tickmarkr 2.6.1 → 2.6.3

package/dist/gates/run-gates.js

@@ -1,8 +1,8 @@
-import { randomUUID } from "node:crypto";
+import { createHash, randomUUID } from "node:crypto";
 import { existsSync, mkdtempSync, readFileSync, rmSync, statSync } from "node:fs";
 import { loadavg, tmpdir } from "node:os";
 import { join, posix } from "node:path";
-import { channelKey, shq } from "../adapters/types.js";
+import { channelKey, configuredEffort, shq } from "../adapters/types.js";
 import { TIER_RANK } from "../config/config.js";
 import { getAdapter } from "../adapters/registry.js";
 import { GATE_NAMES } from "../graph/schema.js";
@@ -10,14 +10,14 @@ import { acceptanceGate } from "./acceptance.js";
 import { compareToBaseline, effectiveCeilingMs, waitForCalmWindow, calmWindowReady } from "./baseline.js";
 import { evidenceGate } from "./evidence.js";
 import { captureLlmOutput } from "./llm.js";
-import { disallowedBy } from "../route/preference.js";
+import { disallowedBy, observedSeat } from "../route/preference.js";
 import { marginalCostRank } from "../route/router.js";
 import { carriedAuthorVendors, gateReviewerFloor, pickReviewer, reviewGate } from "./review.js";
 import { scopeGate } from "./scope.js";
-import { evaluateManifestedTest, isVitestTestCommand } from "./test-manifest.js";
+import { discoverTestManifest, evaluateManifestedTest, isVitestTestCommand, VITEST_CACHE_ENV, worktreeVitestCache } from "./test-manifest.js";
 import { executionSignal } from "../run/execution-budget.js";
 import { failureDisposition } from "../run/recovery.js";
-import { dependencyLinkRefusal, preserveWorktree, producerFields, shGit, resolvedCapacity, verificationProtocol } from "../run/git.js";
+import { dependencyLinkRefusal, FORK_CAP_ENV, preserveWorktree, producerFields, ROUTING_ENV_SEAMS, shGit, resolvedCapacity, SUITE_PARENT_ENV, verificationProtocol } from "../run/git.js";
 import { withJudgeInvocationEvidence } from "../run/journal.js";
 import { computeVerificationIdentity, verificationIdentityKey, formatReusedRow, getVerdictStore, isInfraResult, resolveStateDir, reusedIdentity, } from "./cache.js";
 const productionLoadProvider = () => loadavg()[0] ?? 0;
@@ -43,13 +43,14 @@ function instrumentLlmAdapter(adapter, clocks) {
     return new Proxy(adapter, {
         get(target, property) {
             if (property === "headlessCommand") {
-                return (promptFile, model) => {
-                    const command = target.headlessCommand(promptFile, model);
+                return (promptFile, model, effort) => {
+                    const command = target.headlessCommand(promptFile, model, effort);
                     const dir = mkdtempSync(join(tmpdir(), "tickmarkr-gate-invocation-"));
                     const startedAtPath = join(dir, "started-at");
                     const completedAtPath = join(dir, "completed-at");
                     clocks.push({
                         channel: channelKey({ adapter: target.id, model }),
+                        ...(effort ? { effort } : {}),
                         preparedAt: Date.now(),
                         startedAtPath,
                         completedAtPath,
@@ -85,7 +86,7 @@ function finishLlmDispatches(clocks) {
         finally {
             rmSync(clock.dir, { recursive: true, force: true });
         }
-        return { channel: clock.channel, durationMs: completedAt - startedAt };
+        return { channel: clock.channel, ...(clock.effort ? { effort: clock.effort } : {}), durationMs: completedAt - startedAt };
     });
 }
 async function captureLlmDispatches(adapters, run) {
@@ -214,6 +215,44 @@ export function testCommandForFiles(testCmd, files) {
     const fwd = wrapped && !/\s--\s/.test(testCmd) ? " --" : "";
     return `${testCmd}${fwd} ${files.map(shq).join(" ")}`;
 }
+/** OBS-635: a screen costing at least this share of the full suite runs the full suite instead. */
+export const SCREEN_PROMOTION_RATIO = 0.75;
+/**
+ * The screen's share of the full suite's cost, from the per-file durations the harness measured at
+ * baseline capture — never a worker's timing. Undefined (unknown) unless every selected file has a
+ * measured duration and the measured total is positive; unknown keeps the conservative screen path.
+ */
+export function screenCostRatio(baseline, selected) {
+    const entry = baseline.commands.test;
+    const files = entry?.infra ? undefined : entry?.fileDurations;
+    if (!files?.length)
+        return undefined;
+    const cost = new Map(files.map((f) => [f.file, f.durationMs]));
+    const total = files.reduce((sum, f) => sum + f.durationMs, 0);
+    if (!(total > 0) || selected.some((file) => !cost.has(file)))
+        return undefined;
+    return selected.reduce((sum, file) => sum + cost.get(file), 0) / total;
+}
+/** OBS-635: the full manifest the runner lists NOW, under the environment evaluateManifestedTest's own
+ * discovery receives (test-manifest.ts manifestEnvironment), so it compares with the one a verdict
+ * certified. Undefined when the runner cannot list — an unlisted manifest certifies nothing. */
+async function listFullManifest(cmd, worktree) {
+    const env = { ...process.env, PATH: `${join(worktree, "node_modules/.bin")}:${process.env.PATH ?? ""}`,
+        [VITEST_CACHE_ENV]: worktreeVitestCache(worktree),
+        [FORK_CAP_ENV]: String(resolvedCapacity().forkCap), [SUITE_PARENT_ENV]: String(process.pid) };
+    for (const key of [...ROUTING_ENV_SEAMS, "VITEST", "TEST", "VITEST_WORKER_ID", "VITEST_POOL_ID"])
+        delete env[key];
+    const dir = mkdtempSync(join(tmpdir(), "tickmarkr-full-manifest-"));
+    try {
+        return (await discoverTestManifest(cmd, worktree, { dir, nonce: randomUUID(), env })).files;
+    }
+    catch {
+        return undefined;
+    }
+    finally {
+        rmSync(dir, { recursive: true, force: true });
+    }
+}
 /** The manifest-report path for a detected vitest test command — never the stdout-count/file-count path. */
 async function runVitestManifestGate(worktree, cmd, baseline, selected, artifactDir, retry = {}, retried = false) {
     const entry = baseline.commands.test;
@@ -248,6 +287,25 @@ async function runVitestManifestGate(worktree, cmd, baseline, selected, artifact
         meta: { ...outcome.meta, reportPath, ...(selected ? { selectedTests: [...selected] } : {}) },
     };
 }
+/** A retry base no runner invocation parses. evaluateManifestedTest builds its stranded single-fork
+ * retry from the base it is handed, and one it cannot parse throws before any spawn — so this base
+ * disables that inner recovery: a worker-RPC-stranded re-observation comes back infra (the caller
+ * parks it as ambiguous) instead of launching a second execution. */
+export const REOBSERVATION_RETRY_BASE = "tickmarkr-reobservation-refuses-stranded-retry";
+/** OBS-1106 residual: ONE isolated re-observation of a timeout-shaped red's attributed failing files on
+ * the same checkout, narrowed exactly as a screen is. Never cached and never a verdict: the caller
+ * keeps the original red and reads this only to decide whether that red is chargeable. Exactly one
+ * execution — the bounded infra/host-starved retries and the stranded single-fork recovery are all
+ * refused, so a diagnostic never buys more. */
+export async function reobserveTestFiles(worktree, testCmd, baseline, files, artifactDir) {
+    const cmd = testCommandForFiles(testCmd, files);
+    if (!isVitestTestCommand(testCmd, worktree))
+        return (await compareToBaseline(worktree, { test: cmd }, baseline, ["test"], { selected: files, authorizeRetry: () => false }))[0];
+    const r = await runVitestManifestGate(worktree, cmd, baseline, files, artifactDir, { retryBaseCommand: REOBSERVATION_RETRY_BASE });
+    // fail closed whatever a recovery did: a re-observation never reads a recovered verdict
+    return r.meta?.recovery === undefined ? r
+        : { ...r, pass: false, meta: { ...r.meta, classification: "infra", infra: true, retryable: false, recoveryRefused: true } };
+}
 const SIGNAL_EXIT_RE = /\b(?:SIGTERM|SIGKILL|signal\s+(?:9|15)|exit(?:s|ed|\s+code)?\s+(?:137|143))\b/i;
 const FAILURE_IDENTITY_RE = /\b(?:AssertionError|FAIL\s+\S|Tests?\s+\d+\s+failed|expected\s+.+\s+to\s+)\b/i;
 /** D1: apply the daemon's signal-only rider before either battery cache read or write. Its onGate
@@ -261,12 +319,56 @@ function classifySignalOnlyTest(g) {
         return;
     g.meta = { ...g.meta, classification: "infra", infra: true, retryable: false, kind: "signal-exit" };
 }
+/** OBS-1151: a criterion's comparable subject — its canonical text, the task's declared bounds and the
+ * operator context. The cited files' blobs are compared separately, over the union of both citations. */
+export function judgmentSubjectKey(task, criterion, operatorContext) {
+    return createHash("sha256").update(JSON.stringify([criterion, [...task.files].sort(), [...(task.outOfScope ?? [])].sort(), operatorContext ?? ""])).digest("hex");
+}
+async function blobAt(worktree, commit, path) {
+    const r = await shGit(`git rev-parse --verify --quiet ${shq(`${commit}:${path}`)}`, worktree);
+    return r.code === 0 && r.stdout.trim() ? r.stdout.trim() : undefined;
+}
+/** OBS-1151: the criteria whose fresh ruling reverses the newest prior ruling on the same subject key whose
+ * cited paths hold the identical blob at both commits (older comparable priors are still found behind a
+ * newer prior on different blobs). A citation-less side or an
+ * unreadable blob is unknown, never identical — that criterion's fresh ruling is simply fresh. */
+export async function judgeContradictions(worktree, head, fresh, priors) {
+    const found = [];
+    for (const c of fresh) {
+        if (!c.paths.length)
+            continue;
+        // C-3 (D-669): the comparable prior is the NEWEST one on identical cited blobs, not the newest one
+        // carrying the key — PASS(A) → FAIL(B) → fresh FAIL(A) must still be adjudicated against PASS(A).
+        for (const prior of priors) {
+            const was = prior.criteria.find((q) => q.key === c.key);
+            if (!was || !was.paths.length)
+                continue;
+            const paths = [...new Set([...was.paths, ...c.paths])].sort();
+            let identical = true;
+            for (const path of paths) {
+                const [before, now] = await Promise.all([blobAt(worktree, prior.commit, path), blobAt(worktree, head, path)]);
+                if (!before || !now || before !== now) {
+                    identical = false;
+                    break;
+                }
+            }
+            if (!identical)
+                continue;
+            if (was.met !== c.met)
+                found.push({ id: c.id, met: c.met, priorMet: was.met, priorCommit: prior.commit, paths });
+            break;
+        }
+    }
+    return found;
+}
 export async function runGates(task, ctx) {
     const results = [];
     const evidence = {
         artifactDir: ctx.artifactDir,
         runId: ctx.buildReceiptIdentity?.runId ?? ctx.artifactDir ?? "standalone",
         taskId: task.id, attempt: ctx.buildReceiptIdentity?.attempt ?? 0,
+        // OBS-1140: task and standalone gates honour the configured quota, not the built-in default.
+        quotaBytes: ctx.cfg.gates?.evidenceQuotaBytes,
         ...ctx.evidence,
     };
     // Receipt identity belongs to this round, never to a cached verdict. Each call from the shell
@@ -304,6 +406,9 @@ export async function runGates(task, ctx) {
         await receiptNotes;
     };
     let selectionDecision;
+    // OBS-635: the identity a full suite measured inside the battery, revalidated after semantics.
+    let fullInBattery = false;
+    let batteryFullIdentity;
     let commits = [];
     // Check before cache identity, npm policy probes, or any gate command.
     const dependencyRefusal = dependencyLinkRefusal(ctx.worktree);
@@ -337,14 +442,41 @@ export async function runGates(task, ctx) {
             payload: { gate, reason: "cached-red-discarded", ...(reason === "recheck" ? {} : { bypass: reason }) }, result: hit });
         return true;
     };
+    // OBS-1168(c): every judge/review seat this round opens, so a failed sibling can cancel the other.
+    // A cancelled round dispatches, re-routes and publishes nothing more; its closed seats' own errors
+    // are consequences of the cancel, never a second failure.
+    // A seat whose creation was still pending at the cancel is refused before dispatch: onSlot runs inside
+    // llm.ts's launch guard, so the throw closes (and awaits) that half-launched pane and never runs it.
+    const semanticSlots = new Set();
+    const closing = [];
+    const via = ctx.via && {
+        ...ctx.via, onSlot: (slot) => {
+            semanticSlots.add(slot);
+            ctx.via.onSlot?.(slot);
+            if (cancelled)
+                throw new Error("semantic round cancelled before dispatch");
+        },
+    };
+    let cancelled = false;
+    const cancelSemantic = () => {
+        if (cancelled)
+            return;
+        cancelled = true;
+        for (const slot of semanticSlots)
+            closing.push(via.driver.close(slot).catch(() => { }));
+    };
     const shapeGates = ctx.cfg.gates.byShape?.[task.shape];
     const enabled = (g) => task.gates.includes(g) && (g !== "acceptance" && g !== "review" || shapeGates?.[g] !== false);
     const failed = () => results.some((r) => !r.pass);
     // T4 (OBS-265): a GREEN selected-test run is a screen, not the round's verdict — the merge-candidate
-    // round re-runs the full suite on the same commit and THAT is what the round reports. Held here so
-    // exactly one `test` gate-result ever leaves a round, always carrying which suite spoke for it.
+    // round re-runs the full suite on the same commit and THAT is what the round reports. With no
+    // semantic gate to act on it, the screen is held so its full suite speaks for it in one row.
     // (A RED screen IS the verdict: the round ends there, so it is recorded immediately.)
     let heldTest;
+    // OBS-1176: when acceptance/review WILL act on a green screen, the screen is published before they
+    // start, as its own selected row. The full suite afterwards is a second invocation on its own row —
+    // it carries only its own receipts and interval, so it neither erases nor re-counts the screen.
+    const publishScreen = enabled("acceptance") || enabled("review");
     // v2.0 T2 (OBS-554): this round's per-gate measurement. Every interval a gate actually spends
     // executing is added HERE, at the call site that runs it, so a gate that runs twice (the test
     // gate's screen and its full suite) sums to its own cost and never to the span between them.
@@ -634,6 +766,28 @@ export async function runGates(task, ctx) {
             },
         };
     };
+    const fullTestIdentity = () => computeVerificationIdentity({
+        worktree: ctx.worktree,
+        gate: "test",
+        scope: ctx.verificationScope,
+        command: ctx.commands.test,
+        baseline: ctx.baseline,
+        selectedSet: undefined,
+        capacity: resolvedCapacity(),
+    });
+    // OBS-635: a full green answers only for the manifest it certified. The tree identity cannot see an
+    // ignored generated test the runner would collect, so every full-green reuse rediscovers the
+    // runner's listing and requires the verdict's to equal it; a runner without a listing is bound by
+    // its identity alone. Reads before the semantic gates share one listing; `listing` resets after them.
+    let listing;
+    const certifiesFullManifest = async (verdict) => {
+        if (!verdict.pass || !isVitestTestCommand(ctx.commands.test, ctx.worktree))
+            return true;
+        const certified = verdict.meta?.manifest;
+        const current = await (listing ??= listFullManifest(ctx.commands.test, ctx.worktree));
+        return Array.isArray(certified) && current !== undefined
+            && certified.length === current.length && [...certified].sort().every((file, i) => file === current[i]);
+    };
     // shell tools vs the shared baseline
     const retryOptions = (identity) => ctx.authorizeInfraRetry
         ? { authorizeRetry: (cause) => ctx.authorizeInfraRetry(identity ? verificationIdentityKey(identity) : "", cause === "infra" ? "infrastructure" : "host-starved") }
@@ -663,7 +817,7 @@ export async function runGates(task, ctx) {
                 if (hit)
                     classifySignalOnlyTest(hit); // Older entries predate classification at the write seam.
                 if (hit && identity && !isInfraResult(hit) && (hit.pass || (ctx.verificationScope ?? "battery") === "battery")
-                    && !(await discardCachedRed(g, hit))) {
+                    && !(await discardCachedRed(g, hit)) && (g !== "test" || selected !== undefined || await certifiesFullManifest(hit))) {
                     r = formatReusedRow(hit, identity);
                     cached = true;
                     if (g === "build")
@@ -690,6 +844,10 @@ export async function runGates(task, ctx) {
             // the screen's interval IS the test gate's first interval, so the split needs no second clock
             if (g === "test" && selected)
                 selectedDurationMs = spans.get("test")?.durationMs ?? 0;
+            if (g === "test" && !selected) {
+                fullInBattery = true;
+                batteryFullIdentity = identity;
+            }
             // The pre-battery check proves the tree clean ONCE; a command that exits 0 having rewritten a
             // tracked file makes it dirty again, and every gate after it — including the next shell gate,
             // which would then run against bytes HEAD does not hold — inherits that. So re-check after each
@@ -711,6 +869,13 @@ export async function runGates(task, ctx) {
                 const screened = { ...r, meta: { ...r.meta, selectedTests: selected } };
                 if (!screened.pass)
                     await record(screened);
+                else if (publishScreen) {
+                    await record(screened);
+                    // The screen's interval now lives on its own row; the full suite measures from zero.
+                    spans.delete("test");
+                    loadSamples.delete("test");
+                    selectedDurationMs = undefined;
+                }
                 else {
                     heldTest = withTelemetry(screened);
                     results.push(heldTest);
@@ -790,7 +955,9 @@ export async function runGates(task, ctx) {
         // v1.87 T2: the judge is a configured seat like any other — check it against the operator's
         // policy BEFORE spending a dispatch on it. disallowedBy carries the whole deny grammar (adapter,
         // model, or adapter:model), so a model-scoped deny cannot slip past an adapter-id-only read.
-        const judgeDenied = disallowedBy({ adapter: ctx.cfg.judge.adapter, model: ctx.cfg.judge.model }, ctx.cfg.routing, "judge");
+        // OBS-1186: under the exact cached identity of that channel, as compile, doctor and route read it.
+        const judgeSeat = observedSeat(ctx.health, ctx.cfg.judge.adapter, ctx.cfg.judge.model);
+        const judgeDenied = disallowedBy(judgeSeat, ctx.cfg.routing, "judge");
         if (judgeDenied) {
             return {
                 result: {
@@ -803,8 +970,8 @@ export async function runGates(task, ctx) {
             };
         }
         const judgeAdapter = getAdapter(ctx.cfg.judge.adapter, ctx.adapters);
-        const jvia = ctx.via
-            ? { driver: ctx.via.driver, keep: ctx.via.keep, onSlot: ctx.via.onSlot, name: ctx.via.nameFor("judge", judgeAdapter.id), label: ctx.via.labelFor("judge") }
+        const jvia = via
+            ? { driver: via.driver, keep: via.keep, onSlot: via.onSlot, name: via.nameFor("judge", judgeAdapter.id), label: via.labelFor("judge") }
             : undefined;
         // v1.19 (T2): testCmd threads the detected test runner to the gate so named-test oracles run
         // deterministically (filtered via -t) before any LLM judge dispatch.
@@ -815,8 +982,8 @@ export async function runGates(task, ctx) {
         // Separate from `invocations` above deliberately: that array is transcript evidence and records
         // one entry per CAPTURED OUTPUT, so a dispatch that produced none contributes nothing to it.
         const invocationSpans = [];
-        const invokeJudge = async (adapter, model, via) => {
-            const captured = await captureLlmDispatches([adapter], ([instrumented]) => acceptanceGate(task, ctx.worktree, ctx.baseRef, { adapter: instrumented, model }, via, { testCmd: ctx.commands.test, diffCap: ctx.cfg.gates.diffCap }));
+        const invokeJudge = async (adapter, model, via, effort) => {
+            const captured = await captureLlmDispatches([adapter], ([instrumented]) => acceptanceGate(task, ctx.worktree, ctx.baseRef, { adapter: instrumented, model, effort }, via, { testCmd: ctx.commands.test, diffCap: ctx.cfg.gates.diffCap }));
             // The instrumented adapter is reached only by runLlm. Deterministic oracles and diff-cap exits
             // never call headlessCommand, so they produce no clock and cannot manufacture an invocation.
             invocationSpans.push(...captured.invocations);
@@ -838,7 +1005,58 @@ export async function runGates(task, ctx) {
             }
             return captured.value;
         };
-        let a = await invokeJudge(judgeAdapter, ctx.cfg.judge.model, jvia);
+        const judgePool = () => (ctx.judgeChannels ?? []).filter((c) => disallowedBy(c, ctx.cfg.routing, "judge") === null);
+        const rankJudges = (pool) => [...pool]
+            .sort((x, y) => TIER_RANK[y.tier] - TIER_RANK[x.tier] || marginalCostRank(x) - marginalCostRank(y));
+        // OBS-1151 (+add.1): the fresh judgment always stands on its own reading — a prior PASS is never
+        // reused. Only a criterion that REVERSES the newest prior ruling on the same subject key over
+        // identical cited blobs needs a second, distinct judge; agreement stands (a sound FAIL included),
+        // and a split, no distinct eligible seat or an unreadable adjudication parks for the operator.
+        const adjudicate = async (fresh) => {
+            const primary = String(fresh.meta?.judge ?? channelKey({ adapter: ctx.cfg.judge.adapter, model: ctx.cfg.judge.model }));
+            const head = await shGit("git rev-parse HEAD", ctx.worktree);
+            const commit = head.code === 0 ? head.stdout.trim() : "";
+            const criteria = fresh.meta.judgment.map((c) => ({
+                id: c.id, key: judgmentSubjectKey(task, c.criterion, ctx.operatorContext), met: c.met, paths: c.paths,
+            }));
+            const record = { commit, judge: primary, criteria };
+            const stamped = { ...fresh, meta: { ...fresh.meta, judgment: record } };
+            if (!commit || !ctx.priorJudgments?.length)
+                return commit ? stamped : { ...fresh, meta: { ...fresh.meta, judgment: undefined } };
+            const disputed = await judgeContradictions(ctx.worktree, commit, criteria, ctx.priorJudgments);
+            if (!disputed.length)
+                return stamped;
+            await ctx.onGate?.({ phase: "note", gate: "acceptance", name: "judge-disagreement", payload: { primary, commit, disputed } });
+            const park = (why, adjudicator) => ({
+                gate: "acceptance", pass: false,
+                details: `judge disagreement on ${disputed.map((d) => d.id).join(", ")}: ${primary} reverses an earlier ruling over identical cited blobs (${[...new Set(disputed.flatMap((d) => d.paths))].join(", ")}) — ${why}; parked for an operator ruling, no worker charge`,
+                meta: { classification: "infra", infra: true, retryable: false, cause: "judge-disagreement", judge: primary,
+                    judgeDisagreement: { primary, ...(adjudicator ? { adjudicator } : {}), disputed, outcome: why } },
+            });
+            const primaryAdapter = primary.slice(0, primary.indexOf(":"));
+            // One DISTINCT seat: never the primary channel, a different adapter when the pool has one.
+            const pool = judgePool().filter((c) => channelKey(c) !== primary);
+            const seat = rankJudges(pool.filter((c) => c.adapter !== primaryAdapter))[0] ?? rankJudges(pool)[0];
+            if (!seat)
+                return park("no distinct eligible judge is available");
+            const adjudicator = channelKey(seat);
+            const seatAdapter = getAdapter(seat.adapter, ctx.adapters);
+            const seatVia = via
+                ? { driver: via.driver, keep: via.keep, onSlot: via.onSlot, name: via.nameFor("judge", seatAdapter.id) + "-r2", label: via.labelFor("judge") }
+                : undefined;
+            const second = await invokeJudge(seatAdapter, seat.model, seatVia, configuredEffort(ctx.cfg, seat));
+            const rulings = Array.isArray(second.meta?.judgment) ? second.meta.judgment : undefined;
+            // A citation-less ruling cannot be compared, so it confirms nothing (invented evidence is already unparseable,
+            // and an internally inconsistent verdict carries no judgment rows at all).
+            if (second.meta?.unparseable === true || !rulings)
+                return park("the adjudicating judge returned no readable verdict", adjudicator);
+            const agreed = disputed.every((d) => rulings.some((r) => r.id === d.id && r.met === d.met && r.paths.length > 0));
+            if (!agreed)
+                return park("the adjudicating judge split from the fresh ruling", adjudicator);
+            return { ...stamped, meta: { ...stamped.meta, adjudication: { primary, adjudicator, criteria: disputed.map((d) => d.id), agreed: true } } };
+        };
+        // OBS-1182: every judge seat launches at its OWN configured effort, never the worker's.
+        let a = await invokeJudge(judgeAdapter, ctx.cfg.judge.model, jvia, configuredEffort(ctx.cfg, ctx.cfg.judge));
         // GATE-09: an unparseable judge verdict retries the JUDGE exactly once on a failover channel — never
         // the worker (run-20260711-185020 P43-03 L70-72 billed a judge flake as a worker attempt). The flaked
         // first verdict NEVER enters results (no false gate-result journal event, no operator notify, no stale
@@ -852,7 +1070,7 @@ export async function runGates(task, ctx) {
         // If no other adapter is live, the exclusion degrades to a channel-level reroute within the same
         // adapter so a single-adapter fleet still retries (matching the daemon's unknown-excludeAdapter
         // degradation path).
-        if (a.meta?.unparseable === true && typeof a.meta.judge === "string") {
+        if (!cancelled && a.meta?.unparseable === true && typeof a.meta.judge === "string") {
             const flakedKey = a.meta.judge;
             const flakedAdapter = flakedKey.slice(0, flakedKey.indexOf(":"));
             const pick = (pool) => pool
@@ -866,17 +1084,23 @@ export async function runGates(task, ctx) {
             const sameAdapter = pick(judgePool.filter((c) => c.adapter === flakedAdapter && channelKey(c) !== flakedKey));
             // Prefer a different adapter; if the fleet only has one adapter, retry on a different channel of
             // that adapter; if the fleet has only one channel, fall back to the original judge config.
-            const retry = crossAdapter ?? sameAdapter ?? { adapter: ctx.cfg.judge.adapter, model: ctx.cfg.judge.model };
+            const retry = crossAdapter ?? sameAdapter ?? judgeSeat;
             const retryAdapter = getAdapter(retry.adapter, ctx.adapters);
-            const retryJvia = ctx.via
+            const retryJvia = via
                 // unconditional -r1 suffix: under keepPanes:forever a same-channel retry cannot collide with the
                 // still-open first pane (herdr agent_name_taken regression, research Pitfall 4)
-                ? { driver: ctx.via.driver, keep: ctx.via.keep, onSlot: ctx.via.onSlot, name: ctx.via.nameFor("judge", retryAdapter.id) + "-r1", label: ctx.via.labelFor("judge") }
+                ? { driver: via.driver, keep: via.keep, onSlot: via.onSlot, name: via.nameFor("judge", retryAdapter.id) + "-r1", label: via.labelFor("judge") }
                 : undefined;
             // the retry IS a second acceptanceGate call: one code path, one parser, zero new parse leniency.
-            a = await invokeJudge(retryAdapter, retry.model, retryJvia);
+            a = await invokeJudge(retryAdapter, retry.model, retryJvia, configuredEffort(ctx.cfg, retry));
             a = { ...a, meta: { ...a.meta, judgeRetry: { flaked: flakedKey, retried: channelKey({ adapter: retry.adapter, model: retry.model }) } } };
         }
+        // OBS-1168(b): the re-routed seat could not launch either — no seat produced a verdict, so this is
+        // an infra park over whatever the deterministic gates proved, never a charge against the worker.
+        if (a.meta?.cause === "seat-launch-failed")
+            a = { ...a, meta: { ...a.meta, classification: "infra", infra: true, retryable: false } };
+        else if (!cancelled && Array.isArray(a.meta?.judgment))
+            a = await adjudicate(a);
         // No dispatch, no key: a deterministic-oracle round writes no `invocations` field rather than an
         // empty array a reader could mistake for "measured, and it cost nothing".
         return { result: invocationSpans.length ? { ...a, meta: { ...a.meta, invocations: invocationSpans } } : a, invocations };
@@ -895,6 +1119,8 @@ export async function runGates(task, ctx) {
             const captured = await captureLlmDispatches(ctx.adapters, run);
             invocations.push(...captured.invocations);
             const rv = captured.value;
+            if (cancelled)
+                return rv; // a cancelled seat's non-answer says nothing about the seat
             if (rv.meta?.noVerdict === true || rv.meta?.unparseable === true) {
                 await ctx.onGate?.({ phase: "note", gate: "review", name: "review-no-verdict", payload: { ...rv.meta }, result: rv });
                 if (typeof rv.meta.reviewer === "string") {
@@ -925,7 +1151,7 @@ export async function runGates(task, ctx) {
         const priorReviewers = [...(ctx.priorReviewers ?? []), ...(ctx.excludeReviewers ?? [])];
         const carriedAuthors = ctx.carriedAuthors ?? [];
         let exclusions = [...(ctx.excludeReviewers ?? []), ...retired];
-        let rv = await dispatch((adapters) => reviewGate(task, ctx.worktree, ctx.baseRef, ctx.author, ctx.channels, adapters, ctx.cfg, ctx.via, exclusions, ctx.artifactDir, ctx.reviewHistory, ctx.demotedReviewers, ctx.carriedFindings, priorReviewers, carriedAuthors, ctx.operatorContext));
+        let rv = await dispatch((adapters) => reviewGate(task, ctx.worktree, ctx.baseRef, ctx.author, ctx.channels, adapters, ctx.cfg, via, exclusions, ctx.artifactDir, ctx.reviewHistory, ctx.demotedReviewers, ctx.carriedFindings, priorReviewers, carriedAuthors, ctx.operatorContext));
         // OBS-193/574: an unparseable review verdict retries the REVIEW, preferring a different adapter. Only
         // a single-adapter eligible pool may fall back to another channel on the flaked adapter. The flaked
         // verdict never enters results; an exhausted pool preserves its cause.
@@ -935,9 +1161,29 @@ export async function runGates(task, ctx) {
         let retryPrior = [...priorReviewers];
         const routes = [];
         let hop = 0;
-        while ((rv.meta?.unparseable === true || rv.meta?.noVerdict === true) && typeof rv.meta.reviewer === "string") {
+        // OBS-1196: seats already re-asked after a malformed verdict — once per seat, so never unbounded.
+        const reemitted = new Set();
+        while (!cancelled && (rv.meta?.unparseable === true || rv.meta?.noVerdict === true) && typeof rv.meta.reviewer === "string") {
             hop++;
             const flaked = rv.meta.reviewer;
+            const retryVia = via
+                ? { ...via, nameFor: (role, adapter) => via.nameFor(role, adapter) + `-r${hop}` }
+                : undefined;
+            // OBS-1196: a malformed (unparseable) verdict is a delivery defect of THIS seat, not a reason to drop it. Ask the
+            // same seat once more on the same subject; reviewGate mints a fresh nonce, so only a new, whole,
+            // nonce-bound verdict can answer — the malformed bytes are never salvaged into one.
+            if (rv.meta.cause === "malformed-verdict" && rv.meta.closureInvalid !== true && !reemitted.has(flaked)) {
+                reemitted.add(flaked);
+                const others = ctx.channels.map(channelKey).filter((key) => key !== flaked);
+                const again = await dispatch((adapters) => reviewGate(task, ctx.worktree, ctx.baseRef, ctx.author, ctx.channels, adapters, ctx.cfg, retryVia, others, ctx.artifactDir, ctx.reviewHistory, ctx.demotedReviewers, ctx.carriedFindings, retryPrior, carriedAuthors, ctx.operatorContext));
+                if (again.meta?.noEligibleReviewer !== true) {
+                    await ctx.onGate?.({ phase: "note", gate: "review", name: "review-reemission", payload: { reviewer: flaked, cause: "malformed-verdict",
+                            delivered: again.meta?.unparseable !== true && again.meta?.noVerdict !== true }, result: again });
+                    routes.push(`review re-emission (same seat, fresh nonce): ${flaked} produced a malformed verdict; asked once more`);
+                    rv = { ...again, details: `${routes.join("\n")}\n${again.details}`, meta: { ...again.meta, reviewReemission: { reviewer: flaked } } };
+                    continue;
+                }
+            }
             const emptyOutput = rv.meta.cause === "empty-output";
             if (emptyOutput) {
                 await ctx.onGate?.({
@@ -946,9 +1192,6 @@ export async function runGates(task, ctx) {
                     result: { ...rv, meta: { ...rv.meta, skipped: true } },
                 });
             }
-            const retryVia = ctx.via
-                ? { ...ctx.via, nameFor: (role, adapter) => ctx.via.nameFor(role, adapter) + `-r${hop}` }
-                : undefined;
             const flakedAdapter = flaked.slice(0, flaked.indexOf(":"));
             const adapterExclusions = ctx.channels.filter((c) => c.adapter === flakedAdapter).map(channelKey);
             // RF-1: the retry filters by the floor reviewGate resolves — author tier, task floor, review.floor
@@ -982,11 +1225,14 @@ export async function runGates(task, ctx) {
                 break;
             }
         }
-        if (rv.meta?.noVerdict === true) {
+        if (rv.meta?.noVerdict === true
+            || (rv.meta?.unparseable === true && rv.meta.closureInvalid !== true && typeof rv.meta.reviewer === "string")) {
             // Terminal: every eligible seat returned no verdict. The carried materials stay open — an infra
             // row is not a passing review — and the sibling judge result is untouched beside it.
+            // OBS-1196: an exhausted pool of UNDELIVERED (unparseable) verdicts is the same non-verdict: an infra
+            // park, never a worker retry. A delivered verdict that breaks the closure protocol keeps its path.
             const carried = (ctx.carriedFindings ?? []).filter((f) => f.class === "review:material").map((f) => f.fingerprint);
-            rv = { ...rv, meta: { ...rv.meta, classification: "infra", infra: true, carriedFindings: carried } };
+            rv = { ...rv, meta: { ...rv.meta, noVerdict: true, classification: "infra", infra: true, carriedFindings: carried } };
         }
         return invocations.length ? { ...rv, meta: { ...rv.meta, invocations } } : rv;
     };
@@ -1045,11 +1291,27 @@ export async function runGates(task, ctx) {
         else
             selected = [...new Set([...selected, ...required])].sort();
     }
-    if (ctx.selectionReason)
+    // OBS-635: a screen buys nothing when the full suite that must follow it already has a qualified
+    // green on this exact identity (read BEFORE any screen), or when the harness-measured screen costs
+    // at least 75 % of it — then the full suite runs in the battery instead. Unknown timing keeps the
+    // screen. A selected-only green never answers here: its identity names its selection.
+    let promotion;
+    if (selected) {
+        const hit = verdictStore.get(await fullTestIdentity());
+        const costRatio = screenCostRatio(ctx.baseline, selected);
+        if (hit?.pass === true && !isInfraResult(hit) && await certifiesFullManifest(hit))
+            promotion = { reason: "full-green-cache" };
+        else if (costRatio !== undefined && costRatio >= SCREEN_PROMOTION_RATIO)
+            promotion = { reason: "screen-cost-promoted", costRatio };
+        if (promotion)
+            selected = undefined;
+    }
+    if (ctx.selectionReason || promotion)
         selectionDecision = {
-            scope: selected ? "selected" : "full", reason: selected ? selectionReason
-                : selectionReason === "known-failing-files" ? "unsupported-selection-full-suite" : selectionReason,
+            scope: selected ? "selected" : "full", reason: promotion?.reason ?? (selected ? selectionReason
+                : selectionReason === "known-failing-files" ? "unsupported-selection-full-suite" : selectionReason),
             requiredFiles: [...(ctx.requiredRepairTests ?? [])],
+            ...(promotion?.costRatio !== undefined ? { costRatio: promotion.costRatio } : {}),
         };
     await runBattery(selected ? { ...ctx.commands, test: testCommandForFiles(ctx.commands.test, selected) } : ctx.commands, selected, enabled("test") ? ["test"] : []);
     if (failed())
@@ -1073,27 +1335,73 @@ export async function runGates(task, ctx) {
         // enough: an acceptance-first await withholds a completed review behind a slow/hung judge and a
         // process death can lose that already-earned verdict. The returned result is still sorted into
         // GATE_NAMES order by done(); the event stream truthfully records each independent completion.
-        const judged = judging?.then((outcome) => withJudgeInvocationEvidence(outcome.invocations, () => record(outcome.result)));
-        const reviewed = reviewing?.then((outcome) => record(outcome));
-        if (executionSignal()) {
-            // A cancelled sibling still owns a process until it unwinds; do not settle the task early.
-            const settled = await Promise.allSettled([judged, reviewed]);
-            const rejected = settled.find((result) => result.status === "rejected");
-            if (rejected?.status === "rejected")
-                throw rejected.reason;
-            executionSignal()?.throwIfAborted();
-        }
-        else
-            await Promise.all([judged, reviewed]);
+        // OBS-1168(c): a sibling that throws, or that ends seatless (no seat could launch, so the round can
+        // only park infra), cancels the other — its seats are closed — and the round still AWAITS it, with
+        // or without an execution policy, so no verdict of a settled round publishes after its engagement.
+        const seatless = (r) => r.meta?.cause === "seat-launch-failed" && r.meta?.infra === true;
+        let failure;
+        const fail = (reason) => {
+            if (!cancelled)
+                failure ??= { reason };
+            cancelSemantic();
+        };
+        const judged = judging?.then(async (outcome) => {
+            if (cancelled)
+                return;
+            await withJudgeInvocationEvidence(outcome.invocations, () => record(outcome.result));
+            if (seatless(outcome.result))
+                cancelSemantic();
+        }).catch(fail);
+        const reviewed = reviewing?.then(async (outcome) => {
+            if (cancelled)
+                return;
+            await record(outcome);
+            if (seatless(outcome))
+                cancelSemantic();
+        }).catch(fail);
+        // ponytail: a headless seat has no slot to close; it is awaited to its own timeout, never abandoned.
+        await Promise.all([judged, reviewed]);
+        await Promise.all(closing);
+        if (failure)
+            throw failure.reason;
+        executionSignal()?.throwIfAborted();
         if (failed())
             return done();
     }
+    // OBS-635: the semantic gates ran oracles and vendor CLIs in this worktree after an in-battery full
+    // suite spoke, and its green stands only for the identity it measured. Oracle dirt withdraws it; a
+    // changed or unmeasurable identity — tree, command, baseline, environment, dependency resolution,
+    // capacity, protocol, lifecycle, full manifest — buys a fresh merge-candidate suite below.
+    let rerunFull = false;
+    listing = undefined;
+    if (fullInBattery && ctx.commands.test !== undefined && (enabled("acceptance") || enabled("review"))) {
+        const dirt = await dirtyWorktree();
+        if (dirt) {
+            const refusal = withTelemetry(await dirtyRoundRefusal("test", dirt));
+            results[results.findIndex((r) => r.gate === "test")] = refusal;
+            await ctx.onGate?.({ phase: "end", gate: "test", result: refusal });
+            return done();
+        }
+        const now = await fullTestIdentity();
+        // D-598: an unmeasurable lifecycle on either side is not comparable to anything (VerdictStore R41
+        // refuses it); two `unknown`s hashing equal is not an unchanged identity, so the suite reruns.
+        const measurable = (id) => id?.envParts?.verification?.lifecycle !== "unknown";
+        rerunFull = !now || !batteryFullIdentity || !measurable(now) || !measurable(batteryFullIdentity)
+            || verificationIdentityKey(now) !== verificationIdentityKey(batteryFullIdentity)
+            || !(await certifiesFullManifest(results.find((r) => r.gate === "test")));
+        if (rerunFull) {
+            // The in-battery row keeps its own interval; the replacement measures from zero.
+            spans.delete("test");
+            loadSamples.delete("test");
+        }
+    }
     // The merge-candidate round: every other gate is green, so THIS round is the one that can merge —
     // the full suite runs on the exact gated commit before the pipeline reports green. Nothing merges
-    // on a subset (spec: "nothing merges without a complete green suite"). Its verdict SUPERSEDES the
-    // held screen rather than joining it: one `test` entry in the record, one `test` end event in the
-    // stream, and `fullSuite` says which suite spoke while `selectedTests` keeps what the screen ran.
-    if (selected) {
+    // on a subset (spec: "nothing merges without a complete green suite"). Its verdict replaces the
+    // screen's entry in the returned record (one `test` entry), and `fullSuite` says which suite spoke
+    // while `selectedTests` keeps what the screen ran. In the stream, a held screen is superseded (one
+    // `test` end event); a published screen keeps its own earlier event and this is the second.
+    if (selected || rerunFull) {
         await emitStart("test");
         // This is the last shell command a round can run — the judge's named-test oracle (acceptance.ts)
         // may have run one before it, and every gate between the battery and here reads commits only, so
@@ -1104,20 +1412,12 @@ export async function runGates(task, ctx) {
         let cached = false;
         let identity;
         if (ctx.commands.test !== undefined) {
-            identity = await computeVerificationIdentity({
-                worktree: ctx.worktree,
-                gate: "test",
-                scope: ctx.verificationScope,
-                command: ctx.commands.test,
-                baseline: ctx.baseline,
-                selectedSet: undefined,
-                capacity: resolvedCapacity(),
-            });
+            identity = await fullTestIdentity();
             const hit = verdictStore.get(identity);
             if (hit)
                 classifySignalOnlyTest(hit);
             if (hit && identity && !isInfraResult(hit) && (hit.pass || (ctx.verificationScope ?? "battery") === "battery")
-                && !(await discardCachedRed("test", hit))) {
+                && !(await discardCachedRed("test", hit)) && await certifiesFullManifest(hit)) {
                 full = formatReusedRow(hit, identity);
                 cached = true;
                 await noteReuse("test", full, identity);