@jinn-network/client 0.1.8 → 0.1.9-canary.050a41b1

package/dist/eval/slope.js

@@ -0,0 +1,46 @@
+/**
+ * Ordinary least-squares slope of resolved-rate vs cycle index for the
+ * train-arm slope measurement (issue #822, AC#1).
+ *
+ * The train-arm e2e evaluates a checkpoint against the held-out slate (#817)
+ * at intervals via the eval orchestrator (#818), collecting one
+ * `{ cycleIndex, rate }` point per interval (`rate` = passed / scorable, the
+ * Wilson point estimate). The slope of the least-squares fit is the headline
+ * "is the learner improving across the training sequence" number.
+ *
+ * It is deliberately a thin helper over the closed-form OLS slope
+ * (`cov(x,y) / var(x)`); the per-point confidence intervals come from
+ * `wilson.ts` — this module does NOT reimplement them. The slope sign alone is
+ * never a verdict at small N: a flat or slightly negative slope is "within
+ * noise", which the e2e surfaces via the §4.1 honesty caveat.
+ */
+/**
+ * Least-squares slope of `rate` regressed on `cycleIndex`. Returns 0 for fewer
+ * than two points (no line to fit) and for a degenerate fit where every x is
+ * identical (zero variance — division would be NaN). A flat sequence yields
+ * exactly 0.
+ */
+export function leastSquaresSlope(points) {
+    const n = points.length;
+    if (n < 2)
+        return 0;
+    let sumX = 0;
+    let sumY = 0;
+    for (const { cycleIndex, rate } of points) {
+        sumX += cycleIndex;
+        sumY += rate;
+    }
+    const meanX = sumX / n;
+    const meanY = sumY / n;
+    let cov = 0;
+    let varX = 0;
+    for (const { cycleIndex, rate } of points) {
+        const dx = cycleIndex - meanX;
+        cov += dx * (rate - meanY);
+        varX += dx * dx;
+    }
+    if (varX === 0)
+        return 0;
+    return cov / varX;
+}
+//# sourceMappingURL=slope.js.map

package/dist/eval/slope.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"slope.js","sourceRoot":"","sources":["../../src/eval/slope.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AASH;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAAC,MAAmB;IACnD,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IACxB,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,CAAC,CAAC;IACpB,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,MAAM,EAAE,UAAU,EAAE,IAAI,EAAE,IAAI,MAAM,EAAE,CAAC;QAC1C,IAAI,IAAI,UAAU,CAAC;QACnB,IAAI,IAAI,IAAI,CAAC;IACf,CAAC;IACD,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,CAAC;IACvB,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,CAAC;IACvB,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,MAAM,EAAE,UAAU,EAAE,IAAI,EAAE,IAAI,MAAM,EAAE,CAAC;QAC1C,MAAM,EAAE,GAAG,UAAU,GAAG,KAAK,CAAC;QAC9B,GAAG,IAAI,EAAE,GAAG,CAAC,IAAI,GAAG,KAAK,CAAC,CAAC;QAC3B,IAAI,IAAI,EAAE,GAAG,EAAE,CAAC;IAClB,CAAC;IACD,IAAI,IAAI,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IACzB,OAAO,GAAG,GAAG,IAAI,CAAC;AACpB,CAAC"}

package/dist/eval/train-sequence.d.ts

@@ -0,0 +1,35 @@
+/**
+ * Train-sequence builder with the no-train/test-overlap guard for the
+ * train-arm slope e2e (issue #822, AC#2).
+ *
+ * The learner-full-cycle e2e drives `runCycle` DIRECTLY (it does not post tasks
+ * through the generator), so the generator's `excludeHeldOutSlate` train-stream
+ * chokepoint (#817) is BYPASSED. This builder is therefore the load-bearing
+ * AC#2 mechanism for the e2e: it selects the N distinct training instances from
+ * the pool with the held-out slate excluded, and asserts the resulting
+ * sequence is disjoint from the slate (fail-loud, never a silent drop).
+ *
+ * It reuses `excludeHeldOutSlate` from the #817 primitive rather than
+ * reimplementing the exclusion. Selection is deterministic (instance-id sorted)
+ * so a given pool yields a stable sequence across runs.
+ */
+import type { PoolTask } from '../solver-types/_swe-rebench-v2-pool.js';
+/** Thrown when a chosen training sequence intersects the held-out slate (AC#2). */
+export declare class TrainTestOverlapError extends Error {
+    readonly overlap: string[];
+    constructor(overlap: string[]);
+}
+/** Assert a set of training ids is disjoint from the slate, else throw loud. */
+export declare function assertNoOverlap(trainIds: string[], slateIds: Set<string>): void;
+export declare function buildTrainSequence(args: {
+    pool: PoolTask[];
+    slateIds: Set<string>;
+    /** Number of distinct training tasks (= N training cycles). */
+    count: number;
+    /**
+     * Optional explicit, hand-picked instance_ids (in order). When set, the
+     * builder still runs the no-overlap guard and resolves each id against the
+     * pool — used to fail-loud on a hand-edited sequence that overlaps the slate.
+     */
+    explicitIds?: string[];
+}): PoolTask[];

package/dist/eval/train-sequence.js

@@ -0,0 +1,59 @@
+/**
+ * Train-sequence builder with the no-train/test-overlap guard for the
+ * train-arm slope e2e (issue #822, AC#2).
+ *
+ * The learner-full-cycle e2e drives `runCycle` DIRECTLY (it does not post tasks
+ * through the generator), so the generator's `excludeHeldOutSlate` train-stream
+ * chokepoint (#817) is BYPASSED. This builder is therefore the load-bearing
+ * AC#2 mechanism for the e2e: it selects the N distinct training instances from
+ * the pool with the held-out slate excluded, and asserts the resulting
+ * sequence is disjoint from the slate (fail-loud, never a silent drop).
+ *
+ * It reuses `excludeHeldOutSlate` from the #817 primitive rather than
+ * reimplementing the exclusion. Selection is deterministic (instance-id sorted)
+ * so a given pool yields a stable sequence across runs.
+ */
+import { excludeHeldOutSlate } from '../solver-types/_swe-rebench-v2-held-out-slate.js';
+/** Thrown when a chosen training sequence intersects the held-out slate (AC#2). */
+export class TrainTestOverlapError extends Error {
+    overlap;
+    constructor(overlap) {
+        super(`train/test overlap: training sequence includes held-out slate instance(s) ` +
+            `${overlap.join(', ')} — refusing to train on the eval slate (AC#2). ` +
+            `The slate must stay out-of-sample for the slope to mean anything.`);
+        this.overlap = overlap;
+        this.name = 'TrainTestOverlapError';
+    }
+}
+/** Assert a set of training ids is disjoint from the slate, else throw loud. */
+export function assertNoOverlap(trainIds, slateIds) {
+    const overlap = trainIds.filter((id) => slateIds.has(id));
+    if (overlap.length > 0)
+        throw new TrainTestOverlapError(overlap);
+}
+export function buildTrainSequence(args) {
+    const eligible = excludeHeldOutSlate(args.pool, args.slateIds);
+    const byId = new Map(eligible.map((t) => [t.instance_id, t]));
+    if (args.explicitIds) {
+        // Guard the hand-picked sequence against the slate BEFORE resolving, so a
+        // slate-overlapping id is a TrainTestOverlapError, not a "not eligible".
+        assertNoOverlap(args.explicitIds, args.slateIds);
+        return args.explicitIds.map((id) => {
+            const task = byId.get(id);
+            if (!task) {
+                throw new Error(`explicit training instance ${id} not in the eligible pool`);
+            }
+            return task;
+        });
+    }
+    // Deterministic selection: instance-id sorted, first `count`.
+    const sorted = [...eligible].sort((a, b) => a.instance_id.localeCompare(b.instance_id));
+    if (sorted.length < args.count) {
+        throw new Error(`train sequence needs ${args.count} distinct tasks but only ${sorted.length} eligible ` +
+            `(pool size ${args.pool.length} minus ${args.slateIds.size} held-out slate instance(s))`);
+    }
+    const picked = sorted.slice(0, args.count);
+    assertNoOverlap(picked.map((t) => t.instance_id), args.slateIds);
+    return picked;
+}
+//# sourceMappingURL=train-sequence.js.map

package/dist/eval/train-sequence.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"train-sequence.js","sourceRoot":"","sources":["../../src/eval/train-sequence.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAGH,OAAO,EAAE,mBAAmB,EAAE,MAAM,mDAAmD,CAAC;AAExF,mFAAmF;AACnF,MAAM,OAAO,qBAAsB,SAAQ,KAAK;IAClB;IAA5B,YAA4B,OAAiB;QAC3C,KAAK,CACH,4EAA4E;YAC1E,GAAG,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,iDAAiD;YACtE,mEAAmE,CACtE,CAAC;QALwB,YAAO,GAAP,OAAO,CAAU;QAM3C,IAAI,CAAC,IAAI,GAAG,uBAAuB,CAAC;IACtC,CAAC;CACF;AAED,gFAAgF;AAChF,MAAM,UAAU,eAAe,CAAC,QAAkB,EAAE,QAAqB;IACvE,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC;IAC1D,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC;QAAE,MAAM,IAAI,qBAAqB,CAAC,OAAO,CAAC,CAAC;AACnE,CAAC;AAED,MAAM,UAAU,kBAAkB,CAAC,IAWlC;IACC,MAAM,QAAQ,GAAG,mBAAmB,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;IAC/D,MAAM,IAAI,GAAG,IAAI,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IAE9D,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;QACrB,0EAA0E;QAC1E,yEAAyE;QACzE,eAAe,CAAC,IAAI,CAAC,WAAW,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;QACjD,OAAO,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,EAAE;YACjC,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC1B,IAAI,CAAC,IAAI,EAAE,CAAC;gBACV,MAAM,IAAI,KAAK,CAAC,8BAA8B,EAAE,2BAA2B,CAAC,CAAC;YAC/E,CAAC;YACD,OAAO,IAAI,CAAC;QACd,CAAC,CAAC,CAAC;IACL,CAAC;IAED,8DAA8D;IAC9D,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,CAAC,aAAa,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC;IACxF,IAAI,MAAM,CAAC,MAAM,GAAG,IAAI,CAAC,KAAK,EAAE,CAAC;QAC/B,MAAM,IAAI,KAAK,CACb,wBAAwB,IAAI,CAAC,KAAK,4BAA4B,MAAM,CAAC,MAAM,YAAY;YACrF,cAAc,IAAI,CAAC,IAAI,CAAC,MAAM,UAAU,IAAI,CAAC,QAAQ,CAAC,IAAI,8BAA8B,CAC3F,CAAC;IACJ,CAAC;IACD,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;IAC3C,eAAe,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,CAAC,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;IACjE,OAAO,MAAM,CAAC;AAChB,CAAC"}

package/dist/eval/wilson.d.ts

@@ -0,0 +1,45 @@
+/**
+ * Wilson score interval + resolved-rate comparison for the `jinn eval`
+ * held-out checkpoint orchestrator (issue #818, AC#2).
+ *
+ * The Wilson score interval is a binomial proportion confidence interval that
+ * behaves well at the extremes (p=0, p=1) and for small n — unlike the naive
+ * normal-approximation interval. We write it small (no stats dependency, per
+ * repo convention): the formula is ~10 lines.
+ *
+ * Per log/decisions/2026-05-28-rl-eval-measurement.md §4: this is v1-simple.
+ * Only *large* deltas are trustworthy — we encode that as "the child and
+ * parent intervals do not overlap." No seed control, no multi-run averaging.
+ */
+export interface Interval {
+    /** Observed point estimate, passed / scorable (0 when scorable=0). */
+    p: number;
+    /** Lower bound, clamped to [0, 1]. */
+    lo: number;
+    /** Upper bound, clamped to [0, 1]. */
+    hi: number;
+}
+/**
+ * Wilson score interval for `passed` successes out of `scorable` trials.
+ * `scorable === 0` returns a degenerate `{ p: 0, lo: 0, hi: 0 }` (no NaN).
+ */
+export declare function wilsonInterval(passed: number, scorable: number, z?: number): Interval;
+export type RateVerdict = 'trustworthy' | 'within-noise';
+export interface RateComparison {
+    child: Interval;
+    parent: Interval;
+    /** child.p − parent.p (point-estimate difference, can be negative). */
+    delta: number;
+    /**
+     * 'trustworthy' iff the two Wilson intervals do NOT overlap; otherwise
+     * 'within-noise'. v1-simple: only disjoint intervals justify a claim.
+     */
+    verdict: RateVerdict;
+}
+export declare function compareRates(child: {
+    passed: number;
+    scorable: number;
+}, parent: {
+    passed: number;
+    scorable: number;
+}): RateComparison;

package/dist/eval/wilson.js

@@ -0,0 +1,48 @@
+/**
+ * Wilson score interval + resolved-rate comparison for the `jinn eval`
+ * held-out checkpoint orchestrator (issue #818, AC#2).
+ *
+ * The Wilson score interval is a binomial proportion confidence interval that
+ * behaves well at the extremes (p=0, p=1) and for small n — unlike the naive
+ * normal-approximation interval. We write it small (no stats dependency, per
+ * repo convention): the formula is ~10 lines.
+ *
+ * Per log/decisions/2026-05-28-rl-eval-measurement.md §4: this is v1-simple.
+ * Only *large* deltas are trustworthy — we encode that as "the child and
+ * parent intervals do not overlap." No seed control, no multi-run averaging.
+ */
+/** Two-sided z for a 95% interval (1.96 ≈ Φ⁻¹(0.975)). */
+const DEFAULT_Z = 1.96;
+/**
+ * Wilson score interval for `passed` successes out of `scorable` trials.
+ * `scorable === 0` returns a degenerate `{ p: 0, lo: 0, hi: 0 }` (no NaN).
+ */
+export function wilsonInterval(passed, scorable, z = DEFAULT_Z) {
+    if (scorable === 0)
+        return { p: 0, lo: 0, hi: 0 };
+    const n = scorable;
+    const p = passed / n;
+    const z2 = z * z;
+    const denom = 1 + z2 / n;
+    const centre = p + z2 / (2 * n);
+    const margin = z * Math.sqrt((p * (1 - p)) / n + z2 / (4 * n * n));
+    const lo = (centre - margin) / denom;
+    const hi = (centre + margin) / denom;
+    return {
+        p,
+        lo: Math.max(0, lo),
+        hi: Math.min(1, hi),
+    };
+}
+export function compareRates(child, parent) {
+    const c = wilsonInterval(child.passed, child.scorable);
+    const p = wilsonInterval(parent.passed, parent.scorable);
+    const disjoint = c.lo > p.hi || p.lo > c.hi;
+    return {
+        child: c,
+        parent: p,
+        delta: c.p - p.p,
+        verdict: disjoint ? 'trustworthy' : 'within-noise',
+    };
+}
+//# sourceMappingURL=wilson.js.map

package/dist/eval/wilson.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"wilson.js","sourceRoot":"","sources":["../../src/eval/wilson.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,0DAA0D;AAC1D,MAAM,SAAS,GAAG,IAAI,CAAC;AAWvB;;;GAGG;AACH,MAAM,UAAU,cAAc,CAAC,MAAc,EAAE,QAAgB,EAAE,IAAY,SAAS;IACpF,IAAI,QAAQ,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,EAAE,CAAC;IAClD,MAAM,CAAC,GAAG,QAAQ,CAAC;IACnB,MAAM,CAAC,GAAG,MAAM,GAAG,CAAC,CAAC;IACrB,MAAM,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC;IACjB,MAAM,KAAK,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;IACzB,MAAM,MAAM,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAChC,MAAM,MAAM,GAAG,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IACnE,MAAM,EAAE,GAAG,CAAC,MAAM,GAAG,MAAM,CAAC,GAAG,KAAK,CAAC;IACrC,MAAM,EAAE,GAAG,CAAC,MAAM,GAAG,MAAM,CAAC,GAAG,KAAK,CAAC;IACrC,OAAO;QACL,CAAC;QACD,EAAE,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC;QACnB,EAAE,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC;KACpB,CAAC;AACJ,CAAC;AAgBD,MAAM,UAAU,YAAY,CAC1B,KAA2C,EAC3C,MAA4C;IAE5C,MAAM,CAAC,GAAG,cAAc,CAAC,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;IACvD,MAAM,CAAC,GAAG,cAAc,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,QAAQ,CAAC,CAAC;IACzD,MAAM,QAAQ,GAAG,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAE,CAAC;IAC5C,OAAO;QACL,KAAK,EAAE,CAAC;QACR,MAAM,EAAE,CAAC;QACT,KAAK,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QAChB,OAAO,EAAE,QAAQ,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,cAAc;KACnD,CAAC;AACJ,CAAC"}

package/dist/harnesses/engine/canonical-json.js

@@ -22,9 +22,11 @@
  * Used for manifest signing: produce a deterministic byte string that two
  * independent parties can reproduce from the same object graph.
  */
-import { createRequire } from 'node:module';
-const require = createRequire(import.meta.url);
-const canonicalize = require('canonicalize');
+// canonicalize v3 is ESM-only (no CJS entry); the client is `"type": "module"`
+// so a native ESM default import is the correct interop (replaces the v2-era
+// createRequire shim). v3 is packaging-only vs v2 — its RFC 8785 output is
+// byte-identical, so existing content hashes / manifest signatures are stable.
+import canonicalize from 'canonicalize';
 /**
  * Recursively replace NaN / ±Infinity with null so that canonicalize does not
  * throw — matching the JSON.stringify behaviour that the rest of the codebase

package/dist/harnesses/engine/canonical-json.js.map

@@ -1 +1 @@
-{"version":3,"file":"canonical-json.js","sourceRoot":"","sources":["../../../src/harnesses/engine/canonical-json.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,OAAO,EAAE,aAAa,EAAE,MAAM,aAAa,CAAC;AAC5C,MAAM,OAAO,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;AAC/C,MAAM,YAAY,GAAG,OAAO,CAAC,cAAc,CAA2C,CAAC;AAEvF;;;;;GAKG;AACH,SAAS,eAAe,CAAC,KAAc;IACrC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC/D,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC,GAAG,CAAC,eAAe,CAAC,CAAC;IAC5D,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAChD,mEAAmE;QACnE,qEAAqE;QACrE,sEAAsE;QACtE,WAAW;QACX,MAAM,MAAM,GAAI,KAAoC,CAAC,MAAM,CAAC;QAC5D,IAAI,OAAO,MAAM,KAAK,UAAU,EAAE,CAAC;YACjC,OAAO,eAAe,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;QAC7C,CAAC;QACD,MAAM,GAAG,GAA4B,EAAE,CAAC;QACxC,KAAK,MAAM,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC;YAAE,GAAG,CAAC,CAAC,CAAC,GAAG,eAAe,CAAC,CAAC,CAAC,CAAC;QACxE,OAAO,GAAG,CAAC;IACb,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC;AAED,MAAM,UAAU,aAAa,CAAC,KAAc;IAC1C,MAAM,OAAO,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC;IAEvC,MAAM,MAAM,GAAG,YAAY,CAAC,OAAO,CAAC,CAAC;IACrC,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;QACzB,wEAAwE;QACxE,uEAAuE;QACvE,kCAAkC;QAClC,OAAO,MAAM,CAAC;IAChB,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC"}
+{"version":3,"file":"canonical-json.js","sourceRoot":"","sources":["../../../src/harnesses/engine/canonical-json.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,+EAA+E;AAC/E,6EAA6E;AAC7E,2EAA2E;AAC3E,+EAA+E;AAC/E,OAAO,YAAY,MAAM,cAAc,CAAC;AAExC;;;;;GAKG;AACH,SAAS,eAAe,CAAC,KAAc;IACrC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC/D,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC,GAAG,CAAC,eAAe,CAAC,CAAC;IAC5D,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAChD,mEAAmE;QACnE,qEAAqE;QACrE,sEAAsE;QACtE,WAAW;QACX,MAAM,MAAM,GAAI,KAAoC,CAAC,MAAM,CAAC;QAC5D,IAAI,OAAO,MAAM,KAAK,UAAU,EAAE,CAAC;YACjC,OAAO,eAAe,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;QAC7C,CAAC;QACD,MAAM,GAAG,GAA4B,EAAE,CAAC;QACxC,KAAK,MAAM,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC;YAAE,GAAG,CAAC,CAAC,CAAC,GAAG,eAAe,CAAC,CAAC,CAAC,CAAC;QACxE,OAAO,GAAG,CAAC;IACb,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC;AAED,MAAM,UAAU,aAAa,CAAC,KAAc;IAC1C,MAAM,OAAO,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC;IAEvC,MAAM,MAAM,GAAG,YAAY,CAAC,OAAO,CAAC,CAAC;IACrC,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;QACzB,wEAAwE;QACxE,uEAAuE;QACvE,kCAAkC;QAClC,OAAO,MAAM,CAAC;IAChB,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC"}

package/dist/harnesses/engine/engine.d.ts

@@ -70,6 +70,10 @@ export interface JoinedSolverNetsView {
     } | undefined;
     /** Enumerate all joined manifest CIDs (used for digest-based filtering). */
     manifestCids(): string[];
+    /** Add/replace one joined entry live (used by the hot-apply join applier, #1037). */
+    set(manifestCid: string, entry: {
+        roles: Array<'solver' | 'evaluator'>;
+    }): void;
 }
 /**
  * Build a `JoinedSolverNetsView` from the raw operator-config block.
@@ -82,6 +86,17 @@ export declare function joinedSolverNetsViewFromConfig(joined: Record<string, {
     manifestCid: string;
     roles: Array<'solver' | 'evaluator'>;
 }> | undefined): JoinedSolverNetsView | undefined;
+/**
+ * Mutable `JoinedSolverNetsView` for the running daemon. Unlike
+ * `joinedSolverNetsViewFromConfig` (boot snapshot), the applier
+ * (`daemon/join-applier.ts`, #1037) keeps a handle and calls `set()` when a
+ * join is hot-applied, so the engine's per-task eligibility check sees the new
+ * cid on its next call without a restart.
+ */
+export declare function createMutableJoinedSolverNetsView(initial: Record<string, {
+    manifestCid: string;
+    roles: Array<'solver' | 'evaluator'>;
+}> | undefined): JoinedSolverNetsView;
 /**
  * Resolves a launched SolverNet manifest by IPFS CID.
  *
@@ -576,6 +591,15 @@ export declare class TaskEngine {
      *   TCAttemptAlreadyFinalized, …). We mark the row RACE_LOST and emit a
      *   `kind=race_lost` activity event so operators can audit prunes
      *   without inflating the FAILED counter (#896).
+     * - A transport-transient error (e.g. `AllRpcsFailedError` — every provider
+     *   in the L2 fallback chain failed at once) on a task whose delivery window
+     *   is still open: leave the row in its current in-flight state so the next
+     *   tick re-drives it once the RPCs recover, and emit a `tick_error` (warn)
+     *   event instead of inflating the FAILED counter. Without this the daemon
+     *   stamped the row FAILED, dropping it from `getInFlight()` permanently, so
+     *   L2 work went silent until a manual restart (#912). Past-window transient
+     *   errors still terminalize to avoid churning on work that can no longer
+     *   settle.
      * - Everything else: existing markFailed behaviour. When invoked from
      *   recovery, `contextLabel === 'recovery'` so the failure_reason
      *   carries the `recovery:` prefix the original code path used.

package/dist/harnesses/engine/engine.js

@@ -18,6 +18,7 @@ import { assembleAndSignEnvelope, } from './envelope-assembly.js';
 import { deliverAndClaim, } from './delivery.js';
 import { SafeInnerRevertError, isNonRecoverableInnerRevert, formatDecodedRevert, } from '../../adapters/mech/safe-revert.js';
 import { emitEvent } from '../../observability/emit-event.js';
+import { isRecoverableTransactionError } from '../../tx-retry.js';
 import { SkippableError } from '../types.js';
 import { submitEvaluatorFeedback, codeDigestSha256ToBytes32, encodeExecutionPayload, encodeExecutionPayloadV2, modeStringToFlag, } from '../../erc8004/index.js';
 import { TrajectoryCollector, emitTrajectory } from '../../trajectory/index.js';
@@ -26,6 +27,7 @@ import { VerdictCode } from '../../adapters/mech/verdict-code.js';
 import { buildInfo } from '../../build-info.js';
 import { getSolverNetContract } from '../../vendor/@jinn-network/sdk/dist/solvernets/index.js';
 import { runHarnessWithFreezeFence, } from '../../daemon/freeze-fence.js';
+import { recordLoopTick } from '../../daemon/loop-heartbeat.js';
 import { harnessStateDirName } from '../names.js';
 import { recordTaskCost } from '../../spend/record.js';
 // ── Sentinel error ────────────────────────────────────────────────────────────
@@ -61,6 +63,25 @@ export function joinedSolverNetsViewFromConfig(joined) {
     return {
         get: (cid) => map.get(cid),
         manifestCids: () => [...map.keys()],
+        set: (cid, entry) => { map.set(cid, entry); },
+    };
+}
+/**
+ * Mutable `JoinedSolverNetsView` for the running daemon. Unlike
+ * `joinedSolverNetsViewFromConfig` (boot snapshot), the applier
+ * (`daemon/join-applier.ts`, #1037) keeps a handle and calls `set()` when a
+ * join is hot-applied, so the engine's per-task eligibility check sees the new
+ * cid on its next call without a restart.
+ */
+export function createMutableJoinedSolverNetsView(initial) {
+    const map = new Map();
+    for (const [key, entry] of Object.entries(initial ?? {})) {
+        map.set(entry.manifestCid ?? key, { roles: entry.roles });
+    }
+    return {
+        get: (cid) => map.get(cid),
+        manifestCids: () => [...map.keys()],
+        set: (cid, entry) => { map.set(cid, entry); },
     };
 }
 // ── TaskEngine ─────────────────────────────────────────────────────────
@@ -281,6 +302,7 @@ export class TaskEngine {
             catch (err) {
                 console.error('[harness-engine] tick loop error (continuing):', err instanceof Error ? err.message : err);
             }
+            recordLoopTick(this.store, 'engine-tick'); // #1043 loop watchdog
             if (this.stopped)
                 break;
             await Promise.race([
@@ -761,7 +783,17 @@ export class TaskEngine {
             throw new NotImplementedError('runImpl');
         }
         const runtimePlugins = solverNet?.runtimePlugins ?? [];
-        this.runtimePluginsByRequest.set(task.requestId, runtimePlugins);
+        // #1035: merge harness self-attributed plugins (e.g. claude-code-learner)
+        // into the envelope carrier so they appear in executor.plugins. This is a
+        // SEPARATE array from `runtimePlugins`: the latter still feeds
+        // ctx.runtimePlugins / ctx.solverPluginRoots (which the harness uses to
+        // LOAD solver plugins), and the learner plugin is already loaded by the
+        // harness itself via its own plugin root — adding it there would double-load.
+        const attributedPlugins = [
+            ...runtimePlugins,
+            ...(impl.attributionPlugins?.() ?? []),
+        ];
+        this.runtimePluginsByRequest.set(task.requestId, attributedPlugins);
         const workingDir = task.workingDir ?? join(this.paths.workingDirRoot, task.requestId);
         const kindSeg = solverType.replace(/[.:]/g, '_');
         const implStateDir = task.implStateDir ?? (kindSeg
@@ -846,7 +878,7 @@ export class TaskEngine {
                         informationalClaim: skippedOutput.informational ?? null,
                         solutionOutputsJson: JSON.stringify(skippedOutput),
                         implName: impl.name,
-                        runtimePluginsJson: JSON.stringify(runtimePlugins),
+                        runtimePluginsJson: JSON.stringify(attributedPlugins),
                     });
                     console.log(`[harness-engine] ${task.requestId} RUNNING → POST_SNAPSHOT via impl=${impl.name} (skipped)`);
                     return;
@@ -890,7 +922,7 @@ export class TaskEngine {
                 informationalClaim: output.informational ?? null,
                 solutionOutputsJson: JSON.stringify(output),
                 implName: impl.name,
-                runtimePluginsJson: JSON.stringify(runtimePlugins),
+                runtimePluginsJson: JSON.stringify(attributedPlugins),
             });
         }
         finally {
@@ -1488,10 +1520,7 @@ export class TaskEngine {
             case 'UNRESOLVED':
                 return VerdictCode.Unresolved;
             default:
-                // gatingClaim is null, verdict is absent, or the string is unrecognized.
-                // Return Invalid(3) — not Pass(1). Pass must come from an explicit PASS/SCORED verdict.
-                console.warn(`[harness-engine] verdictCodeForTask: unrecognized gatingClaim.verdict (got=${String(raw)}); defaulting to Invalid(3) — should never happen, indicates the evaluator harness didn't set gatingClaim.verdict before submission`);
-                return VerdictCode.Invalid;
+                throw new Error(`[harness-engine] verdictCodeForTask: missing or unrecognized gatingClaim.verdict (got=${String(raw)}); refusing to claim Invalid(3) on-chain without an explicit evaluator verdict`);
         }
     }
     async _maybePostEvaluatorFeedback(task) {
@@ -1630,6 +1659,15 @@ export class TaskEngine {
      *   TCAttemptAlreadyFinalized, …). We mark the row RACE_LOST and emit a
      *   `kind=race_lost` activity event so operators can audit prunes
      *   without inflating the FAILED counter (#896).
+     * - A transport-transient error (e.g. `AllRpcsFailedError` — every provider
+     *   in the L2 fallback chain failed at once) on a task whose delivery window
+     *   is still open: leave the row in its current in-flight state so the next
+     *   tick re-drives it once the RPCs recover, and emit a `tick_error` (warn)
+     *   event instead of inflating the FAILED counter. Without this the daemon
+     *   stamped the row FAILED, dropping it from `getInFlight()` permanently, so
+     *   L2 work went silent until a manual restart (#912). Past-window transient
+     *   errors still terminalize to avoid churning on work that can no longer
+     *   settle.
      * - Everything else: existing markFailed behaviour. When invoked from
      *   recovery, `contextLabel === 'recovery'` so the failure_reason
      *   carries the `recovery:` prefix the original code path used.
@@ -1652,6 +1690,24 @@ export class TaskEngine {
             return 'race_lost';
         }
         const reason = err instanceof Error ? err.message : String(err);
+        // A transport-transient failure (all RPC providers in the fallback chain
+        // blipped at once, 429s, timeouts, …) is not the task's fault and is not
+        // permanent. Leave the row in its in-flight state — do NOT call
+        // markFailed, which would drop it from getInFlight() forever (#912) — so
+        // the engine-tick loop re-drives it once the RPCs recover. The tick loop
+        // IS the retry; there is no per-task attempt counter. Skip this only once
+        // the delivery window has closed, so we never churn on work that can no
+        // longer settle on-chain.
+        if (task.windowEndTs > Date.now() && isRecoverableTransactionError(err)) {
+            emitEvent(this.store, {
+                kind: 'tick_error',
+                requestId: task.requestId,
+                solverType: task.solverType ?? undefined,
+                outcome: 'warn',
+                detail: `transient RPC failure in ${contextLabel}; left ${task.state} for retry: ${reason}`,
+            }, 'harness-engine');
+            return 'transient';
+        }
         const stamped = contextLabel === 'recovery' ? `recovery: ${reason}` : reason;
         this.persistence.markFailed(task.requestId, stamped);
         return 'failed';
@@ -1695,8 +1751,15 @@ export class TaskEngine {
             if (current && current.state === task.state) {
                 const classification = this._classifyAndMarkTerminal(task, err, 'recovery');
                 const reason = err instanceof Error ? err.message : String(err);
-                const log = classification === 'race_lost' ? console.log : console.error;
-                const verb = classification === 'race_lost' ? 'pruned' : 'failed';
+                // 'transient' leaves the row in-flight (not terminal); the next tick
+                // re-drives it once the RPCs recover (#912). Log it at warn so the
+                // stall is visible without firing the error-level alerting that a
+                // genuine failure does.
+                const { log, verb } = {
+                    race_lost: { log: console.log, verb: 'pruned' },
+                    transient: { log: console.warn, verb: 'deferred (transient RPC)' },
+                    failed: { log: console.error, verb: 'failed' },
+                }[classification];
                 log(`[harness-engine] resume ${verb} for ${task.requestId}: ${reason}`);
             }
             throw err;