mandrel 2.59.0 → 2.61.0

package/.agents/workflows/helpers/acceptance-self-eval.md

@@ -1,7 +1,7 @@
 ---
 description: >-
   Shared include for the bounded acceptance self-eval loop run during Story
-  delivery (`helpers/deliver-story`). Defines the per-round critic mechanic;
+  delivery (`helpers/deliver-story`). Defines the per-round verdict mechanic;
   the caller supplies its gate-decision wrapper (label transitions).
 ---
 
@@ -9,181 +9,103 @@ description: >-
 
 > **Include module.** Not a slash command. Referenced from
 > [`deliver-story.md`](deliver-story.md) at Step 1a. This file is the
-> **single prose home** for the per-round critic mechanic; the caller layers
+> **single prose home** for the per-round verdict mechanic; the caller layers
 > only its wrapper (Story label transitions).
 
 After the implementation commits land and **before** the Story proceeds to
-close, run an explicit, **independent** eval pass that scores the change set
-computed once for this Story and injected into the critic — never one the
-critic re-derives — against **each** `acceptance[]` item
-individually. This is the acceptance gate
-the close-validation chain does not provide: that chain (lint / test / format /
+close, run an explicit eval pass that scores the change set computed once for
+this Story — never one the evaluator re-derives — against **each**
+`acceptance[]` item individually. This is the acceptance gate the
+close-validation chain does not provide: that chain (lint / test / format /
 maintainability / coverage / crap) proves the code is *healthy*, not that it
 satisfies *this Story's* acceptance criteria.
 
 The loop is **always on** (a hard cutover — there is no flag to disable it) and
 **bounded** by `delivery.acceptanceEval.maxRounds` (default 2; `0` means the
 verdict is scored once with no redraft round). It is **distinct from** the
-per-run `sibling-coherence` epilogue
-step (`planRunEpilogue` for N>1): this loop is per-Story, per-criterion,
-mid-delivery, and evaluates the actual work product.
+per-run epilogue (`planRunEpilogue` for N>1): this loop is per-Story,
+per-criterion, mid-delivery, and evaluates the actual work product.
 
 ## Per round
 
-1. **Eval pass — one verdict-owner per cluster.** Exactly
-   **one** pass authors each cluster's verdict: the **fresh-context critic**
-   when the ceremony routing below resolves `fresh` (a sub-agent via the
-   `Agent` tool, *not* a continuation of your implementing turn — the
-   evaluator does not grade its own homework), or the **inline self-eval**
-   when it resolves `inline`. The resolved decision names the owner
-   explicitly (`verdictOwner: 'fresh-critic' | 'inline-self-eval'` from
-   `resolveCeremonyForRisk`). **Never run both**, and never run a
-   preliminary self-assessment pass before dispatching the fresh critic —
-   the redundant pre-pass buys no measurable quality and roughly triples
-   the acceptance-block cost. Step 3's gate is the deterministic **scorer**
-   of the one merged verdict, not a second (or third) pass over the
-   criteria.
+1. **Eval pass — one verdict owner, one verdict file.** Exactly **one** pass
+   authors the Story's verdict, and it covers **every** `acceptance[]` item in
+   one file. Which pass is named by the ceremony decision
+   (`verdictOwner: 'fresh-critic' | 'inline-self-eval'` from
+   `resolveCeremonyForRisk`), and since Story #5343 that follows the
+   **ceremony profile alone**:
 
-   > **Sub-agent type + derived-level ceremony.** When
-   > `delivery.routing.roleScopedAgents` is enabled (the **default**), dispatch
-   > the critic with `subagent_type: acceptance-critic` — it boots on the
-   > role-scoped [`acceptance-critic`](../../agents/acceptance-critic.md) context
-   > (its own system prompt, no `CLAUDE.md` @-closure) that carries the
-   > maker-blind invariant and the verdict schema standalone. When the
-   > kill-switch is **off** (`roleScopedAgents: false`), fall back to
-   > `subagent_type: general-purpose`.
-   >
-   > **Whether to spawn fresh at all is routed off the derived change level**
-   > — the same signal `review-depth.js` resolves depth from, so the two
-   > decisions cannot disagree. Derive it with one script over the Story
-   > branch (Story #5313) — never a hand-carried import block:
-   >
    > ```bash
    > node <main-repo>/.agents/scripts/ceremony-derive.js --story <storyId> --cwd <workCwd>
    > ```
    >
-   > It computes the change set once (`files`), derives the level and
-   > classes, and resolves the ceremony per cluster (`mode`, `reason`,
-   > `verdictOwner`): **`high` (the diff touches a sensitive path registered
-   > in `audit-rules.json`) → `fresh`** (spawn the critic); **`low` (it
-   > touches none) → `inline`** (the contract-identical inline fallback
-   > below); **`null` / unknown (the diff could not be enumerated) → `fresh`
-   > + full ceremony** (fail-safe). This chooses fresh-vs-inline **per
-   > cluster only — it never changes the cluster count**.
-   >
-   > The routing signal is deliberately **not** a planner-authored risk
-   > verdict: a level the plan asserted about itself was exactly the signal
-   > that could *reduce* independent checking, and nothing verified it
-   > against the diff.
-   >
-   > **Inline-critic path (low-level-routed OR nesting-absent harness).** The
-   > verdict is authored **inline** whenever the risk router above resolves to
-   > `inline` (a low-risk cluster), and also as
-   > a **fallback** on any harness that cannot spawn the fresh critic.
-   > Dispatching the critic as a nested `Agent` is the fresh-context shape and
-   > works on any harness that carries `Agent` into sub-agents (Claude Code ≥
-   > 2.1.202). This eval loop itself runs inside a Story delivery
-   > sub-agent, so the nested
-   > critic sits at nesting depth 2. If the host does **not** support nested
-   > `Agent` dispatch at that depth — the tool is absent, or a spawn attempt
-   > returns an unsupported-capability error — do **not** stall the Story
-   > regardless of the risk verdict. Author the verdict **inline**: in a
-   > deliberately scoped,
-   > self-critical pass (re-read only the diff, the `acceptance[]` /
-   > `verify[]` arrays, and the `verify[]` command output — treat the
-   > implementation reasoning as untrusted and score against the criteria
-   > afresh), write the same verdict file described below and hand it to the
-   > same `acceptance-eval.js` gate. The fresh-context isolation is weaker in
-   > the inline path, but the gate, the schema, the round cap, and the
-   > proceed / redraft / block decision are identical — a Story is **never**
-   > stranded on a nesting-absent harness. Note in the blocked/friction
-   > comment (if you block) that the inline fallback was used.
+   > One call computes the change set once (`files`), derives the level and
+   > classes **for review depth**, and resolves the owner (`mode`, `reason`,
+   > `verdictOwner`): **`minimal` / `standard` → `inline`** (the default — you
+   > author the verdict yourself), **`strict` → `fresh`** (dispatch the
+   > maker-blind critic). The derived level no longer routes this decision;
+   > it escalates `review-depth.js` instead, which still resolves `deep` for
+   > any sensitive path.
+
+   **Never run both**, and never run a preliminary self-assessment before
+   dispatching a fresh critic — the redundant pre-pass buys no measurable
+   quality and roughly triples the acceptance-block cost. Step 3's gate is the
+   deterministic **scorer** of that one verdict, not a second pass over the
+   criteria.
 
-   The critic:
-   + Inspects the **change set handed to it in its spawn context** — the one
-     list computed above — and the Story's inline `acceptance[]` / `verify[]`
-     arrays. Pass the file list explicitly when you dispatch the critic; it
-     does not re-enumerate the diff for itself, so a commit
-     landing mid-ceremony cannot leave the critic scoring a different change
-     than the one that routed it.
+   > **Inline owner (the default).** Author the verdict in a deliberately
+   > scoped, self-critical pass: re-read only the diff, the `acceptance[]` /
+   > `verify[]` arrays, and the `verify[]` command output — treat your own
+   > implementation reasoning as untrusted and score each criterion afresh
+   > from the evidence. Write one verdict file covering every item and hand it
+   > to the gate. This is also the **fallback** on any harness that cannot
+   > spawn the fresh critic, so a Story is never stranded: the gate, the
+   > schema, the round cap and the proceed / redraft / block decision are
+   > identical either way. Note in the friction comment (if you block) when
+   > the fallback was used in place of a `strict` critic.
+   >
+   > **Fresh critic (`strict` only).** Dispatch a sub-agent via the `Agent`
+   > tool — *not* a continuation of your implementing turn. When
+   > `delivery.routing.roleScopedAgents` is enabled (the **default**), use
+   > `subagent_type: acceptance-critic`: it boots on the role-scoped
+   > [`acceptance-critic`](../../agents/acceptance-critic.md) context (its own
+   > system prompt, no `CLAUDE.md` @-closure) carrying the maker-blind
+   > invariant and the verdict schema standalone. With the kill-switch off
+   > (`roleScopedAgents: false`), fall back to
+   > `subagent_type: general-purpose`. This loop already runs inside a Story
+   > delivery sub-agent, so the critic sits at nesting depth 2 — supported by
+   > any harness that carries `Agent` into sub-agents (Claude Code ≥ 2.1.202).
+
+   Whichever pass owns it, the verdict:
+   + Inspects the **change set it was handed** — the one `files` list above —
+     and the Story's inline `acceptance[]` / `verify[]` arrays. Pass the file
+     list explicitly when dispatching a fresh critic; it does not re-enumerate
+     the diff for itself, so a commit landing mid-ceremony cannot leave it
+     scoring a different change than the one that routed it.
    + **Runs the `verify[]` commands** and consumes their output as **required
-     evidence** when scoring the relevant acceptance items. `verify[]` is not
-     optional advisory pre-flight — a criterion cannot be scored `met` without
-     the supporting `verify[]` evidence where a `verify[]` command is relevant
-     to it.
+     evidence**. `verify[]` is not optional advisory pre-flight — a criterion
+     cannot be scored `met` without the supporting `verify[]` evidence where a
+     `verify[]` command is relevant to it.
    + **Reuses the credited full-suite run instead of re-paying for it.**
      Before spawning a `verify[]` entry, classify it with `resolveVerifyCredit`
      from
      [`verify-credit.js`](../../scripts/lib/orchestration/verify-credit.js): an
-     entry that is itself a full-suite command (`npm test`, `pnpm run test`,
-     a bare `node --test`) is consulted against the **same stamp close reads**
-     and, when that stamp is fresh, recorded as `pass` with a `detail` naming
-     the credit — **never respawned**. A stale or absent stamp reports
-     `spawn: true` and the command runs for real, so the credit can never
-     manufacture a pass. The gate warns on any such entry: the intended shape
-     is scoped `verify[]` entries **plus** the one credited run
-     ([`deliver-digest.md`](deliver-digest.md) § 5).
-   + **Shares `lint` / `typecheck` evidence with close.** When a
-     `verify[]` command is **byte-identical** to a close-validation gate — in
-     practice only the cheap, command-identical `lint` and `typecheck` gates
-     (`npm run lint` and the resolved `project.commands.typecheck`) — the
-     critic MUST run it through `evidence-gate.js` so a passing run records an
-     evidence entry in the **same keyspace** `close-validation/runner.js`
-     consults. Run it in the **same Story worktree** the close validates (the
-     HEAD-sha key enforces "unchanged HEAD") and pass the exact gate name:
-
-     ```bash
-     node <main-repo>/.agents/scripts/evidence-gate.js \
-       --standalone --scope-id <storyId> --gate lint \
-       --worktree <worktree> -- npm run lint
-     ```
-
-     Close's `shouldSkip` then short-circuits that gate when HEAD is
-     unchanged; a redraft round (HEAD moves) correctly busts it. **Never**
-     run the coverage / CRAP suite through `evidence-gate.js` to stamp it
-     fresh — a false-fresh coverage record without `coverage-final.json`
-     silently weakens the floor. Limit the evidence-share to `lint` and
-     `typecheck`.
-   + Emits a **cluster** verdict file under `temp/` conforming to
+     entry that is itself a full-suite command is consulted against the same
+     stamp close reads and, when that stamp is fresh, recorded as `pass`
+     without being respawned; a stale or absent stamp reports `spawn: true` and
+     the command runs for real. The credited run itself is stated once, in
+     [`deliver-digest.md`](deliver-digest.md) § 5.
+   + Emits **one** verdict file under `temp/` conforming to
      [`acceptance-eval-verdict.schema.json`](../../schemas/acceptance-eval-verdict.schema.json):
      one `{ index, criterion, verdict: met|partial|unmet, evidence,
-     verifyEvidence[] }` record per acceptance item **in that cluster**, each
-     `index` being the item's position in the Story's full `acceptance[]`
-     array. A fresh critic **returns that path to you** rather than calling the
-     gate itself.
-2. **Dispatch the round's clusters in parallel, then merge into one verdict
-   (fresh critics only).** When the verdict owner is the **inline self-eval**,
-   author **one** verdict file covering every `acceptance[]` item and score it
-   in one gate call — the cluster merge below does not apply (Story #5313).
-   For fresh critics, the clusters of a round are independent, so dispatch
-   **all** of the round's critics as N `Agent` calls **in a single assistant
-   turn** — [`parallel-tooling.md`](parallel-tooling.md) **Rule 3** — never
-   serially, and never one round per cluster.
-
-   Then **merge** the cluster verdicts into **one** verdict file under `temp/`:
-   concatenate every cluster's `criteria[]` records and order the merged array
-   by `index`, so it holds exactly one record per `acceptance[]` item in
-   **acceptance-array order**, under a single top-level `storyId`,
-   `schemaVersion`, `round` and `commitSha`. The verdict schema deliberately
-   carries **no `clusterId`** — the round's artifact is the merged verdict, and
-   which critic scored which record is not part of the contract.
-
-   > **Why one gate call and not N.** The round counter is **Story-scoped** —
-   > derived by counting `acceptance-eval` signals in the Story's
-   > `signals.ndjson` — and each cluster verdict has a distinct fingerprint, so
-   > the replay guard never collapses them. A gate call per cluster would spend
-   > one of the (default 2) rounds *per cluster*, so a Story with more than 8
-   > acceptance criteria would exhaust its redraft budget on cluster arithmetic
-   > alone; N concurrent calls would also race that same ledger. Cluster-scoped
-   > round counting exists in
-   > [`acceptance-eval-decision.js`](../../scripts/lib/orchestration/acceptance-eval-decision.js)
-   > but requires an integer `epicId`, which v2 pins `null` — it is not a way
-   > around the merge.
-3. **Decide — exactly one gate call per round.** Run the gate against the
-   **merged** verdict (the caller's Step 1a names the exact invocation — omit
-   `--epic`). The gate **scores the single verdict the round produced** —
-   schema validation, round cap, decision — and never re-scores the criteria
-   itself:
+     verifyEvidence[] }` record per `acceptance[]` item, in acceptance-array
+     order, under a single top-level `storyId`, `schemaVersion`, `round` and
+     `commitSha`. A fresh critic **returns that path to you** rather than
+     calling the gate itself.
+2. **Decide — exactly one gate call per round.** Run the gate against that one
+   verdict file (the caller's Step 1a names the exact invocation — omit
+   `--epic`). The gate **scores the verdict the round produced** — schema
+   validation, round cap, decision — and never re-scores the criteria itself:
 
    ```bash
    node <main-repo>/.agents/scripts/acceptance-eval.js \
@@ -191,10 +113,16 @@ mid-delivery, and evaluates the actual work product.
    ```
 
    The gate reads the Story's `acceptance[]` count itself (Story #5313): a
-   verdict whose `criteria[]` length differs — a single cluster's verdict
-   handed over unmerged — is rejected **before scoring**, with an error naming
-   the merge contract and consuming **no round**. `--expected-criteria` is
-   still accepted but redundant; when passed it must agree with that count.
+   verdict whose `criteria[]` length differs — one covering only part of the
+   Story — is rejected **before scoring**, with an error naming the count and
+   consuming **no round**. `--expected-criteria` is still accepted but
+   redundant; when passed it must agree with that count.
+
+   > **Why one call per round and not several.** The round counter is
+   > **Story-scoped** — derived by counting `acceptance-eval` signals in the
+   > Story's `signals.ndjson` — so a second call in the same round spends one
+   > of the (default 2) rounds for nothing, and concurrent calls race that
+   > ledger.
 
    The gate validates the verdict against the schema, applies the round cap,
    emits the per-criterion `acceptance-eval` signal into the retro / feedback
@@ -209,5 +137,4 @@ mid-delivery, and evaluates the actual work product.
      (transition to `agent::blocked`) and post a `friction` comment naming the
      unmet criteria and their evidence. Never silently proceed to close.
 
-Write both the per-cluster verdicts and the merged verdict under `temp/` only —
-they are scratch artifacts.
+Write the verdict under `temp/` only — it is a scratch artifact.

package/.agents/workflows/helpers/audit-lens-core.md

@@ -152,7 +152,7 @@ Treat the following as **out of scope by construction** and drop any finding
 that rests on one of them:
 
 - **Sanctioned test seams** — exports consumed only by tests, and other
-  patterns the `test-seams` rule blesses. Never a production defect.
+  deliberate test-only patterns. Never a production defect.
 - **Entry points & public API surface** — CLI mains, `bin/` scripts, declared
   `exports` / `bin` / `main`, and barrel contracts consumed out-of-tree. A
   zero in-repo consumer count is not death.

package/.agents/workflows/helpers/code-quality-guardrails.md

@@ -5,11 +5,11 @@ this framework cites. Reviewers, prescriptive auditors, and writing agents all
 read the same numbers from here so a "high cyclomatic complexity" finding in
 `/audit-clean-code` and a pre-commit refusal cite the same threshold.
 
-> Tunable via `delivery.quality.codingGuardrails` in `.agentrc.json` —
-> see [`agentrc-reference.json`](../../docs/agentrc-reference.json) for the framework
-> defaults and [`schemas/agentrc.schema.json`](../../schemas/agentrc.schema.json)
-> for the validated shape. Override per-project; never fork this helper to
-> change a number.
+> The cyclomatic thresholds are fixed framework constants (Story #5382 folded
+> the never-set `delivery.quality.codingGuardrails` block); the MI tolerance
+> is `delivery.quality.gates.maintainability.tolerance` — see
+> [`agentrc-reference.json`](../../docs/agentrc-reference.json). Never fork
+> this helper to change a number.
 
 ## At-keyboard verification
 
@@ -28,14 +28,14 @@ hook calls the same script.
 
 Cyclomatic complexity (CC) is measured per function by `escomplex` (the same
 engine the maintainability axis of `check-baselines.js` runs). Two
-thresholds — the advisory `delivery.quality.codingGuardrails.cyclomaticFlag`
-and the fixed ceiling of 12 (`cyclomaticMustFix` was retired as a config
-key in Story #5313):
+thresholds — the advisory flag of 8 and the fixed ceiling of 12, both
+constants (`cyclomaticMustFix` was retired as a config key in Story #5313,
+`cyclomaticFlag` in Story #5382):
 
 | CC range | Action |
 | --- | --- |
 | ≤ 8 | Pass — no annotation required. |
-| > 8 (default `cyclomaticFlag`) | **Flag** — `quality:preview` counts the function in its `new-method count over c=<flag>` column. The function is allowed to land but the report names it. |
+| > 8 (the advisory flag) | **Flag** — `quality:preview` counts the function in its `new-method count over c=<flag>` column. The function is allowed to land but the report names it. |
 | ≥ 12 | **Advisory in `quality:preview`** — listed by file, method and reading; the preview exits 0 on it. **Ratchet in `check-cyclomatic.js`**: fails when a file gains a function above 12, or when its worst function gets worse than the recorded baseline. |
 
 `check-cyclomatic.js` is a **ratchet**, not a cliff: `baselines/cyclomatic.json`
@@ -57,11 +57,8 @@ source under `tests/<mirrored-path>/<basename>.test.js`. Same-commit pairing
 keeps the bisect honest — a regression and the test that would have caught
 it land or revert together.
 
-When `delivery.quality.codingGuardrails.requireSiblingTest` is `true`,
-the pre-commit hook refuses to commit a staged new source file that
-lacks a sibling test. Default is `false` so legacy repos opt in
-deliberately; once enabled, the structural check replaces the
-review-time prose rule.
+No structural check enforces the pairing — it is a review-time rule
+(Story #5382 removed the never-read `requireSiblingTest` switch).
 
 ## Maintainability-Index drop refactor rule
 
@@ -84,7 +81,7 @@ the commit lands (scoped to `HEAD` by default — the alias passes no
 Renaming a file (or moving it across `src/` directories) detaches its MI and
 CRAP history from the baseline keys. The correct response is a **baseline
 refresh**, not a regression entry: include `baseline-refresh:` (the
-`delivery.quality.crap.refreshTag` default) in the commit subject so the
+fixed refresh tag) in the commit subject so the
 ratchet treats the new key as a clean entry instead of comparing it to
 nothing.
 

package/.agents/workflows/helpers/code-review.md

@@ -87,8 +87,10 @@ review yourself, honor the `depth` semantics above directly.
 1. Resolve `[TICKET_ID]` from `ticketId` (the Story for live `scope: story`).
 2. Resolve `[BASE_REF]` from `baseRef` and `[HEAD_REF]` from `headRef`.
 3. Fetch the Story ticket and resolve the planning context from its own
-   body: folded `## Spec` / `## Slicing`, acceptance criteria, and the
-   `story-plan-state` structured comment when present.
+   body: folded `## Spec` / `## Slicing`, and acceptance criteria. The
+   `story-plan-state` comment beside it carries the operator's plan summary
+   (story set, delivery order, deliver command), not planning context — it
+   holds no machine payload to read.
 4. Read that Spec fully to understand the intended scope, architectural
    decisions, and acceptance criteria. Do **not** look for a parent Epic.
 

package/.agents/workflows/helpers/deliver-digest.md

@@ -25,18 +25,19 @@ rule produces it:
    shape — sub-agent isolation only matters against a *concurrent* sibling
    racing the same checkout, and a one-Story run has none.
 2. **Every other run is `subagent`.** A multi-Story run dispatches every Story
-   as a sub-agent however trivial its shape. Shape still sets ceremony; the
-   `route::lite` label is a human-visible hint, never the control signal.
+   as a sub-agent however trivial its shape — the run's size is the whole
+   premise, and nothing about a Story's own shape enters it.
 
-`inline` removes model-side fan-out only — no `story-worker` boot, no fresh
-acceptance-critic spawn. **`subagent` and `inline` run the same engine**: same
-gates, same PR to `main`, same terminal envelope, byte for byte.
+`inline` removes the `story-worker` boot and nothing else. It does **not**
+change the acceptance verdict owner — that is § 3's decision, and the profile
+alone makes it. **`subagent` and `inline` run the same engine**: same gates,
+same PR to `main`, same terminal envelope, byte for byte.
 
 ## 2. Engine invariants
 
 | Trait | Contract |
 | --- | --- |
-| Ticket type | `type::story` only; an `Epic: #N` footer means **stop and re-plan** |
+| Ticket type | `type::story` only — `resolve-stories.js` validates it |
 | Branch | `story-<id>`, seeded from `project.baseBranch` (`main`) |
 | Merge target | `main` via PR (squash + required checks) — never a direct push |
 | Gates | Every close gate runs regardless of route; no route bypasses one |
@@ -57,38 +58,44 @@ import block:
 node <main-repo>/.agents/scripts/ceremony-derive.js --story <storyId> --cwd <workCwd>
 ```
 
-It prints one JSON object: `files` — the one change set every critic is
+It prints one JSON object: `files` — the one change set the verdict owner is
 handed (`null` when the diff could not be enumerated) — plus `level` and
 `classes` from `review-depth.js`, and `mode`, `reason` and `verdictOwner`
 from `ceremony-routing.js`. Level rules: a sensitive path registered in
 `audit-rules.json` → `high`, none → `low`, an unenumerable diff → `null`.
-Ceremony rules: `minimal` → always inline, `strict` → always fresh,
-`standard` → `high`/`null` → fresh and `low` → inline. An `inline` dispatch
-mode overrides all of it to inline critics. Close's `review-depth.js` reads
-the same derived level, so the two cannot disagree. `--base <ref>` overrides
-`project.baseBranch`.
+The level drives **review depth** only; close's `review-depth.js` reads the
+same derived level, so the two cannot disagree. A sensitive footprint
+therefore buys a **deep review**, not a fresh acceptance critic.
+
+> **The ceremony rule, stated once.** The **profile alone** names the verdict
+> owner (Story #5343, narrowed to that one input by #5366): `minimal` /
+> `standard` → `inline`, `strict` → `fresh`. Nothing else moves it — not the
+> derived change level, not the footprint's sensitivity, and **not the
+> dispatch mode**: an `inline` Story under `strict` still spawns the fresh
+> maker-blind critic, one nesting level shallower than a dispatched one. The
+> only sanctioned inline authoring under `strict` is the harness fallback in
+> [`acceptance-self-eval.md`](acceptance-self-eval.md) — a host that cannot
+> spawn the critic at all, noted in the friction comment if you block.
+
+`--base <ref>` overrides `project.baseBranch`.
 
 ## 4. Acceptance self-eval (Step 1a, required)
 
-**One verdict-owner per cluster** — the fresh critic *or* the inline
-self-eval, named by `verdictOwner`, never both and never a warm-up pass. Each
-scores its cluster's `acceptance[]` items against the change set above, with
-`verify[]` output as evidence. Bounded by `delivery.acceptanceEval.maxRounds`
+**One verdict owner per Story** — named by `verdictOwner`: the inline
+self-eval under `minimal` / `standard` (the default), a fresh maker-blind
+critic under `strict`. Never both, and never a warm-up pass. The owner
+authors **one** verdict file covering every `acceptance[]` item, scored
+against the change set above with `verify[]` output as evidence, and it is
+scored in **one** gate call. Bounded by `delivery.acceptanceEval.maxRounds`
 (default 2; `0` scores once with no redraft).
 
-**Inline owner:** author **one** verdict file covering every `acceptance[]`
-item and score it in **one** gate call — there is no cluster merge.
-**Fresh critics:** one round = N cluster critics → ONE merged verdict → ONE
-gate call. Merge every cluster's records into a single `criteria[]` in
-`acceptance[]` order and score that once; a gate call per cluster spends a
-round *per cluster* and races the round ledger.
-
 `node <main-repo>/.agents/scripts/acceptance-eval.js --story <storyId>
 --verdict <verdict-path>`
 
 The gate reads the Story's `acceptance[]` count itself and rejects a verdict
 whose `criteria[]` length differs **before** scoring, consuming no round;
-`--expected-criteria` is accepted but redundant.
+`--expected-criteria` is accepted but redundant. A second gate call in the
+same round spends a round for nothing and races the Story-scoped ledger.
 
 `proceed` → close. `redraft` → one more round inside the cap. `block` → **do
 not close**: post a `friction` comment and flip `agent::blocked`.
@@ -96,10 +103,12 @@ Per-round mechanics: [`acceptance-self-eval.md`](acceptance-self-eval.md).
 
 ## 5. The one credited suite run
 
-After the self-eval loop's last fix commit, run the suite **once** in the
-worktree through the depositor — it spawns the project's own `npm test`,
-whatever that resolves to, and stamps the result, so any runner earns the
-credit:
+After the self-eval loop's last fix commit, fetch and merge
+`origin/<baseBranch>` into the Story branch **first**, ahead of this run and
+the push: close's base-sync then no-ops, so neither stamp goes stale. Then
+run the suite **once** in the worktree through the depositor — it spawns the
+project's own `npm test`, whatever that resolves to, and stamps the result,
+so any runner earns the credit:
 
 ```bash
 node <main-repo>/.agents/scripts/evidence-gate.js --standalone \