wicked-crew 0.7.25 → 0.7.27

package/dist/api/endpoint-manifest-live.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"endpoint-manifest-live.d.ts","sourceRoot":"","sources":["../../src/api/endpoint-manifest-live.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAQH,OAAO,EAAyB,KAAK,gBAAgB,EAAE,MAAM,wBAAwB,CAAC;AAEtF,wBAAsB,2BAA2B,IAAI,OAAO,CAAC,gBAAgB,CAAC,CAkC7E"}
+{"version":3,"file":"endpoint-manifest-live.d.ts","sourceRoot":"","sources":["../../src/api/endpoint-manifest-live.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAQH,OAAO,EAAyB,KAAK,gBAAgB,EAAE,MAAM,wBAAwB,CAAC;AAEtF,wBAAsB,2BAA2B,IAAI,OAAO,CAAC,gBAAgB,CAAC,CAuC7E"}

package/dist/api/endpoint-manifest-live.js

@@ -40,6 +40,7 @@ export async function collectLiveEndpointManifest() {
         projectsSupported: () => false,
         getSettings: async () => ({}),
         // The daemon's single CoreEvent fan-out subscribes at boot; no event ever arrives here.
+        onLaunch: () => () => undefined, // the launch hook createServer registers (skills keystone, codex round 4)
         onEvent: () => () => { },
     };
     try {
@@ -49,6 +50,10 @@ export async function collectLiveEndpointManifest() {
             projectEvents: { disabled: true },
             interactiveWsRelay: { disabled: true },
             stallWatchdog: { enabled: false },
+            // No skills seam: the manifest documents the routes, and an armed seam would seed the
+            // generating machine's REAL skills root from its installed plugin. The routes register
+            // regardless (they answer 503 unconfigured).
+            skills: { disabled: true },
             // Nonexistent on purpose — headless boot; see the module header.
             studioRoot: join(scratch, 'no-studio-bundle'),
         });

package/dist/api/endpoint-manifest-live.js.map

@@ -1 +1 @@
-{"version":3,"file":"endpoint-manifest-live.js","sourceRoot":"","sources":["../../src/api/endpoint-manifest-live.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AAC9C,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAGjC,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAC3C,OAAO,EAAE,qBAAqB,EAAyB,MAAM,wBAAwB,CAAC;AAEtF,MAAM,CAAC,KAAK,UAAU,2BAA2B;IAC/C,4FAA4F;IAC5F,kFAAkF;IAClF,MAAM,aAAa,GAAG,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;IAC/C,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,GAAG,QAAQ,CAAC;IACpC,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,CAAC,MAAM,EAAE,EAAE,yBAAyB,CAAC,CAAC,CAAC;IACvE,MAAM,OAAO,GAAG;QACd,IAAI,EAAE,IAAI;QACV,iBAAiB,EAAE,GAAG,EAAE,CAAC,KAAK;QAC9B,WAAW,EAAE,KAAK,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC;QAC7B,wFAAwF;QACxF,OAAO,EAAE,GAAG,EAAE,CAAC,GAAG,EAAE,GAAE,CAAC;KACE,CAAC;IAC5B,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,MAAM,YAAY,CAAC,OAAO,EAAE;YACtC,IAAI,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE;YACrB,SAAS,EAAE,IAAI,CAAC,OAAO,EAAE,WAAW,CAAC;YACrC,aAAa,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;YACjC,kBAAkB,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;YACtC,aAAa,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE;YACjC,iEAAiE;YACjE,UAAU,EAAE,IAAI,CAAC,OAAO,EAAE,kBAAkB,CAAC;SAC9C,CAAC,CAAC;QACH,IAAI,CAAC;YACH,MAAM,GAAG,CAAC,KAAK,EAAE,CAAC;YAClB,OAAO,qBAAqB,CAAC,GAAG,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;QAC3D,CAAC;gBAAS,CAAC;YACT,MAAM,GAAG,CAAC,KAAK,EAAE,CAAC;QACpB,CAAC;IACH,CAAC;YAAS,CAAC;QACT,IAAI,aAAa,KAAK,SAAS;YAAE,OAAO,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;;YAC5D,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,GAAG,aAAa,CAAC;QAC9C,MAAM,CAAC,OAAO,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;IACpD,CAAC;AACH,CAAC"}
+{"version":3,"file":"endpoint-manifest-live.js","sourceRoot":"","sources":["../../src/api/endpoint-manifest-live.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AAC9C,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAGjC,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAC3C,OAAO,EAAE,qBAAqB,EAAyB,MAAM,wBAAwB,CAAC;AAEtF,MAAM,CAAC,KAAK,UAAU,2BAA2B;IAC/C,4FAA4F;IAC5F,kFAAkF;IAClF,MAAM,aAAa,GAAG,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;IAC/C,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,GAAG,QAAQ,CAAC;IACpC,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,CAAC,MAAM,EAAE,EAAE,yBAAyB,CAAC,CAAC,CAAC;IACvE,MAAM,OAAO,GAAG;QACd,IAAI,EAAE,IAAI;QACV,iBAAiB,EAAE,GAAG,EAAE,CAAC,KAAK;QAC9B,WAAW,EAAE,KAAK,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC;QAC7B,wFAAwF;QACxF,QAAQ,EAAE,GAAiB,EAAE,CAAC,GAAG,EAAE,CAAC,SAAS,EAAE,0EAA0E;QACzH,OAAO,EAAE,GAAG,EAAE,CAAC,GAAG,EAAE,GAAE,CAAC;KACE,CAAC;IAC5B,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,MAAM,YAAY,CAAC,OAAO,EAAE;YACtC,IAAI,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE;YACrB,SAAS,EAAE,IAAI,CAAC,OAAO,EAAE,WAAW,CAAC;YACrC,aAAa,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;YACjC,kBAAkB,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;YACtC,aAAa,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE;YACjC,sFAAsF;YACtF,uFAAuF;YACvF,6CAA6C;YAC7C,MAAM,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;YAC1B,iEAAiE;YACjE,UAAU,EAAE,IAAI,CAAC,OAAO,EAAE,kBAAkB,CAAC;SAC9C,CAAC,CAAC;QACH,IAAI,CAAC;YACH,MAAM,GAAG,CAAC,KAAK,EAAE,CAAC;YAClB,OAAO,qBAAqB,CAAC,GAAG,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;QAC3D,CAAC;gBAAS,CAAC;YACT,MAAM,GAAG,CAAC,KAAK,EAAE,CAAC;QACpB,CAAC;IACH,CAAC;YAAS,CAAC;QACT,IAAI,aAAa,KAAK,SAAS;YAAE,OAAO,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;;YAC5D,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC,GAAG,aAAa,CAAC;QAC9C,MAAM,CAAC,OAAO,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;IACpD,CAAC;AACH,CAAC"}

package/dist/api/eval-compare.d.ts

@@ -0,0 +1,306 @@
+/**
+ * Cross-run eval comparison (docs/testing/evals-test-plan.md S17) — the OFFLINE, deterministic diff
+ * of two recorded {@link EvalRunDetail}s over the same corpus: which per-sample verdicts flipped,
+ * whether each flip is the kind a tightened store is allowed to produce (`permitted`) or the kind a
+ * release has to explain (`flagged`), whether the two stored summaries reconcile to the per-sample
+ * accounting, and which rules gained or lost exercise (core #394 `rule_coverage`).
+ *
+ * Pure over the two records: no store, no engine, no clock, no I/O — two persisted drilldowns in,
+ * one report out, byte-stable for the same inputs (every list is codepoint-sorted by id).
+ *
+ * # What "comparable" means here
+ *
+ * Plan §3 keys a comparison by five identities. The records carry four of them today (`corpus`,
+ * `rule_store`, `type_filter`, `degraded`; the rule snapshot and the engine build are §5 follow-ups),
+ * so `comparable` is derived from CONTENT, never from a name alone: the same corpus name AND the same
+ * type filter AND the same sample ids AND, per id, the same `kind` AND the same sample PAYLOAD
+ * identity — the `sample.payload_hash` a producer stamped on each result row (`eval-sample.js`
+ * `samplePayloadHash`: sha256 over the canonical JSON of id, description, kind, steering_type,
+ * signals). The engine echoes only id/description/kind/steering_type per row, never the input
+ * `signals`, so a row WITHOUT a payload hash cannot be proven to be the same action as its
+ * namesake in the other run: such a pair is `comparable: false` with `comparable_reason` starting
+ * `unverified: no sample identity` — unverified, which is not the same as "different". A sample
+ * present on one side only, one whose kind changed, or one whose payload hash changed means the
+ * corpus changed under the same name — reported, and the pair is not comparable release over
+ * release. Two runs under DIFFERENT type filters judged different slices of the corpus over
+ * different coverage denominators — not comparable either (`differing-type-filter`, see below).
+ * `comparable_reason` is `null` exactly when `comparable` is true.
+ *
+ * # Flip classification (the S17 rule table)
+ *
+ *   bad   gap → caught                permitted   a rule tightened: a bad behavior is now caught
+ *   bad   caught → gap                flagged     regression: a bad behavior is no longer caught
+ *   good  false_positive → caught     permitted   a good sample is no longer denied
+ *   good  caught → false_positive     flagged     a good sample is newly denied
+ *   anything else                     flagged     not a verdict pair that kind can take (evals.rs
+ *                                                 `evaluate_sample`: a bad sample is caught|gap, a
+ *                                                 good sample is caught|false_positive)
+ *
+ * `good` is a sample KIND, not a verdict — the draft plan's `good → false_positive` spelling was
+ * corrected in revision 3; this module is the executable form of that correction.
+ *
+ * # Coverage transitions vs. rule-set changes — what a record CANNOT tell (codex rounds 6 and 8)
+ *
+ * The wire carries `rule_coverage.exercised` as a COUNT and `unexercised` as a LIST of ids
+ * (api-types `GovernanceEvalRuleCoverage`; wicked-core `crates/wicked-governance/src/evals.rs` on
+ * branch `feat/evals-effect-and-coverage` (core #394/#395, PR #398) at `a87e461` — `RuleCoverage`
+ * lines 323-328: `exercised: usize`, `unexercised: Vec<UnexercisedRule>`, `recall_only: usize`,
+ * `per_type: BTreeMap<String, TypeCoverage>`). The partition is over the ELIGIBLE rules — the
+ * active, effect-bearing rules of the slice (`decide_lane_rules` lines 816-843: `effect.is_some()
+ * && !retired && in_slice`); `recall_only` counts the effect-less active rules OUTSIDE it (docs
+ * lines 317-319, computed at 875-878). A rule is EXERCISED when it appeared in ANY evaluated claim's
+ * `policy_ids`, whatever its effect (`RuleCoverage` docs lines 312-314; `rule_coverage()` lines
+ * 849-880 partitions the eligible rules by `triggered`, the union `run_evals` collects of every
+ * claim's `policy_ids`). A result row's `fired`, however, is the BLOCKING subset only:
+ * `evaluate_sample` lines 906-916 keep the ids whose effect is `Deny`. Hence, for an UNFILTERED run,
+ *
+ *   fired(run) ⊆ exercised(run), and `exercised − |fired|` rules were exercised by a non-blocking
+ *   (`warn`) effect ALONE — counted, but NEVER named anywhere in the record.
+ *
+ * NO field of the wire lists a run's rule set: `exercised` and `recall_only` are counts,
+ * `unexercised` names only the rules nothing exercised, `fired` only the blocking firings. The rule
+ * identities a record enumerates are therefore exactly `unexercised ∪ fired`, and an id one record
+ * enumerates and the other does not is NOT thereby absent from the other's store: it may be one of
+ * that side's unnamed warn-exercised rules, an effect-less (recall-only) or retired rule outside the
+ * eligible partition, under a type filter a rule of another type outside the denominator — or a rule
+ * the store really gained or lost. Round 6 replaced a reconstruction of "the inventory" as
+ * `unexercised ∪ fired` with a completeness inference: `exercised === |fired|` ⇒ every exercised rule
+ * is named ⇒ the silent side's absence was asserted as `added_rules`/`removed_rules`. Round 8 deleted
+ * that inference too — under a type filter a fired id is typed into the slice only by the OTHER run's
+ * `unexercised` row, and a rule's type in run A does not establish its type in run B (codex's
+ * reproduction: A lists R and Q unexercised under `development`; B moves R to `security`, fires R
+ * blocking and exercises Q by `warn`; the cross-run intersection named R, declared B complete and
+ * reported Q — still present and exercised — as `removed_rules`). The delta asserts ONLY what the
+ * records state:
+ *
+ *   gained         unexercised in A (listed) AND blocking-fired in B (listed): a sample now exercises
+ *                  the rule — a statement about the RULE, both ends named, certain
+ *   lost           blocking-fired in A AND unexercised in B (listed): certain
+ *   unidentified   PER RECORD, how many exercised rules of its denominator it does not name:
+ *                  unfiltered `exercised − |fired|`; under a type filter the whole `exercised` (a
+ *                  fired id carries no steering_type, so a record types none of its own firings into
+ *                  the slice) — never reduced by what the OTHER run lists
+ *   added_rules /  asserted only from an explicit rule inventory on BOTH records — the wire carries
+ *   removed_rules  none, so both are EMPTY for every daemon-recorded run and `inventory` is `partial`
+ *   transitions_withheld
+ *                  every id enumerated by one record and not the other, one reason per id saying what
+ *                  it may be besides a rule-set change — WITHHELD, never guessed
+ *
+ * `exercised_delta` is always the difference of the two counts. Reconciliation errors on coverage are
+ * the record contradicting ITSELF (or being malformed — below); two valid reports always `reconcile`.
+ *
+ * # Type filters (codex round 7) — the denominator is the SLICE; the firings are not
+ *
+ * A run recorded under `type_filter: T` was produced by `rules eval --type T`. In the engine that
+ * filter slices the SAMPLES (`run_evals` lines 974-977: only samples of type T are judged) and the
+ * COVERAGE DENOMINATOR (`decide_lane_rules` lines 816-844, `in_slice` at 820: only type-T rules are
+ * eligible; `rule_coverage` 849-880 partitions exactly those; `recall_only` 875-878 is sliced too)
+ * but NOT the gate: `evaluate_sample` runs `select_any` over every active rule whatever its type
+ * (line 901) and a row's `fired` keeps every blocking id it produced (906-916). So under a filter a
+ * row may fire a rule OUTSIDE the denominator — codex's reproduction: a development-filtered run
+ * with `fired: ["SECURITY-DENY"]` and `exercised: 0` is a VALID record (the security rule is not a
+ * development rule), and the unfiltered invariant `exercised ≥ |fired|` does not hold. Hence, per
+ * record ({@link EvalCoverageReconciliation}, reported per side in `coverage_reconciliation`):
+ *
+ *   type_filter null          `'rows'` — every active rule is in the denominator, so the rows'
+ *                             blocking-fired ids are exercised rules: `exercised ≥ |fired|`, and
+ *                             the exercised set is complete iff `exercised === |fired|`.
+ *   type_filter T, per_type   `'per_type'` — reconciled against the engine's OWN row for the slice
+ *   present                   (`rule_coverage.per_type[T]`, `TypeCoverage` lines 300-303):
+ *                             `per_type[T].exercised === exercised`; the rows' fired ids carry no
+ *                             steering type and are NOT a denominator check.
+ *   type_filter T, no         `'n/a (engine reports no per-type coverage)'` — the record carries
+ *   per_type                  nothing to check its slice against; nothing fired-based is asserted.
+ *
+ * Whatever the filter: `per_type`, when present, must sum to `exercised` and agree per type with the
+ * listed `unexercised` rows; under a filter every listed row is of the filter's type; no id is
+ * listed twice; and no listed-unexercised id fired — that row TYPES the rule into the slice, and a
+ * blocking firing is a firing, which makes an eligible rule exercised (862-873) — so this last check
+ * is sound under a filter too (the fix brief's floor was "unfiltered only"; the engine's definition
+ * makes it general).
+ *
+ * Two runs under DIFFERENT filters get NO `rule_coverage_delta` (an inventory diff across
+ * denominators would report the other slice's rules as changes) and are not comparable. Under the
+ * SAME filter T the delta is computed exactly as unfiltered — `gained`/`lost` from ids listed on both
+ * ends (the unexercised row carries `steering_type: T`, the firing is a firing), every one-sided id
+ * withheld — except that `unidentified` is the whole `exercised` count (see above) and each withheld
+ * reason adds that the id may be a rule of another type outside the denominator.
+ *
+ * # Malformed persisted coverage (codex round 8)
+ *
+ * The daemon persists a run's `rule_coverage` verbatim and validates none of it (the script's
+ * `verifyEngineReport` gates only its own reports). A recorded `rule_coverage` that is not the wire
+ * shape — `null`; `unexercised` not an array; `per_type: null`, a row that is not `{ exercised,
+ * unexercised }` of non-negative integers, a key that is no steering type … — cannot be reconciled:
+ * that side is `coverage_reconciliation: 'unverified (malformed rule_coverage)'`
+ * ({@link MALFORMED_RULE_COVERAGE}), a reconciliation error names the run and the defect, no
+ * `rule_coverage_delta` is computed, and the comparison never throws over persisted data.
+ *
+ * # Malformed persisted result rows (Copilot on #475)
+ *
+ * The same store validates a detail's `results` only as "an array" (`EvalRunStore.get()`), so a row
+ * that is not the wire shape (`GovernanceEvalResult` — no `sample`, `sample.id` not a non-empty
+ * string, a `kind` or `verdict` outside its union, `fired` not an array of strings, a row that is no
+ * object at all) can reach the comparison from a corrupted or hand-edited file. Every row is checked
+ * ({@link resultRowProblem}) BEFORE it is indexed, tallied or iterated: a malformed one is ONE
+ * reconciliation error naming the run, the index and the defect, and is EXCLUDED from everything —
+ * the id index, the identity counts, the tally, the coverage `fired` set. The pair is then not
+ * comparable ({@link MALFORMED_RESULT_ROWS}), and that side's stored summary is NOT checked against a
+ * results list the comparison could not fully read (the shortfall is the excluded rows, not a summary
+ * defect — reporting it as one would misattribute it). Never a throw.
+ */
+import type { EvalRunDetail, GovernanceEvalResult, GovernanceEvalSummary, SteeringType } from '../core/types.js';
+export type EvalVerdict = GovernanceEvalResult['verdict'];
+export type EvalSampleKind = GovernanceEvalResult['sample']['kind'];
+export type FlipClass = 'permitted' | 'flagged';
+/** The exact prefix of `comparable_reason` when a side carries result rows without a payload hash. */
+export declare const UNVERIFIED_NO_SAMPLE_IDENTITY = "unverified: no sample identity";
+/** The exact prefix of `comparable_reason` when the two runs were produced under different type filters. */
+export declare const DIFFERING_TYPE_FILTER = "differing-type-filter";
+/** One identity dimension, side by side. */
+export interface IdentityPair<T> {
+    a: T;
+    b: T;
+    same: boolean;
+}
+/** One sample whose verdict differs between the two runs. */
+export interface EvalVerdictFlip {
+    sample_id: string;
+    /** The sample's kind in run B (equal to run A's unless the id is also in `kind_changed`). */
+    kind: EvalSampleKind;
+    from: EvalVerdict;
+    to: EvalVerdict;
+    classification: FlipClass;
+    reason: string;
+}
+/** Whether the two records LIST their rule sets. `complete` requires an explicit rule inventory on
+ *  BOTH records; no field of `GovernanceEvalRuleCoverage` (api-types 0.27.0) carries one — `exercised`
+ *  and `recall_only` are counts, `unexercised` names only the unexercised rules, a row's `fired` only
+ *  the blocking firings — so every comparison of daemon-recorded runs is `partial` and the rule-set
+ *  transitions are withheld by name. See the module doc. */
+export type EvalRuleInventory = 'complete' | 'partial';
+/** The exact `coverage_reconciliation` value of a side whose persisted `rule_coverage` is not the wire
+ *  shape — unverified, not reconciled, no delta computed over it (module doc, "Malformed persisted
+ *  coverage"). */
+export declare const MALFORMED_RULE_COVERAGE = "unverified (malformed rule_coverage)";
+/** The `comparable_reason` prefix of a pair one side of which persisted result rows that are not the
+ *  wire shape (api-types `GovernanceEvalResult`, module doc "Malformed persisted result rows"): each
+ *  is named in `reconciliation_errors` and EXCLUDED, never thrown over — and a comparison over a
+ *  record it could not fully read asserts nothing (Copilot on #475). */
+export declare const MALFORMED_RESULT_ROWS = "unverified: malformed result row(s)";
+/**
+ * How one record's `rule_coverage` was reconciled — which denominator its numbers were checked
+ * against (module doc, "Type filters"):
+ *   - `'rows'` — `type_filter` null: every active rule is in the denominator, so the rows' distinct
+ *     blocking-fired ids are exercised rules and `exercised ≥ |fired|` is checked;
+ *   - `'per_type'` — `type_filter` set and `per_type` present: checked against the engine's own row
+ *     for that type, `per_type[<filter>]`; the rows' fired ids may belong to other types;
+ *   - `'n/a (engine reports no per-type coverage)'` — `type_filter` set and no `per_type`: the
+ *     denominator is the filter's slice and the record carries no number to check it against;
+ *   - {@link MALFORMED_RULE_COVERAGE} — the persisted value is not the wire shape: nothing was
+ *     checked, a reconciliation error names the defect, no delta is computed over this side.
+ */
+export type EvalCoverageReconciliation = 'rows' | 'per_type' | 'n/a (engine reports no per-type coverage)' | typeof MALFORMED_RULE_COVERAGE;
+/** Which rules gained or lost exercise between the runs (present only when BOTH carry WELL-FORMED
+ *  coverage AND both were produced under the same `type_filter`). */
+export interface EvalRuleCoverageDelta {
+    /** `b.rule_coverage.exercised − a.rule_coverage.exercised` — the counts, always. */
+    exercised_delta: number;
+    /** `partial` for every comparison of daemon-recorded runs: the wire carries no rule inventory, so
+     *  `added_rules`/`removed_rules` are never asserted and every one-sided id is in
+     *  `transitions_withheld`. `complete` is reserved for two records that LIST their rule sets. */
+    inventory: EvalRuleInventory;
+    /** Per side, how many exercised rules of its denominator the record counts but does not name —
+     *  a fact about THAT record alone, never reduced by what the other run lists. Unfiltered:
+     *  `rule_coverage.exercised − |distinct blocking-fired ids|` (the rules exercised by a `warn`
+     *  effect alone). Under a type filter: `rule_coverage.exercised` — a fired id carries no
+     *  steering_type, so the record types none of its own firings into the slice. */
+    unidentified: {
+        a: number;
+        b: number;
+    };
+    /** Unexercised in A (listed) AND blocking-fired in B (listed) — a sample now exercises the rule.
+     *  A statement about the RULE, certain on both ends whatever the filter (under one, the id is typed
+     *  into the slice by A's row alone — B's `unidentified` does not subtract it). */
+    gained: string[];
+    /** Blocking-fired in A AND unexercised in B (listed) — a rule nothing exercises any more. Certain. */
+    lost: string[];
+    /** Enumerated by B and not by A AND absent from A's explicit rule inventory — the wire carries no
+     *  inventory, so this is EMPTY for every daemon-recorded run (the candidates are in
+     *  `transitions_withheld`). */
+    added_rules: string[];
+    /** Enumerated by A and not by B AND absent from B's explicit rule inventory — EMPTY for the same
+     *  reason. */
+    removed_rules: string[];
+    /** Every id enumerated (unexercised or blocking-fired) by one record and not the other, one reason
+     *  per id, codepoint-sorted — what the id may be besides a rule-set change: an exercised rule the
+     *  silent side never names (a `warn`-only firing; under a type filter any of its firings), an
+     *  effect-less (recall-only) or retired rule outside the eligible partition, under a type filter a
+     *  rule of another type outside the denominator. WITHHELD, never guessed. */
+    transitions_withheld: string[];
+}
+export interface EvalRunComparison {
+    /** The two run ids, in the order given (A = the baseline, B = the candidate). */
+    a: string;
+    b: string;
+    identity: {
+        corpus: IdentityPair<string | null>;
+        rule_store: IdentityPair<string>;
+        type_filter: IdentityPair<SteeringType | null>;
+        degraded: IdentityPair<'facet-only' | null>;
+    };
+    /** Same corpus name, same type filter, same sample ids, same kinds, same payload identities on
+     *  every row — the same actions were judged over the same slice, provably. */
+    comparable: boolean;
+    /** Why `comparable` is false (`;`-joined when several apply; starts with
+     *  {@link UNVERIFIED_NO_SAMPLE_IDENTITY} when a side carries rows without a payload hash, contains
+     *  {@link DIFFERING_TYPE_FILTER} when the filters differ), or `null` when it is true. */
+    comparable_reason: string | null;
+    /** Every per-sample verdict change, codepoint-sorted by sample id, each classified. */
+    flips: EvalVerdictFlip[];
+    /** Shared samples with the same kind AND the same verdict on both sides, minus those whose payload
+     *  identity is VERIFIED to differ (`payload_changed`). A row unverified on either side (no
+     *  well-formed `payload_hash`) still counts when kind and verdict agree — `unchanged` is a verdict
+     *  statement, not proof of payload identity; `comparable` / `unverified_rows` carry that claim. */
+    unchanged: number;
+    only_in_a: string[];
+    only_in_b: string[];
+    /** Shared ids whose `kind` differs — the corpus changed under this name (always flagged). */
+    kind_changed: string[];
+    /** Shared ids whose `payload_hash` differs on the two sides (both present) — the description,
+     *  steering type or signals changed under the same id and kind: the corpus changed under this
+     *  name (a flip on such an id is always flagged). */
+    payload_changed: string[];
+    /** Result rows WITHOUT a `sample.payload_hash`, per side — the rows whose identity is unverified. */
+    unverified_rows: {
+        a: number;
+        b: number;
+    };
+    /** `b.summary − a.summary`, field by field. */
+    summary_delta: GovernanceEvalSummary;
+    /** Each run's summary equals its own results' tally AND the delta equals the flip/add/remove
+     *  accounting AND each run's coverage is well-formed and does not contradict itself (see
+     *  `coverage_reconciliation` for what it was checked against) — a stored record that disagrees
+     *  with ITSELF, or carries a `rule_coverage` that is not the wire shape, is a defect, not a flip.
+     *  Two valid records always reconcile; an exercised count ABOVE the fired ids is valid (warn-only
+     *  exercise), and so is a fired id outside a filtered run's denominator. */
+    reconciles: boolean;
+    reconciliation_errors: string[];
+    /** Per side, how the record's `rule_coverage` was reconciled ({@link EvalCoverageReconciliation};
+     *  {@link MALFORMED_RULE_COVERAGE} when the persisted value is not the wire shape), or `null` when
+     *  that side carries no coverage (a pre-#394 engine). */
+    coverage_reconciliation: {
+        a: EvalCoverageReconciliation | null;
+        b: EvalCoverageReconciliation | null;
+    };
+    /** Present only when BOTH sides carry WELL-FORMED coverage AND `identity.type_filter.same`. */
+    rule_coverage_delta?: EvalRuleCoverageDelta;
+}
+/** The S17 rule table, as a function. Exported so a caller can classify a single flip. */
+export declare function classifyFlip(kind: EvalSampleKind, from: EvalVerdict, to: EvalVerdict): {
+    classification: FlipClass;
+    reason: string;
+};
+/** Compare two recorded runs — A is the baseline, B the candidate. See the module doc. */
+export declare function compareEvalRuns(a: EvalRunDetail, b: EvalRunDetail): EvalRunComparison;
+//# sourceMappingURL=eval-compare.d.ts.map

package/dist/api/eval-compare.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval-compare.d.ts","sourceRoot":"","sources":["../../src/api/eval-compare.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsJG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,oBAAoB,EAAE,qBAAqB,EAA8B,YAAY,EAAE,MAAM,kBAAkB,CAAC;AAI7I,MAAM,MAAM,WAAW,GAAG,oBAAoB,CAAC,SAAS,CAAC,CAAC;AAC1D,MAAM,MAAM,cAAc,GAAG,oBAAoB,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,CAAC;AACpE,MAAM,MAAM,SAAS,GAAG,WAAW,GAAG,SAAS,CAAC;AAEhD,sGAAsG;AACtG,eAAO,MAAM,6BAA6B,mCAAmC,CAAC;AAE9E,4GAA4G;AAC5G,eAAO,MAAM,qBAAqB,0BAA0B,CAAC;AAE7D,4CAA4C;AAC5C,MAAM,WAAW,YAAY,CAAC,CAAC;IAC7B,CAAC,EAAE,CAAC,CAAC;IACL,CAAC,EAAE,CAAC,CAAC;IACL,IAAI,EAAE,OAAO,CAAC;CACf;AAED,6DAA6D;AAC7D,MAAM,WAAW,eAAe;IAC9B,SAAS,EAAE,MAAM,CAAC;IAClB,6FAA6F;IAC7F,IAAI,EAAE,cAAc,CAAC;IACrB,IAAI,EAAE,WAAW,CAAC;IAClB,EAAE,EAAE,WAAW,CAAC;IAChB,cAAc,EAAE,SAAS,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;CAChB;AAED;;;;4DAI4D;AAC5D,MAAM,MAAM,iBAAiB,GAAG,UAAU,GAAG,SAAS,CAAC;AAEvD;;kBAEkB;AAClB,eAAO,MAAM,uBAAuB,yCAAyC,CAAC;AAE9E;;;wEAGwE;AACxE,eAAO,MAAM,qBAAqB,wCAAwC,CAAC;AAE3E;;;;;;;;;;;GAWG;AACH,MAAM,MAAM,0BAA0B,GAAG,MAAM,GAAG,UAAU,GAAG,2CAA2C,GAAG,OAAO,uBAAuB,CAAC;AAE5I;qEACqE;AACrE,MAAM,WAAW,qBAAqB;IACpC,oFAAoF;IACpF,eAAe,EAAE,MAAM,CAAC;IACxB;;oGAEgG;IAChG,SAAS,EAAE,iBAAiB,CAAC;IAC7B;;;;qFAIiF;IACjF,YAAY,EAAE;QAAE,CAAC,EAAE,MAAM,CAAC;QAAC,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IACvC;;sFAEkF;IAClF,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,sGAAsG;IACtG,IAAI,EAAE,MAAM,EAAE,CAAC;IACf;;mCAE+B;IAC/B,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB;kBACc;IACd,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB;;;;iFAI6E;IAC7E,oBAAoB,EAAE,MAAM,EAAE,CAAC;CAChC;AAED,MAAM,WAAW,iBAAiB;IAChC,iFAAiF;IACjF,CAAC,EAAE,MAAM,CAAC;IACV,CAAC,EAAE,MAAM,CAAC;IACV,QAAQ,EAAE;QACR,MAAM,EAAE,YAAY,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC;QACpC,UAAU,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC;QACjC,WAAW,EAAE,YAAY,CAAC,YAAY,GAAG,IAAI,CAAC,CAAC;QAC/C,QAAQ,EAAE,YAAY,CAAC,YAAY,GAAG,IAAI,CAAC,CAAC;KAC7C,CAAC;IACF;kFAC8E;IAC9E,UAAU,EAAE,OAAO,CAAC;IACpB;;6FAEyF;IACzF,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,uFAAuF;IACvF,KAAK,EAAE,eAAe,EAAE,CAAC;IACzB;;;uGAGmG;IACnG,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,6FAA6F;IAC7F,YAAY,EAAE,MAAM,EAAE,CAAC;IACvB;;yDAEqD;IACrD,eAAe,EAAE,MAAM,EAAE,CAAC;IAC1B,qGAAqG;IACrG,eAAe,EAAE;QAAE,CAAC,EAAE,MAAM,CAAC;QAAC,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1C,+CAA+C;IAC/C,aAAa,EAAE,qBAAqB,CAAC;IACrC;;;;;gFAK4E;IAC5E,UAAU,EAAE,OAAO,CAAC;IACpB,qBAAqB,EAAE,MAAM,EAAE,CAAC;IAChC;;6DAEyD;IACzD,uBAAuB,EAAE;QAAE,CAAC,EAAE,0BAA0B,GAAG,IAAI,CAAC;QAAC,CAAC,EAAE,0BAA0B,GAAG,IAAI,CAAA;KAAE,CAAC;IACxG,+FAA+F;IAC/F,mBAAmB,CAAC,EAAE,qBAAqB,CAAC;CAC7C;AAED,0FAA0F;AAC1F,wBAAgB,YAAY,CAAC,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,WAAW,EAAE,EAAE,EAAE,WAAW,GAAG;IAAE,cAAc,EAAE,SAAS,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAYpI;AAED,0FAA0F;AAC1F,wBAAgB,eAAe,CAAC,CAAC,EAAE,aAAa,EAAE,CAAC,EAAE,aAAa,GAAG,iBAAiB,CAuJrF"}