion`` matches neither an AC id nor AC text.
 
     Short-circuits to ``[]`` (never touching ``task.acceptance_criteria*``)
     when no finding supplies a criterion — the common case — so a task
@@ -139,9 +168,7 @@ def unknown_finding_criteria(task: Any, findings: list[Finding]) -> list[str]:
     criteria = [f.criterion for f in findings if f.criterion]
     if not criteria:
         return []
-    valid_ids = set(getattr(task, "acceptance_criteria_ids", None) or [])
-    valid_texts = set(getattr(task, "acceptance_criteria", None) or [])
-    return [c for c in criteria if c not in valid_ids and c not in valid_texts]
+    return unmatched_criteria(task, criteria)
 
 
 def criterion_mismatch_rejection(task: Any, unknown: list[str]) -> Envelope:

roboco/services/gateway/choreographer/pr_gate.py

@@ -690,17 +690,18 @@ class PRGateMixin(_Base):
 
         A reviewer must not PASS an assembled PR whose suite can't run in the
         workspace, that carries unresolved architectural-convention
-        violations, or whose CI is red/pending/unscheduled/unresolvable;
-        pr_fail stays available for all three. Returns ``(rejection, None)``
-        to block, or ``(None, ci_note)`` to proceed — ``ci_note`` is a
-        non-None evidence stamp only when the CI guard passed through a
-        project with no CI configured at all. The toolchain/conventions
-        guards are inert when their flag is off; the CI guard fails open on
-        an unresolvable gate-level slug/PR number (``None`` from
-        ``_resolve_ci_status``) and also passes through — with an evidence
-        stamp — when ``get_pr_ci_status`` itself classifies a missing
-        project/git_url/token or an unreachable/nonexistent repo as
-        ``no_ci_configured``.
+        violations, or whose CI is red/pending/unscheduled/unresolvable; only
+        red CI is a code defect pr_fail should act on — an unresolvable CI
+        lookup is a platform blip to retry, never a finding. Returns
+        ``(rejection, None)`` to block, or ``(None, ci_note)`` to proceed —
+        ``ci_note`` is a non-None evidence stamp only when the CI guard
+        passed through a project with no CI configured at all. The
+        toolchain/conventions guards are inert when their flag is off; the
+        CI guard fails open on an unresolvable gate-level slug/PR number
+        (``None`` from ``_resolve_ci_status``) and also passes through —
+        with an evidence stamp — when ``get_pr_ci_status`` itself classifies
+        a missing project/git_url/token or an unreachable/nonexistent repo
+        as ``no_ci_configured``.
         """
         from roboco.config import settings as _settings
 
@@ -806,12 +807,13 @@ class PRGateMixin(_Base):
         """Refuse pr_pass unless CI on the assembled PR's head commit is green.
 
         Failing, pending, unscheduled, or unresolvable-via-API CI states all
-        block with a reviewer-aware remediation pointing at ``pr_fail`` (a
-        reviewer has no ``i_am_blocked``) — pending/unscheduled/error are
-        framed as retryable (wait and call pr_pass again), never as a defect
-        to route back to the dev. A project with no CI configured at all
-        passes through cleanly, returning an evidence note so the caller can
-        stamp the verdict with why the guard did not block.
+        block, but only ``failure`` remediates via ``pr_fail`` (a reviewer
+        has no ``i_am_blocked``); pending/unscheduled/error are framed as
+        retryable (wait and call pr_pass again), never as a defect to route
+        back to the dev — a GitHub API lookup error is a platform blip, not
+        a finding. A project with no CI configured at all passes through
+        cleanly, returning an evidence note so the caller can stamp the
+        verdict with why the guard did not block.
         """
         status = await self._resolve_ci_status(task_id, t)
         if status is None:
@@ -871,11 +873,16 @@ class PRGateMixin(_Base):
                 "green — do NOT pr_pass before any check has run",
             )
         # state == "error" (or an unrecognized value) — a genuine GitHub API
-        # failure resolving the signal; never treat this as green.
+        # lookup failure, never treat this as a property of the PR itself.
+        # Live incident: a reviewer converted this transient blip into an
+        # unwaivable blocker finding whose own fix text said "no code change
+        # required" — a platform hiccup is not a diff defect.
         return (
-            "could not determine CI status for the assembled PR (GitHub API error)",
-            "retry pr_pass shortly once the CI status can be resolved; "
-            "if it persists, pr_fail(issues=[...]) to unwedge the PR",
+            "GitHub API error resolving CI status for the assembled PR — "
+            "this is a transient lookup failure, not a property of the PR",
+            "wait a few minutes and retry pr_pass — do NOT pr_fail over a "
+            "CI-status lookup error; a platform blip is not a code finding, "
+            "findings are for defects in the diff",
         )
 
     def _record_gate_verdict(

roboco/services/gateway/choreographer/qa.py

@@ -58,6 +58,11 @@ from roboco.services.gateway.evidence_builder import build_evidence_for_task
 
 logger = structlog.get_logger()
 
+# Cap on one criteria_verified entry's `evidence` — mirrors Finding.fix's cap
+# (roboco.foundation.policy.content.models._FINDING_FIX_CAP): a pointer
+# (file:line, screenshot ref, test name), not a transcript.
+_CRITERION_EVIDENCE_CAP = 500
+
 if TYPE_CHECKING:
     from uuid import UUID
 
@@ -427,13 +432,13 @@ class QAMixin(_Base):
     def _qa_ac_coverage_check(
         cls, task: Any, ac_verdicts: list[str] | None
     ) -> Envelope | None:
-        """Per-acceptance-criterion verification gate for pass_review.
+        """Legacy count-only per-AC gate — superseded by ``criteria_verified``.
 
-        QA may not pass a task until it has recorded a verification for EVERY
-        acceptance criterion. A single gestalt "looks good" approval is how a
-        silently-unbuilt criterion slips through; requiring one verdict per
-        criterion forces QA to check each individually. If a criterion does not
-        hold, the QA fails the review instead of passing a partial.
+        No longer wired into ``pass_review`` (a count of arbitrary strings
+        never verified they actually named the right criterion — the live gap
+        ``_validate_criteria_verified`` closes). Kept for ``ac_verdicts``'
+        existing callers/tests; ``ac_verdicts`` itself still folds into the
+        persisted notes when supplied.
         """
         criteria = list(getattr(task, "acceptance_criteria", None) or [])
         if not criteria:
@@ -545,36 +550,181 @@ class QAMixin(_Base):
             verb=verb,
         )
 
+    @staticmethod
+    def _parse_criterion_entry(entry: Any, idx: int) -> tuple[str, str] | Envelope:
+        """Validate one ``criteria_verified`` entry into a (criterion, evidence)
+        pair, or an ``Envelope`` rejection on any structural problem."""
+        criterion = entry.get("criterion") if isinstance(entry, dict) else None
+        evidence = entry.get("evidence") if isinstance(entry, dict) else None
+        if not isinstance(criterion, str) or not criterion.strip():
+            return Envelope.invalid_state(
+                message=f"criteria_verified[{idx}] is missing a `criterion` string",
+                remediate=(
+                    "each entry needs {criterion, evidence} naming one "
+                    "acceptance criterion"
+                ),
+            )
+        if not isinstance(evidence, str) or not evidence.strip():
+            return Envelope.invalid_state(
+                message=(
+                    f"criteria_verified[{idx}] ({criterion!r}) is missing `evidence`"
+                ),
+                remediate=(
+                    "state concrete evidence: file:line, screenshot ref, "
+                    "rendered-frame path, test name"
+                ),
+            )
+        if len(evidence) > _CRITERION_EVIDENCE_CAP:
+            return Envelope.invalid_state(
+                message=(
+                    f"criteria_verified[{idx}] evidence exceeds "
+                    f"{_CRITERION_EVIDENCE_CAP} chars"
+                ),
+                remediate="keep evidence concise — a pointer, not a transcript",
+            )
+        return criterion.strip(), evidence.strip()
+
+    @classmethod
+    def _parse_criteria_verified_entries(
+        cls, criteria_verified: list[dict[str, Any]]
+    ) -> tuple[list[tuple[str, str]], Envelope | None]:
+        """Shape + soup validation for every ``criteria_verified`` entry.
+
+        Pure parsing — AC matching/coverage is the caller's job. Split out
+        of ``_validate_criteria_verified`` to keep its return count under
+        the complexity bound.
+        """
+        pairs: list[tuple[str, str]] = []
+        for idx, entry in enumerate(criteria_verified):
+            parsed = cls._parse_criterion_entry(entry, idx)
+            if isinstance(parsed, Envelope):
+                return [], parsed
+            pairs.append(parsed)
+        soup = cls._free_text_soup(
+            checks=(("criteria_verified.evidence", [e for _, e in pairs], 8),)
+        )
+        if soup is not None:
+            return [], soup
+        return pairs, None
+
+    @classmethod
+    def _validate_criteria_verified(
+        cls, t: Any, criteria_verified: list[dict[str, Any]] | None
+    ) -> tuple[list[tuple[str, str]], Envelope | None]:
+        """Mandatory per-AC verification gate for pass_review.
+
+        Returns ``(pairs, rejection)`` — ``pairs`` is ``[]`` and ``rejection``
+        non-None on any failure: none supplied (lists every AC verbatim),
+        a malformed or soupy entry, an entry naming a criterion absent from
+        the task (names the valid criteria), or a task AC left uncovered
+        (names the gap). No task ACs imposes no requirement (mirrors the
+        legacy ``_qa_ac_coverage_check``). A gestalt "looks good" is no
+        longer enough — every criterion needs its own matched, evidenced
+        entry, or QA must call ``fail_review`` instead of passing a partial.
+        """
+        criteria = list(getattr(t, "acceptance_criteria", None) or [])
+        if not criteria:
+            return [], None
+        if not criteria_verified:
+            return [], Envelope.invalid_state(
+                message=(
+                    "pass_review needs criteria_verified naming every "
+                    f"acceptance criterion; none supplied. Unverified: {criteria!r}"
+                ),
+                remediate=(
+                    "re-run the review and call pass_review with "
+                    "criteria_verified=[{criterion, evidence}, ...] — stamp "
+                    "EACH criterion with concrete evidence (file:line, "
+                    "screenshot ref, rendered-frame path, test name)"
+                ),
+            )
+        pairs, bad = cls._parse_criteria_verified_entries(criteria_verified)
+        if bad is not None:
+            return [], bad
+        provided = [c for c, _ in pairs]
+        if unknown := findings_lib.unmatched_criteria(t, provided):
+            return [], Envelope.invalid_state(
+                message=(
+                    f"criteria_verified names criteria not on this task: {unknown!r}"
+                ),
+                remediate=(
+                    "each entry's criterion must match one of the task's "
+                    f"acceptance criteria (by id or exact text): {criteria!r}"
+                ),
+            )
+        if uncovered := findings_lib.uncovered_acceptance_criteria(t, provided):
+            return [], Envelope.invalid_state(
+                message=(
+                    "criteria_verified is missing these acceptance criteria: "
+                    f"{uncovered!r}"
+                ),
+                remediate=(
+                    "stamp every criterion with concrete evidence, or call "
+                    "fail_review with the specific gap if one does not hold"
+                ),
+            )
+        return pairs, None
+
+    @staticmethod
+    def _render_criteria_verified(pairs: list[tuple[str, str]]) -> list[str]:
+        """One '[AC] <criterion> — verified: <evidence>' line per entry.
+
+        Style-matched to the findings ledger's '[F-<id8>] ...' bracket-tag
+        rendering (``findings_lib.render_finding_line``).
+        """
+        return [
+            f"[AC] {criterion} — verified: {evidence}" for criterion, evidence in pairs
+        ]
+
+    @classmethod
+    def _merge_criteria_verified_into_notes(
+        cls, notes: str, pairs: list[tuple[str, str]]
+    ) -> str:
+        """Fold the per-AC verification lines into the persisted QA notes.
+
+        Mirrors ``_merge_ac_verdicts_into_notes`` — keeps the per-criterion
+        verification in the audit trail (qa_notes) so PM/CEO see exactly how
+        QA verified each acceptance criterion.
+        """
+        lines = cls._render_criteria_verified(pairs)
+        if not lines:
+            return notes
+        return f"{notes}\n\n" + "\n".join(lines)
+
     async def _qa_pass_final_gates(
         self,
         qa_agent_id: UUID,
         task_id: UUID,
         t: Any,
         role_str: str,
-        ac_verdicts: list[str] | None,
-    ) -> Envelope | None:
-        """AC-coverage + toolchain-runnability gates for pass_review.
+        criteria_verified: list[dict[str, Any]] | None,
+    ) -> tuple[Envelope | None, list[tuple[str, str]]]:
+        """Per-AC verification + toolchain-runnability gates for pass_review.
 
-        Returns the first rejection (already emitted), else None. QA must not
-        PASS on a workspace that cannot run the suite — that is a source-read
+        Returns ``(rejection, pairs)`` — the first emitted rejection (else
+        None) and the validated ``criteria_verified`` (criterion, evidence)
+        pairs for the caller to render into notes. QA must not PASS on a
+        workspace that cannot run the suite — that is a source-read
         "verification"; fail_review is unaffected.
         """
-        ac_rejection = self._qa_ac_coverage_check(t, ac_verdicts)
-        if ac_rejection is not None:
-            return await self._emit_rejection(
-                ac_rejection.with_introspection(task=t, role=role_str),
+        pairs, bad = self._validate_criteria_verified(t, criteria_verified)
+        if bad is not None:
+            rejection = await self._emit_rejection(
+                bad.with_introspection(task=t, role=role_str),
                 agent_id=qa_agent_id,
                 task_id=task_id,
                 verb="pass_review",
             )
+            return rejection, []
         if toolchain := await self._toolchain_broken_guard(qa_agent_id, t):
-            return await self._emit_rejection(
+            rejection = await self._emit_rejection(
                 toolchain.with_introspection(task=t, role=role_str),
                 agent_id=qa_agent_id,
                 task_id=task_id,
                 verb="pass_review",
             )
-        return None
+            return rejection, []
+        return None, pairs
 
     async def pass_review(
         self,
@@ -582,6 +732,7 @@ class QAMixin(_Base):
         task_id: UUID,
         notes: str,
         ac_verdicts: list[str] | None = None,
+        criteria_verified: list[dict[str, Any]] | None = None,
     ) -> Envelope:
         """QA passes the task; transitions awaiting_qa → awaiting_documentation.
 
@@ -596,6 +747,16 @@ class QAMixin(_Base):
         The composed atomic ``qa_pass`` is then dispatched through
         ``VerbRunner.run_intent``, after which the verb body reassigns
         the documenter for handoff.
+
+        ``criteria_verified`` ({criterion, evidence} entries) is the
+        mandatory per-AC verification gate (``_validate_criteria_verified``):
+        every one of the task's acceptance criteria must be named by exactly
+        one entry — matched by AC id or exact text, the same match
+        ``fail_review``'s findings ledger uses for its own ``criterion``
+        field — carrying substantive evidence, or the pass is refused. A
+        gestalt "looks good" is no longer enough; QA must walk each
+        criterion. ``ac_verdicts`` (legacy, count-only) still folds into the
+        persisted notes when supplied but no longer gates the pass.
         """
         rejection, t = await self._verify_qa_owner(qa_agent_id, task_id, "pass_review")
         if rejection is not None:
@@ -621,18 +782,22 @@ class QAMixin(_Base):
             soup_checks=(("notes", notes, 8),),
         ):
             return gate_rejection
-        if rej := await self._qa_pass_final_gates(
-            qa_agent_id, task_id, t, role_str, ac_verdicts
-        ):
-            return rej
+        final_rejection, criteria_pairs = await self._qa_pass_final_gates(
+            qa_agent_id, task_id, t, role_str, criteria_verified
+        )
+        if final_rejection is not None:
+            return final_rejection
 
         briefing = await self._briefing_for(qa_agent_id, task_id)
+        merged_notes = self._merge_criteria_verified_into_notes(
+            self._merge_ac_verdicts_into_notes(notes, ac_verdicts), criteria_pairs
+        )
         spec_ctx = spec_module.Context(
             actor_id=qa_agent_id,
             actor_slug=getattr(agent, "slug", None) if agent is not None else None,
             agent_team=str(agent.team) if agent is not None and agent.team else None,
             original_developer_slug=_extract_original_developer(t),
-            notes=self._merge_ac_verdicts_into_notes(notes, ac_verdicts),
+            notes=merged_notes,
         )
         self._store_qa_note(t, notes, ac_verdicts, passed=True)
         runner = self._verb_runner()

roboco/services/notification.py

@@ -125,6 +125,40 @@ class NotificationService:
             )
         )
 
+    async def send_block_flip_notification(
+        self,
+        task_id: str,
+        flip_count: int,
+        to_agent: str = "ceo",
+    ) -> None:
+        """Alert an overseer that a task keeps flip-flopping block/unblock.
+
+        Raised from the choreographer's ``unblock`` once the per-task flip
+        counter crosses the threshold — a resolver keeps unblocking a task
+        that keeps re-blocking, which usually means a structural wedge (e.g.
+        an escalate_up/unblock cycle with no forward progress) rather than a
+        one-off block.
+        """
+        logger.info(
+            "Sending block-flip notification", task_id=task_id, flip_count=flip_count
+        )
+        body = (
+            f"Task {task_id} has been blocked and unblocked {flip_count} times. "
+            "This flip-flop usually means a structural wedge rather than a "
+            "one-off block — please investigate."
+        )
+        await self._create_notification(
+            CreateNotificationParams(
+                notification_type=NotificationType.BLOCKER_ESCALATION,
+                priority=NotificationPriority.HIGH,
+                from_agent="system",
+                to_agents=[to_agent],
+                subject=f"Task {task_id} flip-flopping block/unblock ({flip_count}x)",
+                body=body,
+                related_task_id=task_id,
+            )
+        )
+
     async def send_qa_ready_notification(
         self,
         task_id: str,

roboco/services/video_engine.py

@@ -54,6 +54,18 @@ _AUTHORING_ACCEPTANCE_CRITERIA = [
 ]
 _POST_ACCEPTANCE_CRITERIA = ["CEO approves or rejects the draft"]
 
+# Mirrors task_completeness._AC_MAX_ITEMS / _AC_MAX_ITEM_CHARS (mig 068) —
+# duplicated locally per this file's no-cross-service-internals idiom (see
+# vault_intake_engine.py / api/schemas/v1/flow.py for the same pattern).
+_AC_MAX_ITEMS = 7
+_AC_MAX_ITEM_CHARS = 200
+
+# reauthor_from_rejection's fallback AC when the original brief named no
+# enumerable features — the rejection reason is already verbatim in the brief.
+_REAUTHOR_FEEDBACK_CRITERION = (
+    "Every point in the CEO rejection feedback is visibly addressed in the rendered cut"
+)
+
 _CHAT_TIMEOUT_SECONDS = 60.0
 
 # The whole CHANGELOG section for a release, not one bullet — capped so a
@@ -146,6 +158,47 @@ def _release_video_brief(version: str, changelog: str, highlights: list[str]) ->
     return "\n\n".join(parts)
 
 
+def _scene_criterion(features: list[str]) -> str | None:
+    """Turn an enumerable brief feature list into its own gate-checkable AC.
+
+    The live failure this closes: a brief named N features, the task's ACs
+    stayed generic, a dev shipped fewer scenes, and every gate passed because
+    "N features" existed only in prose. None for an empty/absent list. The
+    joined list is bounded to the AC per-item char cap — truncated with an
+    "… (+N more)" tail rather than overrunning it.
+    """
+    trimmed = [f.strip() for f in features if f and f.strip()]
+    if not trimmed:
+        return None
+    prefix = "Every brief-named feature appears as its own fully readable scene: "
+    budget = _AC_MAX_ITEM_CHARS - len(prefix)
+    kept = list(trimmed)
+    while kept:
+        omitted = len(trimmed) - len(kept)
+        tail = f"… (+{omitted} more)" if omitted else ""
+        body = "; ".join(kept) + tail
+        if len(body) <= budget:
+            return prefix + body
+        kept.pop()
+    return prefix + "…"  # pathological: even one (huge) feature name overruns
+
+
+def _authoring_criteria(
+    suggested_input_props: dict[str, Any] | None,
+    fallback_acceptance_criterion: str | None,
+) -> list[str]:
+    """The base authoring ACs plus one derived/fallback criterion: the scene
+    criterion when ``suggested_input_props["highlights"]`` is a real list,
+    else ``fallback_acceptance_criterion`` (when given), else nothing."""
+    criteria = list(_AUTHORING_ACCEPTANCE_CRITERIA)
+    features = (suggested_input_props or {}).get("highlights")
+    criterion = _scene_criterion(features if isinstance(features, list) else [])
+    criterion = criterion or fallback_acceptance_criterion
+    if criterion is not None and len(criteria) < _AC_MAX_ITEMS:
+        criteria.append(criterion)
+    return criteria
+
+
 def _reauthor_brief(reason: str, draft: dict[str, Any]) -> str:
     """The revision brief for a CEO-rejected cut: the verbatim rejection
     feedback, a revise-in-place pointer at the existing composition (when
@@ -270,6 +323,7 @@ class VideoEngine(BaseService):
         brief: str,
         suggested_input_props: dict[str, Any] | None = None,
         project_id: UUID | None = None,
+        fallback_acceptance_criterion: str | None = None,
     ) -> TaskTable | None:
         """Originate ONE UX/UI authoring task for a bespoke video, or None.
 
@@ -292,6 +346,14 @@ class VideoEngine(BaseService):
         "highlights": [...]}`` from the release caller) is seeded onto the
         marker as-is so the dev copies real structured data into
         ``propose_video``'s ``input_props`` instead of hand-typing facts.
+
+        When ``suggested_input_props`` carries a ``highlights`` list, it
+        becomes its OWN acceptance criterion (``_scene_criterion``) — a
+        prose-only feature list used to pass gates even when a dev shipped
+        fewer scenes than named. Absent highlights,
+        ``fallback_acceptance_criterion`` (when supplied) is appended instead
+        — ``reauthor_from_rejection`` uses this for its
+        feedback-must-be-addressed criterion.
         """
         if not settings.video_engine_enabled:
             return None
@@ -316,6 +378,9 @@ class VideoEngine(BaseService):
 
         assignee = self._select_ux_dev(open_tasks)
         enriched_brief = await self._enrich_brief(brief)
+        acceptance_criteria = _authoring_criteria(
+            suggested_input_props, fallback_acceptance_criterion
+        )
         # Savepoint-isolate the insert: a DBAPI error here (FK, deadlock,
         # dropped connection) must roll back ONLY this insert, never poison the
         # shared session — whose next commit is the caller's release-publish
@@ -326,7 +391,7 @@ class VideoEngine(BaseService):
                     TaskCreateRequest(
                         title=f"Video: {occasion}",
                         description=enriched_brief,
-                        acceptance_criteria=list(_AUTHORING_ACCEPTANCE_CRITERIA),
+                        acceptance_criteria=acceptance_criteria,
                         team=Team.UX_UI,
                         assigned_to=assignee,
                         created_by=_foundation.AGENTS["system"].uuid,
@@ -533,6 +598,10 @@ class VideoEngine(BaseService):
                     draft.get("input_props") or draft.get("suggested_input_props")
                 ),
                 project_id=project_id,
+                # Original had highlights -> the scene criterion regenerates
+                # from them; no highlights -> this fallback names the reason
+                # (already verbatim in the brief) as the checkable outcome.
+                fallback_acceptance_criterion=_REAUTHOR_FEEDBACK_CRITERION,
             )
         except Exception as exc:
             self.log.warning(

tests/e2e_smoke/arcs.py

@@ -516,6 +516,13 @@ def qa_arc(stack: E2EStack, company: Company, task_id: Any) -> None:
             ac_verdicts=[
                 f"{c} — verified against the PR diff on the origin." for c in criteria
             ],
+            criteria_verified=[
+                {
+                    "criterion": c,
+                    "evidence": "verified against the PR diff on the origin",
+                }
+                for c in criteria
+            ],
         ),
         "qa pass_review",
     )

tests/integration/test_full_lifecycle_real_db.py

@@ -517,6 +517,10 @@ async def test_full_chain_through_doc_handoff(
         task.id,
         notes=qa_notes,
         ac_verdicts=[f"verified: {crit}" for crit in task.acceptance_criteria],
+        criteria_verified=[
+            {"criterion": crit, "evidence": f"verified against the PR diff: {crit}"}
+            for crit in task.acceptance_criteria
+        ],
     )
     assert env.error is None, f"pass_review failed: {env.message}"
     assert env.status == "awaiting_documentation"

tests/integration/test_lifecycle_real_db.py

@@ -513,6 +513,10 @@ async def test_qa_pass_path(
         task.id,
         notes=_QA_PASS_NOTES,
         ac_verdicts=[f"verified: {crit}" for crit in task.acceptance_criteria],
+        criteria_verified=[
+            {"criterion": crit, "evidence": f"verified against the PR diff: {crit}"}
+            for crit in task.acceptance_criteria
+        ],
     )
     assert env.error is None, f"pass_review failed: {env.message}"
     assert env.status == Status.AWAITING_DOCUMENTATION.value

tests/integration/v1/test_full_pending_to_completed.py

@@ -105,6 +105,7 @@ class _MockChoreographer:
         _task_id: object,
         _notes: object,
         _ac_verdicts: object = None,
+        _criteria_verified: object = None,
     ) -> Envelope:
         self._state["task_status"] = "awaiting_documentation"
         return Envelope.ok(

tests/unit/api/test_schemas_v1_flow.py

@@ -8,6 +8,7 @@ from uuid import uuid4
 import pytest
 from pydantic import ValidationError
 from roboco.api.schemas.v1.flow import (
+    _APPROACH_MAX_CHARS,
     DelegateRequest,
     IWillPlanRequest,
     IWillWorkOnRequest,
@@ -165,13 +166,35 @@ def test_i_will_plan_request_rejects_overlong_plan() -> None:
     assert "plan" in str(exc.value)
 
 
-def test_i_will_plan_request_rejects_overlong_approach() -> None:
-    """approach >800 chars is rejected — no ceiling was the bloat bug."""
+def test_i_will_plan_request_truncates_overlong_approach() -> None:
+    """approach >800 chars is truncated to 800 (797 + "..."), never rejected —
+    a hard 422 here used to throw away an otherwise-good plan and degrade the
+    PM to a bare unblock (live incident)."""
+    req = IWillPlanRequest(
+        task_id=uuid4(),
+        plan="plan",
+        approach="a" * (_APPROACH_MAX_CHARS + 1),
+    )
+    assert len(req.approach) == _APPROACH_MAX_CHARS
+    assert req.approach.endswith("...")
+    assert req.approach == "a" * (_APPROACH_MAX_CHARS - 3) + "..."
+
+
+def test_i_will_plan_request_approach_exactly_800_untruncated() -> None:
+    """Exactly the ceiling is the boundary — passes through byte-for-byte."""
+    approach = "a" * _APPROACH_MAX_CHARS
+    req = IWillPlanRequest(task_id=uuid4(), plan="plan", approach=approach)
+    assert req.approach == approach
+    assert not req.approach.endswith("...")
+
+
+def test_i_will_plan_request_rejects_thin_approach() -> None:
+    """approach <150 chars is still a hard reject — a thin plan IS a defect."""
     with pytest.raises(ValidationError) as exc:
         IWillPlanRequest(
             task_id=uuid4(),
             plan="plan",
-            approach="a" * 801,
+            approach="a" * 149,
         )
     assert "approach" in str(exc.value)
 

tests/unit/foundation/policy/content/test_markers.py

@@ -6,6 +6,9 @@ from types import SimpleNamespace
 
 from roboco.foundation.policy.content import markers as m
 
+# Named constant — ruff PLR2004 forbids magic-value comparisons.
+_TWO = 2
+
 
 def _task(om: dict | None = None) -> SimpleNamespace:
     return SimpleNamespace(orchestration_markers=om)
@@ -130,3 +133,25 @@ def test_docs_sync_release_version_roundtrip() -> None:
     assert m.get_docs_sync_release_version(t) is None
     m.set_docs_sync_release_version(t, "0.23.0")
     assert m.get_docs_sync_release_version(t) == "0.23.0"
+
+
+def test_resubmit_unchanged_head_roundtrip() -> None:
+    t = _task()
+    assert m.get_resubmit_unchanged_head(t) is None
+    m.set_resubmit_unchanged_head(t, "aaaa1111bbbb2222")
+    assert m.get_resubmit_unchanged_head(t) == "aaaa1111bbbb2222"
+
+
+def test_block_flip_count_bump_and_notify() -> None:
+    t = _task()
+    assert m.get_block_flip_count(t) == 0
+    assert m.is_block_flip_notified(t) is False
+    assert m.bump_block_flip_count(t) == 1
+    assert m.bump_block_flip_count(t) == _TWO
+    assert m.get_block_flip_count(t) == _TWO
+    # notified stays False across bumps until explicitly marked.
+    assert m.is_block_flip_notified(t) is False
+    m.mark_block_flip_notified(t)
+    assert m.is_block_flip_notified(t) is True
+    # Marking notified must not reset the counter.
+    assert m.get_block_flip_count(t) == _TWO

tests/unit/gateway/test_choreographer_delegate_guards.py

@@ -137,6 +137,7 @@ async def test_delegate_allows_when_parent_in_progress_and_owned() -> None:
         assigned_to=pm_id,
         team="backend",
         quick_context="Decomposition planned; cells implement their slice next.",
+        acceptance_criteria=[],
     )
     new_task = MagicMock(id=uuid4())
     task_svc = AsyncMock()
@@ -194,6 +195,7 @@ async def test_delegate_allows_when_subtask_cap_within_soft_zone() -> None:
         assigned_to=pm_id,
         team="backend",
         quick_context="Decomposition planned; cells implement their slice next.",
+        acceptance_criteria=[],
     )
     many = [MagicMock(id=uuid4()) for _ in range(10)]
     new_task = MagicMock(id=uuid4())
@@ -223,6 +225,7 @@ async def test_delegate_allows_at_zero_subtasks() -> None:
         assigned_to=pm_id,
         team="backend",
         quick_context="Decomposition planned; cells implement their slice next.",
+        acceptance_criteria=[],
     )
     new_task = MagicMock(id=uuid4())
     task_svc = AsyncMock()
@@ -313,6 +316,7 @@ async def test_delegate_past_max_depth_returns_invalid_state_not_500() -> None:
         assigned_to=pm_id,
         team="backend",
         quick_context="Decomposition planned; cells implement their slice next.",
+        acceptance_criteria=[],
     )
     depth_msg = (
         "Task hierarchy would exceed MAX_TASK_DEPTH=4. Create this work as a "

tests/unit/gateway/test_choreographer_pm_extras.py

@@ -675,6 +675,8 @@ async def test_delegate_main_pm_to_cell_pm_creates_subtask() -> None:
         status="in_progress",
         assigned_to=main_pm_id,
         quick_context="Decomposition planned; cells implement their slice next.",
+        # This test is about role/chain wiring, not AC coverage.
+        acceptance_criteria=[],
     )
     new_task = MagicMock(id=uuid4())
     task_svc = AsyncMock()
@@ -719,6 +721,8 @@ async def test_delegate_cell_pm_to_team_dev_creates_subtask() -> None:
         assigned_to=cell_pm_id,
         team="backend",
         quick_context="Decomposition planned; cells implement their slice next.",
+        # This test is about role/chain wiring, not AC coverage.
+        acceptance_criteria=[],
     )
     new_task = MagicMock(id=uuid4())
     task_svc = AsyncMock()
@@ -1185,6 +1189,9 @@ async def test_delegate_main_pm_to_cell_pm_accepts_planning_subtask() -> None:
         status="in_progress",
         assigned_to=main_pm_id,
         quick_context="Decomposition planned; cells implement their slice next.",
+        # This test is about the code-vs-planning task_type guard, not AC
+        # coverage.
+        acceptance_criteria=[],
     )
     new_task = MagicMock(id=uuid4())
     task_svc = AsyncMock()

tests/unit/gateway/test_choreographer_qa.py

@@ -320,6 +320,94 @@ async def test_pass_review_succeeds_and_transitions() -> None:
     a2a_svc.send.assert_awaited_once()
 
 
+@pytest.mark.asyncio
+async def test_pass_review_rejects_without_criteria_verified_when_acs_present() -> None:
+    """A task with real acceptance criteria demands criteria_verified — a
+    gestalt "looks good" notes string alone is no longer enough."""
+    qa_id = uuid4()
+    task_id = uuid4()
+    t = _qa_owned_task(
+        task_id, qa_id, acceptance_criteria=["returns 200", "includes timestamp"]
+    )
+    task_svc = AsyncMock()
+    task_svc.get.return_value = t
+    task_svc.agent_for.return_value = _qa_agent_mock(qa_id)
+    journal_svc = AsyncMock()
+    journal_svc.has_learning_for_task.return_value = True
+    deps = _make_deps(task=task_svc, journal=journal_svc)
+    c = Choreographer(deps)
+
+    notes = "x" * 100
+    env = await c.pass_review(qa_id, task_id, notes=notes)
+    body = env.as_dict()
+    assert body["error"] == "invalid_state", body
+    assert "returns 200" in body["message"]
+    assert "includes timestamp" in body["message"]
+
+
+@pytest.mark.asyncio
+async def test_pass_review_renders_criteria_verified_into_notes() -> None:
+    """Happy path: every AC matched + evidenced renders '[AC] ...' lines into
+    the persisted qa_notes and the transition still fires."""
+    qa_id = uuid4()
+    task_id = uuid4()
+    t = _qa_owned_task(
+        task_id, qa_id, acceptance_criteria=["returns 200", "includes timestamp"]
+    )
+    after = MagicMock(
+        id=task_id,
+        status="awaiting_documentation",
+        assigned_to=qa_id,
+        team="backend",
+        pr_url="https://x/pr/8",
+        qa_evidence_inspected=True,
+    )
+    task_svc = AsyncMock()
+    task_svc.get.return_value = t
+    task_svc.agent_for.return_value = _qa_agent_mock(qa_id)
+    task_svc.qa_pass.return_value = after
+    task_svc.documenter_for_team.return_value = MagicMock(id=uuid4())
+    task_svc.session = MagicMock()
+    task_svc.session.begin_nested = MagicMock(
+        return_value=MagicMock(
+            __aenter__=AsyncMock(return_value=None),
+            __aexit__=AsyncMock(return_value=False),
+        )
+    )
+    _stub_empty_ledger(task_svc.session)
+    journal_svc = AsyncMock()
+    journal_svc.has_learning_for_task.return_value = True
+    a2a_svc = AsyncMock()
+    deps = _make_deps(task=task_svc, journal=journal_svc, a2a=a2a_svc)
+    c = Choreographer(deps)
+
+    notes = (
+        "Reviewed PR carefully. Rendered every scene and checked each frame "
+        "against the brief before approving."
+    )
+    env = await c.pass_review(
+        qa_id,
+        task_id,
+        notes=notes,
+        criteria_verified=[
+            {"criterion": "returns 200", "evidence": "test_healthz asserts 200"},
+            {
+                "criterion": "includes timestamp",
+                "evidence": "frame diff shows ts field at README.md line 12",
+            },
+        ],
+    )
+    assert env.error is None, env.as_dict()
+    assert env.status == "awaiting_documentation"
+    task_svc.qa_pass.assert_awaited_once()
+    persisted_notes = task_svc.qa_pass.call_args.args[2]
+    assert "[AC] returns 200 — verified: test_healthz asserts 200" in persisted_notes
+    assert (
+        "[AC] includes timestamp — verified: frame diff shows ts field at "
+        "README.md line 12" in persisted_notes
+    )
+
+
 @pytest.mark.asyncio
 async def test_pass_review_not_assigned_returns_not_authorized() -> None:
     qa_id = uuid4()

tests/unit/gateway/test_delegate_ac_coverage_gate.py

@@ -0,0 +1,213 @@
+"""Decomposition-coverage gate: delegate rejects a child that doesn't map to
+the parent's acceptance criteria, and a successful delegate reports the
+parent's remaining coverage gaps in evidence.
+
+Live failure this closes (see CLAUDE.md "delegate" section): unmapped
+children were only ever caught late, at submit_up's roll-up gate — after a
+whole wave of subtasks had already run. Moving the mapping check to delegate
+time surfaces "child covers nothing" / "ref doesn't match a real criterion"
+before any subtask is created.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+from uuid import uuid4
+
+import pytest
+from roboco.services.gateway.choreographer import (
+    Choreographer,
+    ChoreographerDeps,
+    DelegateInputs,
+)
+
+
+def _make_deps(**overrides: Any) -> ChoreographerDeps:
+    base: dict[str, Any] = {
+        "task": AsyncMock(),
+        "work_session": AsyncMock(),
+        "git": AsyncMock(),
+        "a2a": AsyncMock(),
+        "journal": AsyncMock(),
+        "audit": AsyncMock(),
+        "evidence_repo": AsyncMock(),
+    }
+    base.update(overrides)
+    repo = base["evidence_repo"]
+    for method in (
+        "list_unread_a2a",
+        "list_unread_mentions",
+        "list_pending_notifications",
+        "task_metadata_gaps",
+        "recent_team_activity",
+        "blockers_in_lane",
+        "journal_highlights_for_task",
+    ):
+        getattr(repo, method).return_value = []
+    _ldef = base["journal"].latest_decision_at.return_value
+    if type(_ldef).__name__ in ("MagicMock", "AsyncMock"):
+        base["journal"].latest_decision_at.return_value = datetime.now(UTC)
+    return ChoreographerDeps(**base)
+
+
+def _parent_with_criteria(pm_id: Any) -> MagicMock:
+    return MagicMock(
+        id=uuid4(),
+        project_id=uuid4(),
+        status="in_progress",
+        assigned_to=pm_id,
+        team="backend",
+        quick_context="Decomposition planned; cells implement their slice next.",
+        acceptance_criteria=["Criterion A", "Criterion B"],
+        acceptance_criteria_ids=["id-a", "id-b"],
+    )
+
+
+def _inputs(**kw: Any) -> DelegateInputs:
+    base: dict[str, Any] = {
+        "title": "Implement endpoint",
+        "description": "Add /v1/foo endpoint with tests",
+        "assigned_to": "be-dev-1",
+        "team": "backend",
+        "task_type": "code",
+        "nature": "technical",
+        "acceptance_criteria": ["GET /v1/foo returns 200 with body"],
+        "intends_to_touch": ["backend/api/routers/foo.py"],
+    }
+    base.update(kw)
+    return DelegateInputs(**base)
+
+
+@pytest.mark.asyncio
+async def test_delegate_rejects_child_with_no_mapping() -> None:
+    """A parent with real ACs rejects a child that maps to none of them."""
+    pm_id = uuid4()
+    parent = _parent_with_criteria(pm_id)
+    task_svc = AsyncMock()
+    task_svc.get.return_value = parent
+    task_svc.agent_for.return_value = MagicMock(role="cell_pm", team="backend")
+    task_svc.get_subtasks.return_value = []
+    deps = _make_deps(task=task_svc)
+    c = Choreographer(deps)
+
+    env = await c.delegate(pm_id, parent.id, _inputs(title="Orphan slice"))
+    body = env.as_dict()
+    assert body["error"] == "invalid_state", body
+    assert "Orphan slice" in body["message"]
+    assert "no covers_parent_criteria" in body["message"]
+    assert "Criterion A" in body["remediate"] and "Criterion B" in body["remediate"]
+    task_svc.create_subtask.assert_not_awaited()
+    task_svc.unknown_ac_refs.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_delegate_rejects_unresolvable_ref_lists_valid_criteria() -> None:
+    """A ref matching neither a criterion id nor its exact text is rejected,
+    with the parent's real criteria named so the PM can pick a valid one."""
+    pm_id = uuid4()
+    parent = _parent_with_criteria(pm_id)
+    task_svc = AsyncMock()
+    task_svc.get.return_value = parent
+    task_svc.agent_for.return_value = MagicMock(role="cell_pm", team="backend")
+    task_svc.get_subtasks.return_value = []
+    task_svc.unknown_ac_refs = MagicMock(return_value=["bogus-ref"])
+    deps = _make_deps(task=task_svc)
+    c = Choreographer(deps)
+
+    env = await c.delegate(
+        pm_id,
+        parent.id,
+        _inputs(title="Endpoint slice", covers_parent_criteria=["bogus-ref"]),
+    )
+    body = env.as_dict()
+    assert body["error"] == "invalid_state", body
+    assert "Endpoint slice" in body["message"]
+    assert "bogus-ref" in body["message"]
+    assert "Criterion A" in body["remediate"] and "Criterion B" in body["remediate"]
+    task_svc.create_subtask.assert_not_awaited()
+    task_svc.unknown_ac_refs.assert_called_once_with(parent, ["bogus-ref"])
+
+
+@pytest.mark.asyncio
+async def test_delegate_rejects_multiple_unresolvable_refs_in_one_envelope() -> None:
+    """Every unresolvable ref is named in the one rejection, not just the first."""
+    pm_id = uuid4()
+    parent = _parent_with_criteria(pm_id)
+    task_svc = AsyncMock()
+    task_svc.get.return_value = parent
+    task_svc.agent_for.return_value = MagicMock(role="cell_pm", team="backend")
+    task_svc.get_subtasks.return_value = []
+    task_svc.unknown_ac_refs = MagicMock(return_value=["bogus-one", "bogus-two"])
+    deps = _make_deps(task=task_svc)
+    c = Choreographer(deps)
+
+    env = await c.delegate(
+        pm_id,
+        parent.id,
+        _inputs(covers_parent_criteria=["bogus-one", "bogus-two"]),
+    )
+    body = env.as_dict()
+    assert body["error"] == "invalid_state", body
+    assert "bogus-one" in body["message"]
+    assert "bogus-two" in body["message"]
+    task_svc.create_subtask.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_delegate_success_evidence_carries_covered_and_uncovered() -> None:
+    """A resolvable mapping creates the subtask and reports the parent's
+    coverage split in evidence, using the same primitive submit_up checks."""
+    pm_id = uuid4()
+    parent = _parent_with_criteria(pm_id)
+    new_task = MagicMock(id=uuid4())
+    task_svc = AsyncMock()
+    task_svc.get.return_value = parent
+    task_svc.agent_for.return_value = MagicMock(role="cell_pm", team="backend")
+    task_svc.get_subtasks.return_value = []
+    task_svc.unknown_ac_refs = MagicMock(return_value=[])
+    task_svc.create_subtask.return_value = new_task
+    task_svc.uncovered_parent_acceptance_criteria.return_value = ["Criterion B"]
+    deps = _make_deps(task=task_svc)
+    c = Choreographer(deps)
+
+    env = await c.delegate(pm_id, parent.id, _inputs(covers_parent_criteria=["id-a"]))
+    body = env.as_dict()
+    assert body["error"] is None, body
+    assert body["status"] == "created"
+    assert body["evidence"]["parent_ac_coverage"] == {
+        "covered": ["Criterion A"],
+        "uncovered": ["Criterion B"],
+    }
+
+
+@pytest.mark.asyncio
+async def test_delegate_wave_leaving_acs_uncovered_still_succeeds() -> None:
+    """No full-coverage hard gate at delegate: a wave may leave criteria for a
+    later delegate call — the child is still created, gaps just get listed."""
+    pm_id = uuid4()
+    parent = _parent_with_criteria(pm_id)
+    parent.acceptance_criteria = ["Criterion A", "Criterion B", "Criterion C"]
+    parent.acceptance_criteria_ids = ["id-a", "id-b", "id-c"]
+    new_task = MagicMock(id=uuid4())
+    task_svc = AsyncMock()
+    task_svc.get.return_value = parent
+    task_svc.agent_for.return_value = MagicMock(role="cell_pm", team="backend")
+    task_svc.get_subtasks.return_value = []
+    task_svc.unknown_ac_refs = MagicMock(return_value=[])
+    task_svc.create_subtask.return_value = new_task
+    task_svc.uncovered_parent_acceptance_criteria.return_value = [
+        "Criterion B",
+        "Criterion C",
+    ]
+    deps = _make_deps(task=task_svc)
+    c = Choreographer(deps)
+
+    env = await c.delegate(pm_id, parent.id, _inputs(covers_parent_criteria=["id-a"]))
+    body = env.as_dict()
+    assert body["error"] is None, body
+    coverage = body["evidence"]["parent_ac_coverage"]
+    assert coverage["covered"] == ["Criterion A"]
+    assert coverage["uncovered"] == ["Criterion B", "Criterion C"]
+    task_svc.create_subtask.assert_awaited_once()

tests/unit/gateway/test_delegate_incomplete_input.py

@@ -65,6 +65,8 @@ def _parent_in_progress(pm_id: Any) -> MagicMock:
         priority=2,
         # delegate obligates the PM's quick_context resumption section.
         quick_context="Decomposition planned; cells implement their slice next.",
+        # These tests are about task_completeness, not AC coverage.
+        acceptance_criteria=[],
     )
 
 

tests/unit/gateway/test_delegate_parent_lock.py

@@ -63,6 +63,8 @@ def _parent(pm_id: object) -> MagicMock:
         team="backend",
         # delegate obligates the PM's quick_context resumption section.
         quick_context="Decomposition planned; cells implement their slice next.",
+        # These tests are about the parent-lock ordering, not AC coverage.
+        acceptance_criteria=[],
     )
 
 

tests/unit/gateway/test_delegate_project_routing.py

@@ -51,6 +51,8 @@ def _parent(pm_id: Any, product_id: Any = None, project_id: Any = None) -> Magic
         team="backend",
         # delegate obligates the PM's quick_context resumption section.
         quick_context="Decomposition planned; cells implement their slice next.",
+        # These tests are about project routing, not AC coverage.
+        acceptance_criteria=[],
     )
 
 

tests/unit/gateway/test_pr_pass_ci_status_guard.py

@@ -4,12 +4,14 @@ Before this guard, ``pr_pass`` had no CI-status check at all — a reviewer coul
 pass an assembled PR whose CI was red, still running, or not yet scheduled.
 ``_ci_status_guard`` (wired into ``_pr_pass_blocked`` alongside the existing
 toolchain/conventions guards) reads ``GitService.get_pr_ci_status`` and blocks
-on failure/pending/pending_not_scheduled/error with reviewer-aware remediation
-(``pr_fail``, never ``i_am_blocked`` — a reviewer has no such verb). A project
-with no CI configured at all passes through cleanly, stamping the verdict note
-with why the guard did not block. ``pr_fail`` is unaffected by CI state
-entirely, and the separate inbound ``PRReviewerMixin`` surface
-(``claim_pr_review`` / ``post_pr_review``) never consults CI status at all.
+on failure/pending/pending_not_scheduled/error; only ``failure`` remediates
+via ``pr_fail`` (never ``i_am_blocked`` — a reviewer has no such verb) — the
+``error`` state is a transient GitHub API lookup failure and remediates via
+retry only, never ``pr_fail``. A project with no CI configured at all passes
+through cleanly, stamping the verdict note with why the guard did not block.
+``pr_fail`` is unaffected by CI state entirely, and the separate inbound
+``PRReviewerMixin`` surface (``claim_pr_review`` / ``post_pr_review``) never
+consults CI status at all.
 """
 
 from __future__ import annotations
@@ -162,6 +164,7 @@ async def test_pr_pass_blocked_on_github_api_error() -> None:
     assert env.error == "invalid_state"
     assert "GitHub API error" in (env.message or "")
     assert "retry" in (env.remediate or "").lower()
+    assert "do NOT pr_fail" in (env.remediate or "")
 
 
 @pytest.mark.asyncio

tests/unit/gateway/test_qa_criteria_verified.py

@@ -0,0 +1,175 @@
+"""pass_review requires a matched, evidenced verification per acceptance
+criterion — not just a count of arbitrary strings.
+
+Live failure this closes: QA passed a rendered video shipping 3 of the
+brief's 4 named scenes because nothing forced the reviewer to walk each
+acceptance criterion individually. Mirrors the test idiom in
+test_qa_ac_coverage.py, one level stricter: criteria_verified entries must
+each match a real AC (by id or exact text) and carry substantive evidence.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+from roboco.services.gateway.choreographer import Choreographer
+
+_EVIDENCE_CAP = 500
+
+
+def _task(criteria: list[str], ids: list[str] | None = None) -> SimpleNamespace:
+    return SimpleNamespace(
+        acceptance_criteria=criteria,
+        acceptance_criteria_ids=ids or [],
+    )
+
+
+def test_no_criteria_imposes_no_requirement() -> None:
+    pairs, rej = Choreographer._validate_criteria_verified(_task([]), None)
+    assert pairs == []
+    assert rej is None
+
+
+def test_none_supplied_lists_every_criterion_verbatim() -> None:
+    criteria = [
+        "scene 1 renders",
+        "scene 2 renders",
+        "scene 3 renders",
+        "scene 4 renders",
+    ]
+    t = _task(criteria)
+    pairs, rej = Choreographer._validate_criteria_verified(t, None)
+    assert pairs == []
+    assert rej is not None
+    body = rej.as_dict()
+    assert body["error"] == "invalid_state", body
+    for crit in criteria:
+        assert crit in body["message"]
+
+
+def test_empty_list_is_treated_as_none_supplied() -> None:
+    t = _task(["a"])
+    pairs, rej = Choreographer._validate_criteria_verified(t, [])
+    assert pairs == []
+    assert rej is not None
+
+
+def test_partial_coverage_names_the_missing_criterion() -> None:
+    t = _task(["a", "b", "c"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t,
+        [
+            {"criterion": "a", "evidence": "frame 1 shows a rendered"},
+            {"criterion": "b", "evidence": "frame 2 shows b rendered"},
+        ],
+    )
+    assert pairs == []
+    assert rej is not None
+    assert "c" in rej.as_dict()["message"]
+
+
+def test_unmatched_criterion_is_rejected_naming_valid_ones() -> None:
+    t = _task(["a", "b"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t,
+        [
+            {"criterion": "a", "evidence": "frame 1 shows a rendered"},
+            {"criterion": "not-a-real-ac", "evidence": "frame 2 shows something"},
+        ],
+    )
+    assert pairs == []
+    assert rej is not None
+    body = rej.as_dict()
+    assert "not-a-real-ac" in body["message"]
+    assert "a" in body["remediate"] and "b" in body["remediate"]
+
+
+def test_missing_criterion_key_is_rejected() -> None:
+    t = _task(["a"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t, [{"evidence": "frame 1 shows a rendered"}]
+    )
+    assert pairs == []
+    assert rej is not None
+
+
+def test_blank_evidence_is_rejected() -> None:
+    t = _task(["a"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t, [{"criterion": "a", "evidence": "   "}]
+    )
+    assert pairs == []
+    assert rej is not None
+
+
+def test_soup_evidence_is_rejected() -> None:
+    t = _task(["a"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t, [{"criterion": "a", "evidence": "wip"}]
+    )
+    assert pairs == []
+    assert rej is not None
+
+
+def test_overlong_evidence_is_rejected() -> None:
+    t = _task(["a"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t, [{"criterion": "a", "evidence": "x" * (_EVIDENCE_CAP + 100)}]
+    )
+    assert pairs == []
+    assert rej is not None
+    assert str(_EVIDENCE_CAP) in rej.as_dict()["message"]
+
+
+def test_full_coverage_by_exact_text_passes() -> None:
+    t = _task(["a", "b"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t,
+        [
+            {"criterion": "a", "evidence": "frame 1 shows a rendered fully"},
+            {"criterion": "b", "evidence": "frame 2 shows b rendered fully"},
+        ],
+    )
+    assert rej is None
+    assert pairs == [
+        ("a", "frame 1 shows a rendered fully"),
+        ("b", "frame 2 shows b rendered fully"),
+    ]
+
+
+def test_full_coverage_by_ac_id_passes() -> None:
+    t = _task(["scene renders"], ids=["AC-1"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t, [{"criterion": "AC-1", "evidence": "rendered-frame path: out/frame3.png"}]
+    )
+    assert rej is None
+    assert pairs == [("AC-1", "rendered-frame path: out/frame3.png")]
+
+
+def test_extra_entries_beyond_the_ac_set_are_allowed() -> None:
+    t = _task(["a"])
+    pairs, rej = Choreographer._validate_criteria_verified(
+        t,
+        [{"criterion": "a", "evidence": "frame 1 shows a rendered fully"}],
+    )
+    assert rej is None
+    assert len(pairs) == 1
+
+
+def test_render_criteria_verified_matches_style() -> None:
+    lines = Choreographer._render_criteria_verified(
+        [("scene 1 renders", "frame 12 shows scene 1 fully")]
+    )
+    assert lines == ["[AC] scene 1 renders — verified: frame 12 shows scene 1 fully"]
+
+
+def test_merge_criteria_verified_into_notes() -> None:
+    merged = Choreographer._merge_criteria_verified_into_notes(
+        "base review", [("a", "evidence a"), ("b", "evidence b")]
+    )
+    assert "[AC] a — verified: evidence a" in merged
+    assert "[AC] b — verified: evidence b" in merged
+
+
+def test_merge_with_no_pairs_returns_notes_unchanged() -> None:
+    assert Choreographer._merge_criteria_verified_into_notes("base", []) == "base"

tests/unit/gateway/test_resubmit_unchanged_head_exemption.py

@@ -0,0 +1,233 @@
+"""The one-shot resubmit exemption on the unchanged-PR loop-stopper.
+
+Live wedge: a cell task looped block/unblock for hours because the freshness
+guard (``_unchanged_pr_guard``, shared by ``submit_root``/``submit_up``)
+demands a new commit after ANY ``pr_fail`` — a structural deadlock when the
+rejection round's findings require no code change (e.g. a transient
+CI-lookup error ledgered as a finding, then waived/addressed). Once every
+ledger finding is resolved (``_open_finding_ids`` empty), ONE resubmission is
+exempted per head sha (``markers.resubmit_unchanged_head``); a second attempt
+at the SAME head still refuses. Covers both call sites: ``submit_root``
+(root) and ``submit_up`` (cell).
+
+Note on test style: the upstream ``FINDINGS_ADDRESSED`` tracing gate
+(``_check_submit_up_gates``, shared by both verbs) already refuses the whole
+verb with ``tracing_gap`` whenever findings are open — by the time
+``_unchanged_pr_guard`` runs, open findings are already impossible through the
+public ``submit_root``/``submit_up`` entrypoints. The "findings still open"
+and "ambiguous case short-circuits before the findings check" scenarios are
+therefore exercised by calling the guard method directly (defense-in-depth on
+the guard's own logic); the exemption-grant and one-shot-per-head scenarios
+drive the full verb end-to-end to prove the real wiring.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+from uuid import uuid4
+
+import pytest
+from roboco.foundation.policy.content import markers
+from roboco.services.gateway.choreographer import Choreographer, ChoreographerDeps
+
+SHA_OLD = "aaaa1111bbbb2222cccc3333dddd4444eeee5555"
+SHA_NEW = "9999888877776666555544443333222211110000"
+
+
+def _make_deps(**overrides: Any) -> ChoreographerDeps:
+    base: dict[str, Any] = {
+        "task": AsyncMock(),
+        "work_session": AsyncMock(),
+        "git": AsyncMock(),
+        "a2a": AsyncMock(),
+        "journal": AsyncMock(),
+        "audit": AsyncMock(),
+        "evidence_repo": AsyncMock(),
+    }
+    base.update(overrides)
+    base["journal"].has_decision_for_task.return_value = True
+    base["journal"].latest_decision_at.return_value = datetime.now(UTC)
+    base["journal"].has_reflect_for_task.return_value = True
+    return ChoreographerDeps(**base)
+
+
+def _unchanged_notes(head_sha: str = SHA_OLD) -> dict[str, Any]:
+    return {"pr_review": {"verdict": "failed", "head_sha": head_sha, "summary": "..."}}
+
+
+async def _call_guard(
+    c: Choreographer, kind: str, t: Any, briefing: dict[str, Any]
+) -> Any:
+    if kind == "root":
+        return await c._submit_root_unchanged_pr_guard(t, briefing)
+    return await c._submit_up_unchanged_pr_guard(t, briefing)
+
+
+# ---------------------------------------------------------------------------
+# Direct guard tests — isolate ``_unchanged_pr_guard`` from the upstream
+# FINDINGS_ADDRESSED gate, which already forbids open findings from ever
+# reaching here through the real submit_root/submit_up flow.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["root", "cell"])
+async def test_refused_when_findings_still_open(kind: str) -> None:
+    """Unchanged head + open findings -> refused (behavior unchanged)."""
+    c = Choreographer(_make_deps())
+    cc: Any = c
+    cc._current_pr_head_sha = AsyncMock(return_value=SHA_OLD)
+    cc._open_finding_ids = AsyncMock(return_value=("abcd1234",))
+    t = MagicMock(
+        id=uuid4(), notes_structured=_unchanged_notes(), orchestration_markers=None
+    )
+
+    env = await _call_guard(c, kind, t, {})
+
+    assert env is not None
+    assert env.error == "invalid_state", env.as_dict()
+    assert "unchanged" in (env.message or "").lower()
+    assert markers.get_resubmit_unchanged_head(t) is None
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["root", "cell"])
+async def test_allowed_when_head_advanced_before_findings_are_even_checked(
+    kind: str,
+) -> None:
+    """A different current head SHA -> allowed; the ambiguity check short-
+    circuits before the findings/exemption logic ever runs."""
+    c = Choreographer(_make_deps())
+    cc: Any = c
+    cc._current_pr_head_sha = AsyncMock(return_value=SHA_NEW)
+    open_findings_spy = AsyncMock(return_value=("abcd1234",))
+    cc._open_finding_ids = open_findings_spy
+    t = MagicMock(
+        id=uuid4(), notes_structured=_unchanged_notes(), orchestration_markers=None
+    )
+
+    env = await _call_guard(c, kind, t, {})
+
+    assert env is None
+    open_findings_spy.assert_not_awaited()
+
+
+# ---------------------------------------------------------------------------
+# End-to-end tests — drive the real submit_root/submit_up verb to prove the
+# exemption's marker stamp actually lets the assembled PR proceed into the
+# gate. Zero open findings satisfies both the upstream FINDINGS_ADDRESSED gate
+# and the new exemption check, mirroring test_submit_root_unchanged_pr_guard.py
+# / test_submit_up_unchanged_pr_guard.py's fixture shape.
+# ---------------------------------------------------------------------------
+
+
+def _resubmit(
+    kind: str,
+    *,
+    notes_structured: dict[str, Any] | None,
+    pr_number: int | None = 139,
+) -> tuple[Choreographer, Any, Any]:
+    pm_id = uuid4()
+    task_id = uuid4()
+    if kind == "root":
+        role, team, parent = "main_pm", "main_pm", None
+        branch = "feature/main_pm/c80e19ff"
+    else:
+        role, team, parent = "cell_pm", "backend", uuid4()
+        branch = "feature/backend/cell-task"
+    in_prog = MagicMock(
+        id=task_id,
+        status="in_progress",
+        assigned_to=pm_id,
+        pr_number=pr_number,
+        branch_name=branch,
+        parent_task_id=parent,
+        batch_id=None,
+        team=team,
+        notes_structured=notes_structured,
+        orchestration_markers=None,
+    )
+    gated = MagicMock(**{**in_prog.__dict__, "status": "awaiting_pr_review"})
+    task_svc = AsyncMock()
+    task_svc.get.return_value = in_prog
+    task_svc.submit_for_review.return_value = gated
+    task_svc.all_subtasks_terminal.return_value = True
+    task_svc.uncovered_parent_acceptance_criteria.return_value = []
+    task_svc.agent_for.return_value = MagicMock(role=role, team=team)
+    task_svc.session.begin_nested = MagicMock(
+        return_value=MagicMock(__aenter__=AsyncMock(), __aexit__=AsyncMock())
+    )
+    c = Choreographer(_make_deps(task=task_svc, git=AsyncMock()))
+    cc: Any = c
+    cc._project_slug_for = AsyncMock(return_value="proj-slug")
+    # Zero open findings throughout — satisfies the upstream FINDINGS_ADDRESSED
+    # gate so the flow reaches the unchanged-PR guard under test.
+    cc._open_finding_ids = AsyncMock(return_value=())
+    return c, pm_id, task_id
+
+
+async def _call_submit(
+    c: Choreographer, kind: str, pm_id: Any, task_id: Any, notes: str
+) -> Any:
+    if kind == "root":
+        return await c.submit_root(pm_id, task_id, notes=notes)
+    return await c.submit_up(pm_id, task_id, notes=notes)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["root", "cell"])
+async def test_exemption_granted_when_no_open_findings_and_no_prior_marker(
+    kind: str,
+) -> None:
+    """Unchanged head + zero open findings + no marker -> allowed once, marker
+    stamped with the current head."""
+    c, pm_id, task_id = _resubmit(kind, notes_structured=_unchanged_notes())
+    c.git.get_pr_head_sha = AsyncMock(return_value=SHA_OLD)
+
+    env = await _call_submit(
+        c, kind, pm_id, task_id, "re-submitting; CI blip only, no code change needed"
+    )
+
+    assert env.error is None, env.as_dict()
+    assert env.status == "awaiting_pr_review"
+    c.task.submit_for_review.assert_awaited_once()
+    t = await c.task.get(task_id)
+    assert markers.get_resubmit_unchanged_head(t) == SHA_OLD
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["root", "cell"])
+async def test_second_attempt_at_same_head_is_refused(kind: str) -> None:
+    """The exemption is one-shot: a second resubmit at the SAME head, still
+    with no open findings, refuses — the marker already recorded this head."""
+    c, pm_id, task_id = _resubmit(kind, notes_structured=_unchanged_notes())
+    c.git.get_pr_head_sha = AsyncMock(return_value=SHA_OLD)
+
+    first = await _call_submit(c, kind, pm_id, task_id, "first resubmit; CI blip only")
+    assert first.error is None, first.as_dict()
+
+    second = await _call_submit(
+        c, kind, pm_id, task_id, "second resubmit; still the same unchanged head"
+    )
+    assert second.error == "invalid_state", second.as_dict()
+    assert "unchanged" in (second.message or "").lower()
+    assert "already used" in (second.message or "").lower()
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["root", "cell"])
+async def test_allowed_when_head_advanced(kind: str) -> None:
+    """A different current head SHA -> allowed (existing fail-open behavior),
+    end to end through the real verb."""
+    c, pm_id, task_id = _resubmit(kind, notes_structured=_unchanged_notes())
+    c.git.get_pr_head_sha = AsyncMock(return_value=SHA_NEW)
+
+    env = await _call_submit(
+        c, kind, pm_id, task_id, "resubmitting after the real fix landed"
+    )
+
+    assert env.error is None, env.as_dict()
+    assert env.status == "awaiting_pr_review"
+    c.task.submit_for_review.assert_awaited_once()

tests/unit/gateway/test_submit_root_unchanged_pr_guard.py

@@ -105,8 +105,15 @@ def _resubmit_root(
 
 @pytest.mark.asyncio
 async def test_submit_root_refuses_unchanged_pr_after_pr_fail() -> None:
-    """The loop-stopper: prior pr_fail stamped head SHA X, the PR head is still
-    X (no new cell work on the root branch) → refuse, do not open the gate."""
+    """The loop-stopper still holds past the one-shot exemption: prior
+    pr_fail stamped head SHA X, the PR head is still X (no new cell work on
+    the root branch). The findings ledger here fail-opens to "nothing open"
+    (mock session, no real query) — the same signal ``_check_submit_up_gates``
+    upstream already reads for FINDINGS_ADDRESSED — so the first resubmit at
+    this head is the one-shot exemption (see
+    test_resubmit_unchanged_head_exemption.py for full exemption coverage);
+    a second resubmit at the SAME head refuses, so the loop still can't run
+    forever."""
     c, main_pm_id, root_task_id = _resubmit_root(
         notes_structured={
             "pr_review": {"verdict": "failed", "head_sha": SHA_OLD, "summary": "..."}
@@ -114,9 +121,14 @@ async def test_submit_root_refuses_unchanged_pr_after_pr_fail() -> None:
     )
     c.git.get_pr_head_sha = AsyncMock(return_value=SHA_OLD)
 
-    env = await c.submit_root(
+    first = await c.submit_root(
         main_pm_id, root_task_id, notes="re-submitting the root after the fix"
     )
+    assert first.error is None, first.as_dict()
+
+    env = await c.submit_root(
+        main_pm_id, root_task_id, notes="re-submitting again; still unchanged"
+    )
 
     assert env.error is not None, env.as_dict()
     assert env.error == "invalid_state"
@@ -124,8 +136,9 @@ async def test_submit_root_refuses_unchanged_pr_after_pr_fail() -> None:
     remediate = env.remediate or ""
     assert "re-delegate" in remediate
     assert "submit_root" in remediate
-    # The PR was NOT re-opened / re-pushed — the runner never ran.
-    c.task.submit_for_review.assert_not_awaited()
+    # The second attempt's PR was NOT re-opened / re-pushed — the runner ran
+    # only for the first (exempted) call.
+    c.task.submit_for_review.assert_awaited_once()
 
 
 @pytest.mark.asyncio

tests/unit/gateway/test_submit_up_unchanged_pr_guard.py

@@ -80,9 +80,15 @@ def _resubmit_cell(
 
 @pytest.mark.asyncio
 async def test_submit_up_refuses_unchanged_pr_after_pr_fail() -> None:
-    """The loop-stopper: prior pr_fail stamped head SHA X, the cell PR head is
-    still X (no new dev work on the cell branch) → refuse, do not re-open the
-    gate."""
+    """The loop-stopper still holds past the one-shot exemption: prior
+    pr_fail stamped head SHA X, the cell PR head is still X (no new dev work
+    on the cell branch). The findings ledger here fail-opens to "nothing
+    open" (mock session, no real query) — the same signal
+    ``_check_submit_up_gates`` upstream already reads for FINDINGS_ADDRESSED
+    — so the first resubmit at this head is the one-shot exemption (see
+    test_resubmit_unchanged_head_exemption.py for full exemption coverage);
+    a second resubmit at the SAME head refuses, so the loop still can't run
+    forever."""
     c, cell_pm_id, cell_task_id = _resubmit_cell(
         notes_structured={
             "pr_review": {"verdict": "failed", "head_sha": SHA_OLD, "summary": "..."}
@@ -90,17 +96,23 @@ async def test_submit_up_refuses_unchanged_pr_after_pr_fail() -> None:
     )
     c.git.get_pr_head_sha = AsyncMock(return_value=SHA_OLD)
 
-    env = await c.submit_up(
+    first = await c.submit_up(
         cell_pm_id, cell_task_id, notes="re-submitting the cell after the fix"
     )
+    assert first.error is None, first.as_dict()
+
+    env = await c.submit_up(
+        cell_pm_id, cell_task_id, notes="re-submitting again; still unchanged"
+    )
 
     assert env.error is not None, env.as_dict()
     assert env.error == "invalid_state", env.as_dict()
     assert "unchanged" in (env.message or "").lower()
     remediate = env.remediate or ""
     assert "submit_up" in remediate
-    # The cell PR was NOT re-opened / re-pushed — the runner never ran.
-    c.task.submit_for_review.assert_not_awaited()
+    # The second attempt's PR was NOT re-opened / re-pushed — the runner ran
+    # only for the first (exempted) call.
+    c.task.submit_for_review.assert_awaited_once()
 
 
 @pytest.mark.asyncio

tests/unit/gateway/test_unblock_flip_breaker.py

@@ -0,0 +1,151 @@
+"""The block/unblock flip-flop breaker.
+
+Live wedge: fe-pm's escalate_up auto-blocks a task and main_pm's unblock
+resolves it, repeat — 10 flips, 43 spawns, no forward progress, no cycle
+breaker. ``unblock`` now stamps a per-task flip counter
+(``markers.block_flip_count``) and, at exactly the 3rd flip, best-effort
+alerts the CEO once — the unblock itself always still succeeds.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+from uuid import uuid4
+
+import pytest
+from roboco.foundation.policy.content import markers
+from roboco.services.gateway.choreographer import Choreographer, ChoreographerDeps
+from roboco.services.notification import NotificationService
+
+# Named constants — ruff PLR2004 forbids magic-value comparisons.
+_TWO_FLIPS = 2
+_THREE_FLIPS = 3
+_FOUR_FLIPS = 4
+
+
+def _make_deps(**overrides: Any) -> ChoreographerDeps:
+    base: dict[str, Any] = {
+        "task": AsyncMock(),
+        "work_session": AsyncMock(),
+        "git": AsyncMock(),
+        "a2a": AsyncMock(),
+        "journal": AsyncMock(),
+        "audit": AsyncMock(),
+        "evidence_repo": AsyncMock(),
+    }
+    base.update(overrides)
+    base["journal"].has_decision_for_task.return_value = True
+    base["journal"].latest_decision_at.return_value = datetime.now(UTC)
+    return ChoreographerDeps(**base)
+
+
+def _flip_setup() -> tuple[Choreographer, Any, Any, Any]:
+    """A blocked task whose ``unblock_with_restore`` returns the SAME mock
+    object each call, so the flip-counter marker persists across repeated
+    unblock() calls the way it would on one real ORM row across requests.
+    """
+    pm_id = uuid4()
+    task_id = uuid4()
+    t = MagicMock(
+        id=task_id,
+        status="blocked",
+        pre_block_state="in_progress",
+        pre_block_assignee=uuid4(),
+        pre_block_metadata={},
+        dependency_ids=[],
+        orchestration_markers=None,
+    )
+    task_svc = AsyncMock()
+    task_svc.get.return_value = t
+    task_svc.unblock_with_restore.return_value = t
+    task_svc.unmet_dependency_ids.return_value = []
+    c = Choreographer(_make_deps(task=task_svc))
+    return c, pm_id, task_id, t
+
+
+async def _unblock_once(c: Choreographer, pm_id: Any, task_id: Any, t: Any) -> Any:
+    """Re-block before each call — a fresh flip in the cycle."""
+    t.status = "blocked"
+    return await c.unblock(pm_id, task_id, "resolved upstream; restoring")
+
+
+@pytest.mark.asyncio
+async def test_first_and_second_unblock_do_not_notify() -> None:
+    c, pm_id, task_id, t = _flip_setup()
+    cc: Any = c
+    notify = AsyncMock()
+    cc._notify_ceo_block_flip = notify
+
+    for _ in range(2):
+        env = await _unblock_once(c, pm_id, task_id, t)
+        assert env.error is None, env.as_dict()
+
+    notify.assert_not_awaited()
+    assert markers.get_block_flip_count(t) == _TWO_FLIPS
+
+
+@pytest.mark.asyncio
+async def test_third_unblock_notifies_ceo_once() -> None:
+    c, pm_id, task_id, t = _flip_setup()
+    cc: Any = c
+    notify = AsyncMock()
+    cc._notify_ceo_block_flip = notify
+
+    for _ in range(3):
+        env = await _unblock_once(c, pm_id, task_id, t)
+        assert env.error is None, env.as_dict()
+
+    notify.assert_awaited_once_with(task_id, _THREE_FLIPS)
+    assert markers.is_block_flip_notified(t) is True
+
+
+@pytest.mark.asyncio
+async def test_fourth_unblock_does_not_renotify() -> None:
+    c, pm_id, task_id, t = _flip_setup()
+    cc: Any = c
+    notify = AsyncMock()
+    cc._notify_ceo_block_flip = notify
+
+    for _ in range(4):
+        env = await _unblock_once(c, pm_id, task_id, t)
+        assert env.error is None, env.as_dict()
+
+    notify.assert_awaited_once()
+    assert markers.get_block_flip_count(t) == _FOUR_FLIPS
+
+
+@pytest.mark.asyncio
+async def test_notification_failure_does_not_fail_unblock(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """The real ``_notify_ceo_block_flip`` swallows a notify-service failure —
+    unblock still succeeds on the 3rd flip."""
+    c, pm_id, task_id, t = _flip_setup()
+    monkeypatch.setattr(
+        NotificationService,
+        "send_block_flip_notification",
+        AsyncMock(side_effect=RuntimeError("notification service down")),
+    )
+
+    env = None
+    for _ in range(3):
+        env = await _unblock_once(c, pm_id, task_id, t)
+        assert env.error is None, env.as_dict()
+
+    assert env is not None
+    assert env.error is None
+    assert markers.is_block_flip_notified(t) is True
+
+
+@pytest.mark.asyncio
+async def test_counter_persists_via_marker_across_calls() -> None:
+    c, pm_id, task_id, t = _flip_setup()
+    cc: Any = c
+    cc._notify_ceo_block_flip = AsyncMock()
+
+    await _unblock_once(c, pm_id, task_id, t)
+    assert markers.get_block_flip_count(t) == 1
+    await _unblock_once(c, pm_id, task_id, t)
+    assert markers.get_block_flip_count(t) == _TWO_FLIPS

tests/unit/services/test_video_engine.py

@@ -36,6 +36,9 @@ ONE = 1
 TWO = 2
 THREE = 3
 FOUR = 4
+FIVE = 5
+SEVEN = 7
+_AC_ITEM_CHAR_CAP = 200  # mirrors task_completeness._AC_MAX_ITEM_CHARS
 
 
 async def _seed(session: AsyncSession) -> None:
@@ -896,3 +899,152 @@ async def test_reauthor_from_rejection_missing_draft_returns_none(
 
     result = await engine.reauthor_from_rejection(task, "some reason")
     assert result is None
+
+
+# --------------------------------------------------------------------------- #
+# scene criterion — a brief-named feature list becomes its own gate-checkable
+# AC (delegation detail-fidelity: prose feature counts used to pass gates
+# even when a dev shipped fewer scenes than the brief named)
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_draft_release_video_carries_scene_criterion_for_highlights(
+    db_session: AsyncSession, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    await _seed(db_session)
+    _enable(monkeypatch, video_on_release=True)
+    _mock_local_model(monkeypatch, "shipped!")
+    engine = video_engine_module.VideoEngine(db_session)
+    task = await engine.draft_release_video(version="1.0.0", changelog=_CHANGELOG)
+    assert task is not None
+    assert len(task.acceptance_criteria) == FIVE
+    scene_ac = task.acceptance_criteria[FOUR]
+    assert scene_ac.startswith("Every brief-named feature appears as its own")
+    assert "a huge new release" in scene_ac
+
+
+@pytest.mark.asyncio
+async def test_open_video_task_no_highlights_keeps_ac_count_unchanged(
+    db_session: AsyncSession, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    await _seed(db_session)
+    _enable(monkeypatch)
+    engine = video_engine_module.VideoEngine(db_session)
+    task = await engine.open_video_task(
+        occasion="spotlight no-highlights",
+        script="s",
+        platforms=["x"],
+        brief="b",
+        suggested_input_props={"version": "1.0.0"},  # no "highlights" key
+    )
+    assert task is not None
+    assert len(task.acceptance_criteria) == FOUR  # no scene criterion appended
+
+
+@pytest.mark.asyncio
+async def test_reauthor_from_rejection_with_highlights_regenerates_scene_criterion(
+    db_session: AsyncSession, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    await _seed(db_session)
+    _enable(monkeypatch)
+    engine = video_engine_module.VideoEngine(db_session)
+    source_task = await engine.open_video_task(
+        occasion="release v3.0.0",
+        script="s",
+        platforms=["x"],
+        brief="b",
+        suggested_input_props={
+            "version": "3.0.0",
+            "highlights": ["Feature A", "Feature B"],
+        },
+    )
+    assert source_task is not None
+    draft = markers.get_video_draft(source_task) or {}
+    # Simulate propose_video: the dev's real input_props carries the same
+    # highlights forward onto the marker.
+    markers.set_video_draft(
+        source_task,
+        {
+            **draft,
+            "composition_id": "ReleaseIntro",
+            "input_props": {
+                "version": "3.0.0",
+                "highlights": ["Feature A", "Feature B"],
+            },
+        },
+    )
+    source_task.status = TS.COMPLETED
+    await db_session.flush()
+
+    post_task = await engine._originate_video_post(
+        source_task=source_task,
+        mp4_paths={"vertical": "a.mp4", "square": "b.mp4"},
+        captions={"x": "cap"},
+        platforms=["x"],
+    )
+    post_task.status = TS.CANCELLED
+    await db_session.flush()
+
+    revision = await engine.reauthor_from_rejection(post_task, "Logo too small")
+    assert revision is not None
+    assert len(revision.acceptance_criteria) == FIVE
+    scene_ac = revision.acceptance_criteria[FOUR]
+    assert "Feature A" in scene_ac
+    assert "Feature B" in scene_ac
+
+
+@pytest.mark.asyncio
+async def test_reauthor_from_rejection_without_highlights_adds_feedback_criterion(
+    db_session: AsyncSession, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    await _seed(db_session)
+    _enable(monkeypatch)
+    engine = video_engine_module.VideoEngine(db_session)
+    source_task = await engine.open_video_task(
+        occasion="release v4.0.0", script="s", platforms=["x"], brief="b"
+    )
+    assert source_task is not None
+    draft = markers.get_video_draft(source_task) or {}
+    markers.set_video_draft(source_task, {**draft, "composition_id": "Intro"})
+    source_task.status = TS.COMPLETED
+    await db_session.flush()
+
+    post_task = await engine._originate_video_post(
+        source_task=source_task,
+        mp4_paths={"vertical": "a.mp4", "square": "b.mp4"},
+        captions={"x": "cap"},
+        platforms=["x"],
+    )
+    post_task.status = TS.CANCELLED
+    await db_session.flush()
+
+    revision = await engine.reauthor_from_rejection(post_task, "Logo is cut off")
+    assert revision is not None
+    assert len(revision.acceptance_criteria) == FIVE
+    assert revision.acceptance_criteria[FOUR] == (
+        "Every point in the CEO rejection feedback is visibly addressed in "
+        "the rendered cut"
+    )
+
+
+@pytest.mark.asyncio
+async def test_open_video_task_scene_criterion_truncates_pathological_feature_list(
+    db_session: AsyncSession, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    await _seed(db_session)
+    _enable(monkeypatch)
+    engine = video_engine_module.VideoEngine(db_session)
+    features = [f"Feature number {i}" for i in range(50)]
+    task = await engine.open_video_task(
+        occasion="release pathological",
+        script="s",
+        platforms=["x"],
+        brief="b",
+        suggested_input_props={"highlights": features},
+    )
+    assert task is not None
+    assert len(task.acceptance_criteria) <= SEVEN
+    scene_ac = task.acceptance_criteria[-1]
+    assert len(scene_ac) <= _AC_ITEM_CHAR_CAP
+    assert "more)" in scene_ac  # truncated, not silently dropped