@gaunt-sloth/batch 2.0.0-beta.8 → 2.0.0

package/dist/evalTypes.d.ts

@@ -1,13 +1,15 @@
 /**
  * @packageDocumentation
  * BATCH-2 — the shapes for `gth eval`: a parsed suite/case, deterministic-check results, the
- * judge's verdict, and per-case/suite outcomes. Deliberately separate from {@link ../types.js}
+ * judge's verdict, and per-case/suite outcomes. Deliberately separate from `types.js`
  * (BATCH-1's cell/outcome shapes), which documents itself as scoped to "cells and outcomes" only
  * — eval's shapes layer on top of (not into) that file.
  */
 import type { ApprovalRung } from '@gaunt-sloth/core/config/shell-policy.js';
 import type { PreflightFloorKind } from '@gaunt-sloth/core/core/shell/raterVocabulary.js';
 import type { ToolResultRecord } from '#src/types.js';
+import type { RaterPromptArm } from '#src/raterPromptArm.js';
+import type { AdvertisedToolInventory, ToolCoverageReport, ToolCoverageSpec } from '#src/toolCoverage.js';
 import type { EvalCaseClassification, EvalClassificationReport, EvalClassificationSpec, EvalMetricSpec } from '#src/classificationTypes.js';
 /** The 0-10 judge scale's default pass threshold, matching `review`'s own (unexported)
  * `DEFAULT_PASS_THRESHOLD` in `packages/review/src/middleware/reviewRateMiddleware.ts` — same
@@ -71,7 +73,7 @@ declare const PREFLIGHT_MECHANISM_OF: {
  * The floor is spelled here because it is not a preflight — it is a lexical scan core keeps no list
  * of, consulted by the approvals gate before any rating (every rung but `bypass`) and again by the
  * shell tool at execution time (every rung). The preflights are NOT spelled: they come from
- * {@link PREFLIGHT_MECHANISM_OF}, which is core's own `PREFLIGHT_FLOOR_KINDS` with our `forced_by`
+ * `PREFLIGHT_MECHANISM_OF`, which is core's own `PREFLIGHT_FLOOR_KINDS` with our `forced_by`
  * spelling attached.
  *
  * ## Why a case asserts a mechanism at all (the I1 finding)
@@ -158,7 +160,7 @@ export declare const PREFLIGHT_MECHANISMS: ("open-world-preflight" | "script-env
 export type PreflightMechanism = (typeof PREFLIGHT_MECHANISM_OF)[PreflightFloorKind];
 /**
  * The `forced_by` spelling of one of core's preflight arms — the typed lookup into
- * {@link PREFLIGHT_MECHANISM_OF}, which stays private so this is the only way in.
+ * `PREFLIGHT_MECHANISM_OF`, which stays private so this is the only way in.
  *
  * TOTAL over core's `PreflightFloorKind` and therefore never `undefined`, which is the point: a
  * caller cannot be handed an arm it has no name for, because such an arm would not have compiled.
@@ -178,7 +180,7 @@ export declare function mechanismNeedsPermissiveRating(mechanism: ForcedByMechan
 export interface GthAgentTarget {
     type: 'gth-agent';
     /** Suite-level profile hint. Only `undefined`/`'default'` is accepted (see
-     * {@link ../evalSuite.js}'s `parseEvalSuite`) — per-case/per-identity profile switching is
+     * `evalSuite.js`'s `parseEvalSuite`) — per-case/per-identity profile switching is
      * `identities` scope, not this. */
     profile?: string;
 }
@@ -240,7 +242,7 @@ export interface AgUiAgentTarget {
  * §8 hardline floor — decides some commands outright, with no model in the loop.
  *
  * Unlike the agent targets this one is NOT driven by a `RunCellFn`: it supplies the
- * {@link RunClassifyFn} seam instead (`buildRaterClassifier` in {@link ../raterTarget.js}), which
+ * {@link RunClassifyFn} seam instead (`buildRaterClassifier` in `raterTarget.js`), which
  * the command passes to `runEvalSuite` as `options.classify`.
  */
 export interface RaterTarget {
@@ -252,7 +254,7 @@ export interface RaterTarget {
      * It is the suite's declaration, not the last word: a run whose config declares `approvals`
      * overrides it, which is what makes the `rung × model` sweep work through the existing `config:`
      * axis (a sweep cell cannot reach a `target` field). The override is announced, never silent —
-     * see {@link ../raterTarget.js}.
+     * see `raterTarget.js`.
      */
     rung: ApprovalRung;
 }
@@ -263,7 +265,7 @@ export interface RaterTarget {
  * target), so the target only changes which runner the command builds. */
 export type EvalTarget = GthAgentTarget | AdkAgentTarget | AgUiAgentTarget | RaterTarget;
 /** One `json_path` assertion (BATCH-10): resolve `path` against the answer-parsed-as-JSON and check
- * it. Exactly one of `equals`/`contains` is set (enforced in {@link ../evalSuite.js}'s parse):
+ * it. Exactly one of `equals`/`contains` is set (enforced in `evalSuite.js`'s parse):
  * - `equals` — the resolved value must deep-equal this (any JSON value, incl. `null`).
  * - `contains` — the resolved value must be a string containing this substring. */
 export interface JsonPathCheck {
@@ -275,7 +277,7 @@ export interface JsonPathCheck {
  * One `tool_result_json_path` assertion (BATCH-21): select tool results by name `tool` (exact or
  * glob, the same matcher `must_call` uses), parse each matching result's payload as JSON, and
  * resolve `path` against it (the same minimal dot/`[index]` path `json_path` uses). At most one of
- * `equals`/`contains` may be set (enforced in {@link ../evalSuite.js}'s parse):
+ * `equals`/`contains` may be set (enforced in `evalSuite.js`'s parse):
  * - neither — pure existence check: the path must resolve in a matching result's payload;
  * - `equals` — the resolved value must deep-equal this (any JSON value, incl. `null`);
  * - `contains` — the resolved value must be a string containing this substring.
@@ -359,6 +361,31 @@ export interface EvalExpectation {
      * touching the first.
      */
     expectAction?: string;
+    /**
+     * BATCH-45 — **assert that a MODEL actually rendered a verdict for this round**, i.e. that
+     * {@link ClassifyOutcome.modelLabel} is present. `true` is the only legal value; `undefined` =
+     * this block makes no such claim. `rater` target only.
+     *
+     * It is NOT an assertion about the rung, and NOT an assertion about WHAT the model said. Those
+     * are {@link expectLabel} and {@link expectAction}, and the whole point of this key is that it
+     * says neither: a case can pin "a rater ruled here" without predicting which verdict three
+     * different raters would render, which is a prediction no measurement backs.
+     *
+     * **Why the action column cannot do this job.** When the gate never obtains a rating (timeout,
+     * throw, unparseable answer) it fails closed, and since EXT-171 that ESCALATES rather than
+     * negotiating. A genuine `catastrophic` verdict escalates too. So `expect_action: escalate` is
+     * satisfied identically by a rater that judged the command unnegotiable and by a rater that
+     * never answered at all — and a case asserting only that passes on a run where nothing was
+     * measured. BATCH-45 found ten such cells live in this repo's own approvals corpus.
+     *
+     * **Why `modelLabel` and not the rationale text.** `modelLabel` is suppressed from the call's own
+     * capture (`rating.failClosed`), never from a reading of the verdict's prose. Core's
+     * `isFailClosed` answers a question about the reason TEXT, and the rating prompt tells a rater to
+     * answer `destructive` and say it could not assess a command it is unsure of — so that predicate
+     * scores an obedient judgement as a gate failure. EXT-171 and `raterHealth` both say never to key
+     * on verdict text.
+     */
+    expectRated?: boolean;
 }
 /**
  * One conversational turn (BATCH-12): the `user` message to send, and the {@link EvalExpectation}
@@ -378,7 +405,7 @@ export interface EvalTurn {
      *
      * **Carrying it is not the same as showing it to the rater.** §5.1 admits the justification from
      * round 2 onward, so the round-1 rating withholds it; the target hands the whole context to core's
-     * {@link import('@gaunt-sloth/core/core/shell/negotiation.js').ShellNegotiationState}, which is
+     * {@link @gaunt-sloth/core!core/shell/negotiation.ShellNegotiationState | ShellNegotiationState}, which is
      * the single implementation of that rule.
      */
     justification?: string;
@@ -410,7 +437,7 @@ export interface EvalTurn {
 }
 /**
  * One turn's raw run outcome inside a multi-turn conversation (BATCH-12 Task 2). Structurally the
- * per-turn analogue of BATCH-1's {@link ../types.js CellRunOutcome}: the turn's answer plus the
+ * per-turn analogue of BATCH-1's {@link "types.js"!CellRunOutcome | CellRunOutcome}: the turn's answer plus the
  * tools/tokens captured FOR THAT TURN — a per-invoke GS2-16 delta (each `processMessages` call
  * resets the tally), NOT the cumulative conversation total. `ok:false` with `error` set means that
  * turn's SUT invocation failed (no answer to grade). The conversational runner returns one of these
@@ -425,12 +452,18 @@ export interface TurnRunOutcome {
     /** BATCH-21 — this turn's per-tool-call result records (parallel to {@link tools}; a per-turn
      * delta like everything else here). Only the in-process `gth-agent` runner populates it. */
     toolResults?: ToolResultRecord[];
+    /**
+     * BATCH-32 — the advertised-tool inventory (the coverage denominator). **The one field here that
+     * is NOT a per-turn delta**: a conversation builds its agent once, so the inventory is fixed for
+     * the whole conversation and every turn repeats it.
+     */
+    advertisedTools?: AdvertisedToolInventory;
     error?: string;
 }
 /**
  * Injectable "run one whole scripted conversation" function (BATCH-12 Task 2) — the multi-turn
- * analogue of BATCH-1's {@link ../types.js RunCellFn}, and the seam that lets
- * {@link ../evalRunner.js}'s multi-turn path be unit tested without any real LLM/MCP. Given the
+ * analogue of BATCH-1's {@link "types.js"!RunCellFn | RunCellFn}, and the seam that lets
+ * `evalRunner.js`'s multi-turn path be unit tested without any real LLM/MCP. Given the
  * ordered user messages of ONE (case × identity) conversation, it builds the agent + resolves tools
  * ONCE, runs each turn against the accumulated message history, and returns one {@link TurnRunOutcome}
  * per turn (per-turn answer + per-turn tool delta), cleaning up once. The production wiring
@@ -447,7 +480,7 @@ export type RunConversationFn = (userMessages: string[]) => Promise<TurnRunOutco
  * a round-1 rating, and a reset makes a later round a round-1 rating again — so the target hands the
  * accumulated state to `ShellNegotiationState.contextFor` rather than deciding here. A shape that
  * carried "what the rater sees" would be this package holding a second opinion about §5.1, which is
- * the one thing {@link ../raterTarget.js} exists not to do.
+ * the one thing `raterTarget.js` exists not to do.
  */
 export interface ClassifyRound {
     /** The command the agent proposes in this round. */
@@ -465,7 +498,7 @@ export interface ClassifyRound {
  * ## The Half-B seam, now filled
  *
  * Half A shipped this shape, the runner plumbing and the tests (against an injected fake); Half B
- * supplies the one function — `buildRaterClassifier` in {@link ../raterTarget.js}, the
+ * supplies the one function — `buildRaterClassifier` in `raterTarget.js`, the
  * {@link RaterTarget}'s implementation, which drives the approvals rating prompt + decision mapping
  * at a declared rung. The shape did not have to move to fit it.
  *
@@ -641,7 +674,7 @@ export interface EvalCase {
      */
     modelFree: boolean;
 }
-/** A fully parsed and validated suite — see {@link ../evalSuite.js}'s `parseEvalSuite`. */
+/** A fully parsed and validated suite — see `evalSuite.js`'s `parseEvalSuite`. */
 export interface EvalSuite {
     target: EvalTarget;
     /**
@@ -669,11 +702,21 @@ export interface EvalSuite {
      * label enum has nothing to read).
      */
     metrics: EvalMetricSpec[];
+    /**
+     * BATCH-32 — the suite's `tool_coverage:` declaration: waivers, an optional floor, and required
+     * tools. Absent = coverage is still REPORTED (that is the whole point of the node: a run that
+     * exercises 3 of 41 tools must say so without being asked), but nothing gates on it.
+     *
+     * A top-level key of its own rather than a `metrics:` entry: a metric is a predicate over LABELLED
+     * cases and its machinery assumes a `classification:` block, whereas coverage applies to suites
+     * with no labels at all — most of them.
+     */
+    toolCoverage?: ToolCoverageSpec;
     /**
      * BATCH-25 — the config sweep: named cells the WHOLE suite is run once per, so one corpus
      * produces one comparison table instead of N unrelated runs. Absent = a single run.
      *
-     * Consumed by the CLI, not by {@link ../evalRunner.js runEvalSuite}: a sweep is a run-level
+     * Consumed by the CLI, not by {@link "evalRunner.js"!runEvalSuite | runEvalSuite}: a sweep is a run-level
      * concept (same suite, different config) exactly like BATCH-19's multi-suite loop, so the runner
      * stays about grading and #405's identity matrix is untouched.
      */
@@ -698,6 +741,17 @@ export interface EvalSweepValue {
     model?: string;
     /** Plain-data config overrides deep-merged onto the resolved config (never `llm`). */
     config?: Record<string, unknown>;
+    /**
+     * [[BATCH-31]] — the cell's PROMPT ARM: which of gth's own preflight notes this cell's ratings go
+     * out without, so a note-on / note-off A/B is one suite file rather than two. `rater` targets
+     * only; `{ omit: [] }` is the baseline arm and is required rather than implied.
+     *
+     * **A third override kind, not a third spelling of `config:`.** It is deliberately not a config
+     * key and never merges into one: a config key is exactly what a session could also be given, and
+     * an omission a session can reach is a switch that suppresses safety context in the live approvals
+     * gate. See `raterPromptArm.ts` for the full argument and for how that is enforced.
+     */
+    notes?: RaterPromptArm;
 }
 /**
  * BATCH-25 — the sweep: named axes whose CARTESIAN PRODUCT is the set of runs. The approvals
@@ -719,7 +773,7 @@ export interface DeterministicCheckResult {
     failures: string[];
 }
 /** The judge's structured verdict on one case's answer — matches `review`'s `RateSchema` shape
- * (0-10 `rate` + a reason string) for UX consistency, see {@link ../judge.js}. */
+ * (0-10 `rate` + a reason string) for UX consistency, see `judge.js`. */
 export interface JudgeVerdict {
     rate: number;
     reason: string;
@@ -737,7 +791,7 @@ export interface JudgeOutcome {
     error?: string;
 }
 /** Injectable "grade one answer against one rubric" function — the seam that lets
- * {@link ../evalRunner.js}'s `runEvalSuite` be fully unit tested without any real LLM call, mirror
+ * `evalRunner.js`'s `runEvalSuite` be fully unit tested without any real LLM call, mirror
  * of BATCH-1's `RunCellFn`. The production wiring (`evalCommand.ts`) adapts `judgeEvalCase` to
  * this shape; tests inject a fake that resolves/fails as needed. */
 export type JudgeFn = (answer: string, rubric: string) => Promise<JudgeOutcome>;
@@ -830,5 +884,12 @@ export interface EvalSuiteSummary {
     /** BATCH-25 — the confusion matrices + declared metrics. Omitted entirely for a suite with no
      * `classification:` block, so a pre-BATCH-25 `results.json` is byte-for-byte unchanged. */
     classification?: EvalClassificationReport;
+    /**
+     * BATCH-32 — which of the agent's advertised tools this suite exercised. Omitted entirely when no
+     * cell reported an inventory: an external target, or a run whose SUT never initialised. That
+     * absence is the honest answer, and it is why this is optional rather than a zeroed block — a
+     * `0/0` coverage figure looks like a measurement and is not one.
+     */
+    toolCoverage?: ToolCoverageReport;
 }
 export {};

package/dist/evalTypes.js

@@ -60,7 +60,7 @@ const PREFLIGHT_MECHANISM_OF = {
  * The floor is spelled here because it is not a preflight — it is a lexical scan core keeps no list
  * of, consulted by the approvals gate before any rating (every rung but `bypass`) and again by the
  * shell tool at execution time (every rung). The preflights are NOT spelled: they come from
- * {@link PREFLIGHT_MECHANISM_OF}, which is core's own `PREFLIGHT_FLOOR_KINDS` with our `forced_by`
+ * `PREFLIGHT_MECHANISM_OF`, which is core's own `PREFLIGHT_FLOOR_KINDS` with our `forced_by`
  * spelling attached.
  *
  * ## Why a case asserts a mechanism at all (the I1 finding)
@@ -132,7 +132,7 @@ export const FORCED_BY_ASSERTIONS = {
 export const PREFLIGHT_MECHANISMS = Object.values(PREFLIGHT_MECHANISM_OF);
 /**
  * The `forced_by` spelling of one of core's preflight arms — the typed lookup into
- * {@link PREFLIGHT_MECHANISM_OF}, which stays private so this is the only way in.
+ * `PREFLIGHT_MECHANISM_OF`, which stays private so this is the only way in.
  *
  * TOTAL over core's `PreflightFloorKind` and therefore never `undefined`, which is the point: a
  * caller cannot be handed an arm it has no name for, because such an arm would not have compiled.

package/dist/evalTypes.js.map

@@ -1 +1 @@
-{"version":3,"file":"evalTypes.js","sourceRoot":"","sources":["../src/evalTypes.ts"],"names":[],"mappings":"AAqBA;;;;mGAImG;AACnG,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC;AAE7C;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAG,yBAAyB,CAAC;AAEjE;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,sBAAsB,GAAG;IAC7B,iBAAiB,EAAE,2BAA2B;IAC9C,YAAY,EAAE,sBAAsB;CAC2B,CAAC;AAElE;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG;IAClC,gBAAgB;IAChB,GAAG,MAAM,CAAC,MAAM,CAAC,sBAAsB,CAAC;CAChC,CAAC;AAiBX,mGAAmG;AACnG,MAAM,CAAC,MAAM,gBAAgB,GAAG,WAAW,CAAC;AAE5C;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAsC;IACrE,gBAAgB,EAAE,uBAAuB;IACzC,2BAA2B,EAAE,GAAG,gBAAgB,6BAA6B;IAC7E,sBAAsB,EAAE,GAAG,gBAAgB,wBAAwB;CACpE,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsCG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG,MAAM,CAAC,MAAM,CAAC,sBAAsB,CAAC,CAAC;AAW1E;;;;;;GAMG;AACH,MAAM,UAAU,qBAAqB,CAAC,IAAwB;IAC5D,OAAO,sBAAsB,CAAC,IAAI,CAAC,CAAC;AACtC,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,8BAA8B,CAAC,SAAwC;IACrF,OAAO,CACL,SAAS,KAAK,SAAS;QACtB,oBAAqD,CAAC,QAAQ,CAAC,SAAS,CAAC,CAC3E,CAAC;AACJ,CAAC"}
+{"version":3,"file":"evalTypes.js","sourceRoot":"","sources":["../src/evalTypes.ts"],"names":[],"mappings":"AA6BA;;;;mGAImG;AACnG,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC;AAE7C;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAG,yBAAyB,CAAC;AAEjE;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,sBAAsB,GAAG;IAC7B,iBAAiB,EAAE,2BAA2B;IAC9C,YAAY,EAAE,sBAAsB;CAC2B,CAAC;AAElE;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG;IAClC,gBAAgB;IAChB,GAAG,MAAM,CAAC,MAAM,CAAC,sBAAsB,CAAC;CAChC,CAAC;AAiBX,mGAAmG;AACnG,MAAM,CAAC,MAAM,gBAAgB,GAAG,WAAW,CAAC;AAE5C;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAsC;IACrE,gBAAgB,EAAE,uBAAuB;IACzC,2BAA2B,EAAE,GAAG,gBAAgB,6BAA6B;IAC7E,sBAAsB,EAAE,GAAG,gBAAgB,wBAAwB;CACpE,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsCG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG,MAAM,CAAC,MAAM,CAAC,sBAAsB,CAAC,CAAC;AAW1E;;;;;;GAMG;AACH,MAAM,UAAU,qBAAqB,CAAC,IAAwB;IAC5D,OAAO,sBAAsB,CAAC,IAAI,CAAC,CAAC;AACtC,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,8BAA8B,CAAC,SAAwC;IACrF,OAAO,CACL,SAAS,KAAK,SAAS;QACtB,oBAAqD,CAAC,QAAQ,CAAC,SAAS,CAAC,CAC3E,CAAC;AACJ,CAAC"}

package/dist/index.d.ts

@@ -18,15 +18,21 @@ export { extractClassificationValue, buildConfusionMatrix, collectTags, readRaw,
 export { buildClassificationReport } from '#src/classificationReport.js';
 export { computeMetric, parseMetricPredicate, evaluatePredicate, evaluatePredicates, formatTally, } from '#src/metrics.js';
 export { renderClassificationReport, renderConfusionMatrix, renderMetric, } from '#src/classificationRender.js';
+export { computeToolCoverage, aggregateToolCoverage, gradeRunToolCoverage, hasRunToolCoverageSpec, WAIVER_SHARE_WARN_THRESHOLD, } from '#src/toolCoverage.js';
+export type { AdvertisedToolInventory, AdvertisedToolRecord, RunToolCoverageGrade, RunToolCoverageSource, RunToolCoverageSpec, ToolCoverageInput, ToolCoverageReport, ToolCoverageServerReport, ToolCoverageSpec, } from '#src/toolCoverage.js';
+export { renderToolCoverage } from '#src/toolCoverageRender.js';
+export type { RenderToolCoverageOptions } from '#src/toolCoverageRender.js';
 export { buildBlindExport, diffRelabel, renderRelabelDiff } from '#src/blindExport.js';
 export type { BlindExport, BlindExportCase, RelabelDiff, RelabelEntry } from '#src/blindExport.js';
-export { expandSweep, deepMerge, renderComparison, diffRuns, renderRunDiff, } from '#src/evalCompare.js';
-export type { ComparisonColumn, RunDiff, RunDiffEntry, SweepCell } from '#src/evalCompare.js';
+export { expandSweep, deepMerge, renderComparison, diffRuns, renderRunDiff, parseJudgeDriftFilter, DEFAULT_JUDGE_DRIFT_FILTER, DEFAULT_JUDGE_DRIFT_TOLERANCE, MAX_JUDGE_DRIFT_TOLERANCE, } from '#src/evalCompare.js';
+export type { ComparisonColumn, JudgeDriftEntry, JudgeDriftFilter, JudgeDriftMean, RunDiff, RunDiffEntry, SweepCell, } from '#src/evalCompare.js';
 export { UNRECOGNIZED_LABEL, NO_EXPECTATION } from '#src/classificationTypes.js';
 export type { ClassificationExtractor, ClassifiedCell, EvalCaseClassification, EvalClassificationReport, EvalClassificationSpec, EvalConfusionMatrix, EvalMetricCoverage, EvalMetricResult, EvalMetricSpec, EvalMetricTally, MetricField, MetricPredicate, } from '#src/classificationTypes.js';
 export type { ClassifyOutcome, ClassifyRequest, ClassifyRound, EvalSweep, EvalSweepValue, RaterTarget, RunClassifyFn, } from '#src/evalTypes.js';
 export { buildRaterClassifier, HARDLINE_NO_RATING_REASON, HARDLINE_REFUSAL_MARKER, NEGOTIATION_BOUND_MARKER, NO_RATING_CALL_MARKER, } from '#src/raterTarget.js';
 export type { RaterClassifierOptions } from '#src/raterTarget.js';
+export { RATER_PROMPT_NOTE_NAMES } from '#src/raterPromptArm.js';
+export type { RaterPromptArm, RaterPromptNoteName } from '#src/raterPromptArm.js';
 export { resolveReporters, availableReporterNames } from '#src/reporters/registry.js';
 export { driveReporters } from '#src/reporters/drive.js';
 export { createTextReporter } from '#src/reporters/textReporter.js';

package/dist/index.js

@@ -17,12 +17,20 @@ export { extractClassificationValue, buildConfusionMatrix, collectTags, readRaw,
 export { buildClassificationReport } from '#src/classificationReport.js';
 export { computeMetric, parseMetricPredicate, evaluatePredicate, evaluatePredicates, formatTally, } from '#src/metrics.js';
 export { renderClassificationReport, renderConfusionMatrix, renderMetric, } from '#src/classificationRender.js';
+// BATCH-32 — tool coverage: which of the agent's advertised tools a suite exercised.
+export { computeToolCoverage, aggregateToolCoverage, gradeRunToolCoverage, hasRunToolCoverageSpec, WAIVER_SHARE_WARN_THRESHOLD, } from '#src/toolCoverage.js';
+export { renderToolCoverage } from '#src/toolCoverageRender.js';
 export { buildBlindExport, diffRelabel, renderRelabelDiff } from '#src/blindExport.js';
-export { expandSweep, deepMerge, renderComparison, diffRuns, renderRunDiff, } from '#src/evalCompare.js';
+export { expandSweep, deepMerge, renderComparison, diffRuns, renderRunDiff, parseJudgeDriftFilter, DEFAULT_JUDGE_DRIFT_FILTER, DEFAULT_JUDGE_DRIFT_TOLERANCE, MAX_JUDGE_DRIFT_TOLERANCE, } from '#src/evalCompare.js';
 export { UNRECOGNIZED_LABEL, NO_EXPECTATION } from '#src/classificationTypes.js';
 // BATCH-25 Half B — the `rater` target: the one implementation of the `RunClassifyFn` seam, which
 // drives gth's own approvals rater over a corpus of commands.
 export { buildRaterClassifier, HARDLINE_NO_RATING_REASON, HARDLINE_REFUSAL_MARKER, NEGOTIATION_BOUND_MARKER, NO_RATING_CALL_MARKER, } from '#src/raterTarget.js';
+// [[BATCH-31]] — the prompt arm: the note-on / note-off A/B a rater suite can express. Only the
+// NAMES and the type are exported. `armRaterModel` deliberately is not — nothing outside this
+// package needs to build one, and the narrower the surface the smaller the chance of a caller
+// wiring an omission somewhere the argument in `raterPromptArm.ts` does not cover.
+export { RATER_PROMPT_NOTE_NAMES } from '#src/raterPromptArm.js';
 // BATCH-19 — the `gth eval` reporter facility (A1 seam). These are the public plugin contract an
 // out-of-core `@gaunt-sloth/eval-reporter-*` package implements, exported from the package root so a
 // reporter package can type its factory against ONE import.

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAC7C,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AACtD,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAClD,OAAO,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AACxE,OAAO,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAWlD,OAAO,EAAE,wBAAwB,EAAE,MAAM,eAAe,CAAC;AAEzD,gGAAgG;AAChG,OAAO,EAAE,cAAc,EAAE,MAAM,mBAAmB,CAAC;AACnD,OAAO,EAAE,sBAAsB,EAAE,MAAM,6BAA6B,CAAC;AACrE,OAAO,EACL,aAAa,EACb,kBAAkB,EAClB,iBAAiB,EACjB,6BAA6B,GAC9B,MAAM,eAAe,CAAC;AAEvB,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AAqBrD,OAAO,EAAE,2BAA2B,EAAE,MAAM,mBAAmB,CAAC;AAEhE,mGAAmG;AACnG,8FAA8F;AAC9F,OAAO,EACL,0BAA0B,EAC1B,oBAAoB,EACpB,WAAW,EACX,OAAO,GACR,MAAM,wBAAwB,CAAC;AAChC,OAAO,EAAE,yBAAyB,EAAE,MAAM,8BAA8B,CAAC;AACzE,OAAO,EACL,aAAa,EACb,oBAAoB,EACpB,iBAAiB,EACjB,kBAAkB,EAClB,WAAW,GACZ,MAAM,iBAAiB,CAAC;AACzB,OAAO,EACL,0BAA0B,EAC1B,qBAAqB,EACrB,YAAY,GACb,MAAM,8BAA8B,CAAC;AACtC,OAAO,EAAE,gBAAgB,EAAE,WAAW,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AAEvF,OAAO,EACL,WAAW,EACX,SAAS,EACT,gBAAgB,EAChB,QAAQ,EACR,aAAa,GACd,MAAM,qBAAqB,CAAC;AAE7B,OAAO,EAAE,kBAAkB,EAAE,cAAc,EAAE,MAAM,6BAA6B,CAAC;AAyBjF,kGAAkG;AAClG,8DAA8D;AAC9D,OAAO,EACL,oBAAoB,EACpB,yBAAyB,EACzB,uBAAuB,EACvB,wBAAwB,EACxB,qBAAqB,GACtB,MAAM,qBAAqB,CAAC;AAG7B,iGAAiG;AACjG,qGAAqG;AACrG,4DAA4D;AAC5D,OAAO,EAAE,gBAAgB,EAAE,sBAAsB,EAAE,MAAM,4BAA4B,CAAC;AACtF,OAAO,EAAE,cAAc,EAAE,MAAM,yBAAyB,CAAC;AACzD,OAAO,EAAE,kBAAkB,EAAE,MAAM,gCAAgC,CAAC;AAQpE,kGAAkG;AAClG,6CAA6C;AAC7C,OAAO,EAAE,WAAW,EAAE,MAAM,8BAA8B,CAAC;AAC3D,OAAO,EAAE,4BAA4B,EAAE,MAAM,eAAe,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAC7C,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AACtD,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAClD,OAAO,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AACxE,OAAO,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAWlD,OAAO,EAAE,wBAAwB,EAAE,MAAM,eAAe,CAAC;AAEzD,gGAAgG;AAChG,OAAO,EAAE,cAAc,EAAE,MAAM,mBAAmB,CAAC;AACnD,OAAO,EAAE,sBAAsB,EAAE,MAAM,6BAA6B,CAAC;AACrE,OAAO,EACL,aAAa,EACb,kBAAkB,EAClB,iBAAiB,EACjB,6BAA6B,GAC9B,MAAM,eAAe,CAAC;AAEvB,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AAqBrD,OAAO,EAAE,2BAA2B,EAAE,MAAM,mBAAmB,CAAC;AAEhE,mGAAmG;AACnG,8FAA8F;AAC9F,OAAO,EACL,0BAA0B,EAC1B,oBAAoB,EACpB,WAAW,EACX,OAAO,GACR,MAAM,wBAAwB,CAAC;AAChC,OAAO,EAAE,yBAAyB,EAAE,MAAM,8BAA8B,CAAC;AACzE,OAAO,EACL,aAAa,EACb,oBAAoB,EACpB,iBAAiB,EACjB,kBAAkB,EAClB,WAAW,GACZ,MAAM,iBAAiB,CAAC;AACzB,OAAO,EACL,0BAA0B,EAC1B,qBAAqB,EACrB,YAAY,GACb,MAAM,8BAA8B,CAAC;AACtC,qFAAqF;AACrF,OAAO,EACL,mBAAmB,EACnB,qBAAqB,EACrB,oBAAoB,EACpB,sBAAsB,EACtB,2BAA2B,GAC5B,MAAM,sBAAsB,CAAC;AAY9B,OAAO,EAAE,kBAAkB,EAAE,MAAM,4BAA4B,CAAC;AAEhE,OAAO,EAAE,gBAAgB,EAAE,WAAW,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AAEvF,OAAO,EACL,WAAW,EACX,SAAS,EACT,gBAAgB,EAChB,QAAQ,EACR,aAAa,EACb,qBAAqB,EACrB,0BAA0B,EAC1B,6BAA6B,EAC7B,yBAAyB,GAC1B,MAAM,qBAAqB,CAAC;AAU7B,OAAO,EAAE,kBAAkB,EAAE,cAAc,EAAE,MAAM,6BAA6B,CAAC;AAyBjF,kGAAkG;AAClG,8DAA8D;AAC9D,OAAO,EACL,oBAAoB,EACpB,yBAAyB,EACzB,uBAAuB,EACvB,wBAAwB,EACxB,qBAAqB,GACtB,MAAM,qBAAqB,CAAC;AAG7B,gGAAgG;AAChG,8FAA8F;AAC9F,8FAA8F;AAC9F,mFAAmF;AACnF,OAAO,EAAE,uBAAuB,EAAE,MAAM,wBAAwB,CAAC;AAGjE,iGAAiG;AACjG,qGAAqG;AACrG,4DAA4D;AAC5D,OAAO,EAAE,gBAAgB,EAAE,sBAAsB,EAAE,MAAM,4BAA4B,CAAC;AACtF,OAAO,EAAE,cAAc,EAAE,MAAM,yBAAyB,CAAC;AACzD,OAAO,EAAE,kBAAkB,EAAE,MAAM,gCAAgC,CAAC;AAQpE,kGAAkG;AAClG,6CAA6C;AAC7C,OAAO,EAAE,WAAW,EAAE,MAAM,8BAA8B,CAAC;AAC3D,OAAO,EAAE,4BAA4B,EAAE,MAAM,eAAe,CAAC"}

package/dist/judge.d.ts

@@ -1,6 +1,4 @@
 /**
- * @module judge
- *
  * BATCH-2 — LLM-as-judge grading for `gth eval`. Adapts the *mechanism* of EXT-10's shell-safety
  * judge (`packages/core/src/core/shell/judge.ts`): `model.withStructuredOutput(zodSchema)` for a
  * single non-agentic structured call, raced against a timeout. The failure policy differs on
@@ -18,6 +16,8 @@
  * `judgeShellCommand`'s own default. A separate `--judge <profile>` model is BATCH-2's own
  * "Not in scope" list (identity-matrix/pluggable-target work); grading with the SUT's own model
  * config is a known, real simplification for this first slice.
+ *
+ * @module
  */
 import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import * as z from 'zod';

package/dist/judge.js

@@ -1,6 +1,4 @@
 /**
- * @module judge
- *
  * BATCH-2 — LLM-as-judge grading for `gth eval`. Adapts the *mechanism* of EXT-10's shell-safety
  * judge (`packages/core/src/core/shell/judge.ts`): `model.withStructuredOutput(zodSchema)` for a
  * single non-agentic structured call, raced against a timeout. The failure policy differs on
@@ -18,8 +16,11 @@
  * `judgeShellCommand`'s own default. A separate `--judge <profile>` model is BATCH-2's own
  * "Not in scope" list (identity-matrix/pluggable-target work); grading with the SUT's own model
  * config is a known, real simplification for this first slice.
+ *
+ * @module
  */
 import { HumanMessage, SystemMessage } from '@langchain/core/messages';
+import { CALL_TIMED_OUT, withCallDeadline } from '@gaunt-sloth/core/runtime/abortableCall.js';
 import { structuredOutputBoundary } from '@gaunt-sloth/core/runtime/structuredOutput.js';
 import * as z from 'zod';
 /** Structured verdict the judge model must return — 0-10 `rate` + a `reason`, matching `review`'s
@@ -85,20 +86,19 @@ export async function judgeEvalCase(answer, rubric, model, options) {
         return { attempted: true, ok: false, error: 'No usable judge model configured.' };
     }
     const { system, user } = buildJudgeMessages(answer, rubric);
-    let timer;
     try {
         // EXT-88 — routed through the shared boundary like every other structured-output call. The
         // rubric verdict has no optional field today, so the boundary hands back this very schema by
         // identity; going through it is what stops a later optional field re-introducing the defect.
         const boundary = structuredOutputBoundary(EvalVerdictSchema);
         const structured = model.withStructuredOutput(boundary.wireSchema);
-        const judgePromise = structured.invoke([new SystemMessage(system), new HumanMessage(user)]);
-        const TIMEOUT = Symbol('eval-judge-timeout');
-        const timeoutPromise = new Promise((resolve) => {
-            timer = setTimeout(() => resolve(TIMEOUT), timeoutMs);
-        });
-        const raced = await Promise.race([judgePromise, timeoutPromise]);
-        if (raced === TIMEOUT) {
+        // [[EXT-179]] — the budget aborts the judge call rather than abandoning it. This matters most
+        // in a SWEEP: a suite grades many cases, and every stalled judge used to leave its own request
+        // in flight, so the leaked sockets accumulated across the run and kept the process alive after
+        // the report had been written. The timeout text is unchanged and the case still fails as a
+        // case — an aborted judgement is a judgement not obtained, never an auto-pass.
+        const raced = await withCallDeadline(timeoutMs, (signal) => structured.invoke([new SystemMessage(system), new HumanMessage(user)], { signal }));
+        if (raced === CALL_TIMED_OUT) {
             return { attempted: true, ok: false, error: `Judge timed out after ${timeoutMs}ms.` };
         }
         // withStructuredOutput already coerces to the schema, but re-validate defensively: a fake or
@@ -116,9 +116,5 @@ export async function judgeEvalCase(answer, rubric, model, options) {
             error: error instanceof Error ? error.message : String(error),
         };
     }
-    finally {
-        if (timer)
-            clearTimeout(timer);
-    }
 }
 //# sourceMappingURL=judge.js.map

package/dist/judge.js.map

@@ -1 +1 @@
-{"version":3,"file":"judge.js","sourceRoot":"","sources":["../src/judge.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,EAAE,wBAAwB,EAAE,MAAM,+CAA+C,CAAC;AACzF,OAAO,KAAK,CAAC,MAAM,KAAK,CAAC;AAIzB;;oGAEoG;AACpG,MAAM,CAAC,MAAM,iBAAiB,GAAG,CAAC,CAAC,MAAM,CAAC;IACxC,IAAI,EAAE,CAAC;SACJ,MAAM,EAAE;SACR,GAAG,CAAC,CAAC,CAAC;SACN,GAAG,CAAC,EAAE,CAAC;SACP,QAAQ,CAAC,yDAAyD,CAAC;IACtE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,2CAA2C,CAAC;CACzE,CAAC,CAAC;AAIH;;oGAEoG;AACpG,MAAM,CAAC,MAAM,6BAA6B,GAAG,MAAM,CAAC;AAEpD,MAAM,wBAAwB,GAAG;IAC/B,mCAAmC;IACnC,gGAAgG;IAChG,cAAc;IACd,EAAE;IACF,+FAA+F;IAC/F,gGAAgG;IAChG,mDAAmD;IACnD,6FAA6F;IAC7F,+BAA+B;CAChC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb;;;;GAIG;AACH,MAAM,UAAU,kBAAkB,CAChC,MAAc,EACd,MAAc;IAEd,MAAM,SAAS,GAAG;QAChB,SAAS;QACT,MAAM;QACN,EAAE;QACF,qBAAqB;QACrB,UAAU;QACV,MAAM,IAAI,gBAAgB;QAC1B,WAAW;KACZ,CAAC;IACF,OAAO,EAAE,MAAM,EAAE,wBAAwB,EAAE,IAAI,EAAE,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;AAC1E,CAAC;AAED;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,MAAc,EACd,MAAc,EACd,KAAgC,EAChC,OAAgC;IAEhC,MAAM,SAAS,GAAG,OAAO,EAAE,SAAS,IAAI,6BAA6B,CAAC;IAEtE,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,CAAC,oBAAoB,KAAK,UAAU,EAAE,CAAC;QAC/D,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,mCAAmC,EAAE,CAAC;IACpF,CAAC;IAED,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,GAAG,kBAAkB,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5D,IAAI,KAAgD,CAAC;IACrD,IAAI,CAAC;QACH,2FAA2F;QAC3F,6FAA6F;QAC7F,6FAA6F;QAC7F,MAAM,QAAQ,GAAG,wBAAwB,CAAC,iBAAiB,CAAC,CAAC;QAC7D,MAAM,UAAU,GAAG,KAAK,CAAC,oBAAoB,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;QACnE,MAAM,YAAY,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,IAAI,aAAa,CAAC,MAAM,CAAC,EAAE,IAAI,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAE5F,MAAM,OAAO,GAAG,MAAM,CAAC,oBAAoB,CAAC,CAAC;QAC7C,MAAM,cAAc,GAAG,IAAI,OAAO,CAAiB,CAAC,OAAO,EAAE,EAAE;YAC7D,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,SAAS,CAAC,CAAC;QACxD,CAAC,CAAC,CAAC;QAEH,MAAM,KAAK,GAAG,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC,YAAY,EAAE,cAAc,CAAC,CAAC,CAAC;QACjE,IAAI,KAAK,KAAK,OAAO,EAAE,CAAC;YACtB,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,yBAAyB,SAAS,KAAK,EAAE,CAAC;QACxF,CAAC;QAED,6FAA6F;QAC7F,0DAA0D;QAC1D,MAAM,MAAM,GAAG,QAAQ,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QACzC,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;YACpB,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,oCAAoC,EAAE,CAAC;QACrF,CAAC;QACD,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,CAAC,IAAI,EAAE,CAAC;IAC7D,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,OAAO;YACL,SAAS,EAAE,IAAI;YACf,EAAE,EAAE,KAAK;YACT,KAAK,EAAE,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC;SAC9D,CAAC;IACJ,CAAC;YAAS,CAAC;QACT,IAAI,KAAK;YAAE,YAAY,CAAC,KAAK,CAAC,CAAC;IACjC,CAAC;AACH,CAAC"}
+{"version":3,"file":"judge.js","sourceRoot":"","sources":["../src/judge.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,EAAE,cAAc,EAAE,gBAAgB,EAAE,MAAM,4CAA4C,CAAC;AAC9F,OAAO,EAAE,wBAAwB,EAAE,MAAM,+CAA+C,CAAC;AACzF,OAAO,KAAK,CAAC,MAAM,KAAK,CAAC;AAIzB;;oGAEoG;AACpG,MAAM,CAAC,MAAM,iBAAiB,GAAG,CAAC,CAAC,MAAM,CAAC;IACxC,IAAI,EAAE,CAAC;SACJ,MAAM,EAAE;SACR,GAAG,CAAC,CAAC,CAAC;SACN,GAAG,CAAC,EAAE,CAAC;SACP,QAAQ,CAAC,yDAAyD,CAAC;IACtE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,2CAA2C,CAAC;CACzE,CAAC,CAAC;AAIH;;oGAEoG;AACpG,MAAM,CAAC,MAAM,6BAA6B,GAAG,MAAM,CAAC;AAEpD,MAAM,wBAAwB,GAAG;IAC/B,mCAAmC;IACnC,gGAAgG;IAChG,cAAc;IACd,EAAE;IACF,+FAA+F;IAC/F,gGAAgG;IAChG,mDAAmD;IACnD,6FAA6F;IAC7F,+BAA+B;CAChC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb;;;;GAIG;AACH,MAAM,UAAU,kBAAkB,CAChC,MAAc,EACd,MAAc;IAEd,MAAM,SAAS,GAAG;QAChB,SAAS;QACT,MAAM;QACN,EAAE;QACF,qBAAqB;QACrB,UAAU;QACV,MAAM,IAAI,gBAAgB;QAC1B,WAAW;KACZ,CAAC;IACF,OAAO,EAAE,MAAM,EAAE,wBAAwB,EAAE,IAAI,EAAE,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;AAC1E,CAAC;AAED;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,MAAc,EACd,MAAc,EACd,KAAgC,EAChC,OAAgC;IAEhC,MAAM,SAAS,GAAG,OAAO,EAAE,SAAS,IAAI,6BAA6B,CAAC;IAEtE,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,CAAC,oBAAoB,KAAK,UAAU,EAAE,CAAC;QAC/D,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,mCAAmC,EAAE,CAAC;IACpF,CAAC;IAED,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,GAAG,kBAAkB,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5D,IAAI,CAAC;QACH,2FAA2F;QAC3F,6FAA6F;QAC7F,6FAA6F;QAC7F,MAAM,QAAQ,GAAG,wBAAwB,CAAC,iBAAiB,CAAC,CAAC;QAC7D,MAAM,UAAU,GAAG,KAAK,CAAC,oBAAoB,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;QAEnE,8FAA8F;QAC9F,+FAA+F;QAC/F,+FAA+F;QAC/F,2FAA2F;QAC3F,+EAA+E;QAC/E,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAAC,SAAS,EAAE,CAAC,MAAM,EAAE,EAAE,CACzD,UAAU,CAAC,MAAM,CAAC,CAAC,IAAI,aAAa,CAAC,MAAM,CAAC,EAAE,IAAI,YAAY,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,MAAM,EAAE,CAAC,CACnF,CAAC;QACF,IAAI,KAAK,KAAK,cAAc,EAAE,CAAC;YAC7B,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,yBAAyB,SAAS,KAAK,EAAE,CAAC;QACxF,CAAC;QAED,6FAA6F;QAC7F,0DAA0D;QAC1D,MAAM,MAAM,GAAG,QAAQ,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QACzC,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;YACpB,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,oCAAoC,EAAE,CAAC;QACrF,CAAC;QACD,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,CAAC,IAAI,EAAE,CAAC;IAC7D,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,OAAO;YACL,SAAS,EAAE,IAAI;YACf,EAAE,EAAE,KAAK;YACT,KAAK,EAAE,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC;SAC9D,CAAC;IACJ,CAAC;AACH,CAAC"}

package/dist/output.d.ts

@@ -4,7 +4,7 @@ import type { BatchSummary, CellResult } from '#src/types.js';
  * (pass/fail counts + a per-cell one-liner — a lightweight flake report) into `outputDir`.
  * Creates `outputDir` (and any missing parents) if it doesn't exist.
  *
- * Pure I/O, deliberately separate from {@link runBatchMatrix}: the runner never touches the
+ * Pure I/O, deliberately separate from {@link @gaunt-sloth/batch!"BatchRunner.js".runBatchMatrix | runBatchMatrix}: the runner never touches the
  * filesystem, so unit tests can exercise matrix/concurrency/retry logic without a tmp dir, and this
  * function can be tested in isolation with a fixed set of results.
  *

package/dist/output.js

@@ -6,7 +6,7 @@ import { buildBatchSummary } from '#src/BatchRunner.js';
  * (pass/fail counts + a per-cell one-liner — a lightweight flake report) into `outputDir`.
  * Creates `outputDir` (and any missing parents) if it doesn't exist.
  *
- * Pure I/O, deliberately separate from {@link runBatchMatrix}: the runner never touches the
+ * Pure I/O, deliberately separate from {@link @gaunt-sloth/batch!"BatchRunner.js".runBatchMatrix | runBatchMatrix}: the runner never touches the
  * filesystem, so unit tests can exercise matrix/concurrency/retry logic without a tmp dir, and this
  * function can be tested in isolation with a fixed set of results.
  *

package/dist/parseOver.d.ts

@@ -4,7 +4,7 @@ import type { MatrixRow } from '#src/types.js';
  * `.jsonl`/`.ndjson` → one JSON object per line; anything else (including `.csv`) → CSV.
  *
  * This is content binding only (BATCH-1 scope): every row becomes an object of string fields that
- * {@link bindCellContent} interpolates into the script. A glob-of-binary-files path binding is out
+ * {@link @gaunt-sloth/batch!"interpolate.js".bindCellContent | bindCellContent} interpolates into the script. A glob-of-binary-files path binding is out
  * of scope for this task.
  *
  * Throws a descriptive `Error` on malformed input — the harness-level failure the CLI surface doc

package/dist/parseOver.js

@@ -3,7 +3,7 @@
  * `.jsonl`/`.ndjson` → one JSON object per line; anything else (including `.csv`) → CSV.
  *
  * This is content binding only (BATCH-1 scope): every row becomes an object of string fields that
- * {@link bindCellContent} interpolates into the script. A glob-of-binary-files path binding is out
+ * {@link @gaunt-sloth/batch!"interpolate.js".bindCellContent | bindCellContent} interpolates into the script. A glob-of-binary-files path binding is out
  * of scope for this task.
  *
  * Throws a descriptive `Error` on malformed input — the harness-level failure the CLI surface doc

package/dist/pipelineCli.d.ts

@@ -1,6 +1,4 @@
 /**
- * @module pipelineCli
- *
  * BATCH-9 — the standalone `gth-batch` pipeline runner. A thin entry point that runs the BATCH-1
  * matrix runtime ({@link buildMatrix} + {@link runBatchMatrix}) directly from a shell pipeline,
  * without pulling in the whole `gaunt-sloth` app. It takes a prompt-executable script + `--over`
@@ -16,10 +14,12 @@
  * the pipeline shell already handles) and output is JSONL on stdout (not a directory of files).
  *
  * stdout discipline: the run itself is noisy (the runtime's `display()`/`ProgressIndicator`/token
- * streaming all target `process.stdout`). The bin entry ({@link file://./bin.ts}) redirects
+ * streaming all target `process.stdout`). The bin entry ({@link @gaunt-sloth/batch!"bin.js" | ./bin.ts}) redirects
  * `process.stdout.write` to stderr for the duration and this module writes the machine JSONL
  * straight to fd 1 (`fs.writeSync`), so stdout stays a clean data channel — the same "protocol
  * channel" discipline `packages/app/cli.js` uses for ACP.
+ *
+ * @module
  */
 import type { CommandLineConfigOverrides, GthConfig } from '@gaunt-sloth/core/config.js';
 import type { BatchSummary, MatrixRow, RunCellFn } from '#src/types.js';

package/dist/pipelineCli.js

@@ -1,6 +1,4 @@
 /**
- * @module pipelineCli
- *
  * BATCH-9 — the standalone `gth-batch` pipeline runner. A thin entry point that runs the BATCH-1
  * matrix runtime ({@link buildMatrix} + {@link runBatchMatrix}) directly from a shell pipeline,
  * without pulling in the whole `gaunt-sloth` app. It takes a prompt-executable script + `--over`
@@ -16,10 +14,12 @@
  * the pipeline shell already handles) and output is JSONL on stdout (not a directory of files).
  *
  * stdout discipline: the run itself is noisy (the runtime's `display()`/`ProgressIndicator`/token
- * streaming all target `process.stdout`). The bin entry ({@link file://./bin.ts}) redirects
+ * streaming all target `process.stdout`). The bin entry ({@link @gaunt-sloth/batch!"bin.js" | ./bin.ts}) redirects
  * `process.stdout.write` to stderr for the duration and this module writes the machine JSONL
  * straight to fd 1 (`fs.writeSync`), so stdout stays a clean data channel — the same "protocol
  * channel" discipline `packages/app/cli.js` uses for ACP.
+ *
+ * @module
  */
 import { writeSync } from 'node:fs';
 import { readFileSync } from 'node:fs';