tickmarkr 1.97.0 → 2.1.0

package/dist/run/daemon.js

@@ -8,26 +8,30 @@ import { classifyDeadChannel, NO_TRAILER_SUMMARY, trailerPattern, UNPARSEABLE_TR
 import { allAdapters, getAdapter, probeAll, readDoctor, rolePools } from "../adapters/registry.js";
 import { addUsage, channelKey, matchesTrustDialog, QUOTA_RE } from "../adapters/types.js";
 import { bannerShell, paneDispatchCommand } from "../brand.js";
+import { collateralHits } from "../compile/collateral.js";
 import { DEFAULT_DIFF_CAP, globalConfigDir, loadConfigWithMode, readOverlayFile, repoOverlayPath, } from "../config/config.js";
 import { DeliveryReadinessError } from "../drivers/herdr.js";
 import { herdrSealShellPrefix, SubprocessDriver } from "../drivers/subprocess.js";
 import { formatOwnedName } from "../drivers/types.js";
 import { captureBaseline, detectGateCommands, detectVacuousOracles } from "../gates/baseline.js";
 import { runGates } from "../gates/run-gates.js";
-import { addEvidence, attributeBlocked, blockedTasks, getTask, graphDefinitionHash, loadGraph, pendingTasks, readyTasks, saveGraph, setStatus } from "../graph/graph.js";
+import { addEvidence, attributeBlocked, blockedTasks, getTask, graphDefinitionHash, loadGraph, pendingTasks, readyTasks, saveGraph, setStatus, taskContentDigest } from "../graph/graph.js";
 import { GATE_NAMES } from "../graph/schema.js";
 import { augmentRetryBrief, consult, renderRetryGuidance } from "./consult.js";
 import { runEnvironment } from "./environment.js";
 import { cleanupRunWorktrees, gitHead, linkNodeModules, npmDependencyInstallCommand, npmDependencyManifestChanged, runWithForkBudget, sh, shGit, WORKTREE_LAYOUT_CONTRACT, worktreePath } from "./git.js";
 import { runInteractiveSeed } from "./interactive-seed.js";
-import { activeRetryBan, classifyTaskFailure, classifyWorkerResultCause, engagementComparable, GATE_FINGERPRINT_CAP, GATE_SATISFIED_RELEASE, identicalGateFailures, journaledFailureBrief, Journal, loadRoutingProfile, newRunId, normalizeGateFailure, pendingRepairFindings, phaseForGate, recordedTaskFailureKind, repairsSinceApproval, reviewRoundsSinceApproval, runHasEnded, structuredFindings, upheldFeedbackByTask } from "./journal.js";
+import { activeRetryBan, classifyTaskFailure, classifyWorkerResultCause, engagementComparable, formatPriorFindingEvidence, GATE_FINGERPRINT_CAP, GATE_SATISFIED_RELEASE, identicalGateFailures, journaledFailureBrief, Journal, loadRoutingProfile, newRunId, normalizeGateFailure, pendingRepairFindings, phaseForGate, readPriorRunEvidence, recordedTaskFailureKind, repairsSinceApproval, reviewRoundsSinceApproval, runHasEnded, structuredFindings, upheldFeedbackByTask } from "./journal.js";
 import { isDiffCapPark } from "../gates/review.js";
 import { acquireApprovalSerialization, acquireRunLock, releaseRunLock } from "./lock.js";
 import { ensureIntegration, integrationBranch, integrationHead, mergeTask, verifyIntegrationTip } from "./merge.js";
 import { nextChannel, route } from "../route/router.js";
 import { desiredPanes } from "./reconcile.js";
-import { NUDGEABLE_ADAPTERS, PANE_READ_ROWS, StallProgressTracker, stallSnapshotBannerRows } from "./stall.js";
+import { harvestCpuFlatWindowMs, NUDGEABLE_ADAPTERS, PANE_READ_ROWS, StallProgressTracker, stallSnapshotBannerRows, WorkerTreeCpuAccountant, } from "./stall.js";
 import { armSupervision } from "./supervision.js";
+// Compatibility exports for the daemon liveness tests and existing consumers. The implementation
+// lives in stall.ts so gate dispatch can depend on it without importing the daemon.
+export { harvestCpuFlatWindowMs, resetHarvestCpuFlatMsForTests, setHarvestCpuFlatMsForTests, workerTreeCpuMs, } from "./stall.js";
 const MODE_RANK = { "staff-led": 0, "risk-based": 1, "partner-led": 2 };
 // An override (flag/spec) re-resolves through loadConfigWithMode itself, via a synthesized repo overlay
 // carrying routing.mode — floors, explore, lints, and provenance all come from config.ts's preset
@@ -182,6 +186,10 @@ function narrowRepairBattery(failing) {
         return isOracleFailure(g) && g.meta?.unparseable !== true;
     return false;
 }
+// v2.0 T2 (OBS-554): the measurement keys run-gates stamps on a GateResult, lifted verbatim onto the
+// gate row. One list, one lift — both onGate sites record through the same helper.
+const GATE_TELEMETRY_KEYS = ["durationMs", "load1Start", "load1End", "selectedDurationMs", "fullDurationMs", "invocations"];
+const gateMeasurement = (meta = {}) => Object.fromEntries(GATE_TELEMETRY_KEYS.filter((k) => meta[k] !== undefined).map((k) => [k, meta[k]]));
 /**
  * T4 (OBS-265): the journal with the review objections a round did NOT hinge on removed. Judge and
  * review are now launched together, so a round can journal a failed review that the serial walk would
@@ -350,40 +358,6 @@ export function setHarvestSilentMsForTests(ms) {
 export function resetHarvestSilentMsForTests() {
     harvestSilentMs = HARVEST_SILENT_MS;
 }
-// A CPU delta needs two samples separated in WALL CLOCK, and the CPU clock is QUANTIZED: darwin's
-// `ps` prints hundredths ("0:00.03"), linux's prints whole seconds ("00:00:01"). Equality across a
-// window shorter than the quantum is not evidence of anything — a worker throttled to a low duty
-// cycle accrues less than one tick per sample and reads flat while genuinely working. So the flat
-// observation must span the LARGER of a floor and this many ticks of the clock actually in use:
-// crossing 30 ticks means the tree burned <1 tick in 30, i.e. under ~3% of one core. On a
-// hundredths host that is a 3s window; on a whole-second host it is 30s — still nothing against the
-// ~15m redispatch it replaces. Resolution is read off the sampled rows, never assumed.
-const HARVEST_CPU_FLAT_MS = 3_000;
-const HARVEST_CPU_FLAT_TICKS = 30;
-// Once the flat window opens, retain descendants often enough to observe brief tool processes that
-// can start and exit between the daemon's ordinary wait slices. This sampler exists only during an
-// eligible silence window; it is stopped on progress or as soon as the worker wait concludes.
-const HARVEST_CPU_ACCOUNTING_POLL_MS = 100;
-// T2 review (material): that 100ms cadence forks a shell plus `ps` ten times a second, and on a host
-// where `ps` is unsupported or denied (the managed-sandbox class) EVERY sample fails — tens of
-// thousands of processes per silent attempt, multiplied by daemon concurrency, for a probe that can
-// never conclude anything. Persistent failure is structural, not transient, so the sampler STOPS
-// after this many consecutive unreadable snapshots. It stays stopped for the silence window it was
-// started for: read() then reports no CPU, the triad refuses to conclude and journals the gap, and a
-// later window (after real progress clears the accountant) starts a fresh one that pays the same
-// bounded probe again.
-const HARVEST_CPU_UNMEASURABLE_SAMPLE_CAP = 20;
-let harvestCpuFlatMs;
-export function harvestCpuFlatWindowMs(resolutionMs) {
-    return harvestCpuFlatMs ?? Math.max(HARVEST_CPU_FLAT_MS, resolutionMs * HARVEST_CPU_FLAT_TICKS);
-}
-/** Test seam — pin the flat window so a probe case need not sit through a real one. */
-export function setHarvestCpuFlatMsForTests(ms) {
-    harvestCpuFlatMs = ms;
-}
-export function resetHarvestCpuFlatMsForTests() {
-    harvestCpuFlatMs = undefined;
-}
 // Once the silence gate is met the CPU probe owns the poll cadence: the trailer-wait slice is 30s,
 // so two samples would otherwise cost a minute of wall clock apiece. Below the gate the only rule
 // is not to sleep PAST it — at the shipped 5m gate that changes no slice a worker sees today.
@@ -509,167 +483,6 @@ export async function verifyIntegrationTipCached(intWt, commands, journal, opts
     }
     return tipFailed;
 }
-// `ps` CPU time: "[[dd-]hh:]mm:ss[.frac]" (darwin prints "0:00.03", linux "00:00:01", both print
-// "1-02:03:04" past a day). Anything else is a header or a row this parser must not guess at.
-// `frac` reports whether THIS host prints sub-second digits — the quantum the flat window is sized
-// against, measured rather than assumed (a darwin sample is 10ms, a linux one 1000ms).
-function parsePsCpu(raw) {
-    const m = /^(?:(\d+)-)?(?:(\d+):)?(\d+):(\d+(?:\.\d+)?)$/.exec(raw);
-    if (!m)
-        return undefined;
-    const ms = ((Number(m[1] ?? 0) * 24 + Number(m[2] ?? 0)) * 60 + Number(m[3])) * 60_000 + Math.round(Number(m[4]) * 1000);
-    return { ms, frac: m[4].includes(".") };
-}
-let linuxClockTickMs;
-function linuxProcessCpuMs(pid, cwd) {
-    if (!existsSync("/proc/self/stat"))
-        return Promise.resolve(undefined);
-    // shGit, not sh: the accountant samples this path at a 100ms cadence, and a LOGIN shell would
-    // re-run the operator's profile (nvm/pyenv/direnv side effects included) on every sample.
-    linuxClockTickMs ??= shGit("getconf CLK_TCK", cwd, 15_000).then((r) => {
-        const ticks = r.code === 0 ? Number(r.stdout.trim()) : Number.NaN;
-        return Number.isFinite(ticks) && ticks > 0 ? 1_000 / ticks : undefined;
-    });
-    return linuxClockTickMs.then((resolutionMs) => {
-        if (resolutionMs === undefined)
-            return undefined;
-        try {
-            // `/proc/<pid>/stat` fields 14-17 are user/system jiffies for the process and its waited-for
-            // children. The child totals retain tools that start and exit wholly between live-tree polls.
-            // Split after the LAST ')' because comm may contain spaces or parentheses; field 3 is rest[0].
-            const stat = readFileSync(`/proc/${pid}/stat`, "utf8");
-            const fields = stat.slice(stat.lastIndexOf(")") + 2).trim().split(/\s+/);
-            const ticks = Number(fields[11]) + Number(fields[12]) + Number(fields[13]) + Number(fields[14]);
-            return Number.isFinite(ticks) ? { ms: ticks * resolutionMs, resolutionMs } : undefined;
-        }
-        catch {
-            return undefined; // process exited between ps ancestry capture and the precise CPU read
-        }
-    });
-}
-// T2 (OBS-264): the triad's CPU leg. Every non-seeded process of an attempt descends from that
-// attempt's own dispatch script, whose path is unique — print, argv-interactive and resume launches
-// all start there. (interactive-seed is intentionally fail-open below because its adapter-owned
-// launch bypasses this script.) ONE `ps` snapshot finds the root and all current descendants:
-// the agent CLI is a CHILD of the script's shell, so the root's own TIME never moves while the CLI
-// thinks. `resolutionMs` is the sampled clock's quantum, which sizes the caller's flat window.
-// Returns 0 when nothing matches: a worker whose process tree is gone is the strongest possible
-// "not working". Returns undefined when the snapshot itself failed or parsed to nothing —
-// unmeasurable CPU is never evidence a worker stopped, and the caller refuses to conclude on it.
-async function workerTreeCpuSnapshot(marker, cwd) {
-    // shGit, not sh: same login-shell cost as the CLK_TCK probe above — `ps` needs no profile.
-    const snapshot = await shGit("ps -Awwo pid=,ppid=,time=,command=", cwd, 15_000);
-    if (snapshot.code !== 0)
-        return undefined;
-    const rows = [];
-    for (const line of snapshot.stdout.split("\n")) {
-        const m = /^\s*(\d+)\s+(\d+)\s+(\S+)\s+(.*)$/.exec(line);
-        if (!m)
-            continue;
-        const cpu = parsePsCpu(m[3]);
-        if (cpu !== undefined)
-            rows.push({ pid: m[1], ppid: m[2], cpuMs: cpu.ms, frac: cpu.frac, cmd: m[4] });
-    }
-    if (rows.length === 0)
-        return undefined;
-    const tree = new Set(rows.filter((p) => p.cmd.includes(marker)).map((p) => p.pid));
-    // `ps` output is not topologically ordered — relax the parent→child closure until it stops growing.
-    for (let grew = true; grew;) {
-        grew = false;
-        for (const p of rows) {
-            if (!tree.has(p.pid) && tree.has(p.ppid)) {
-                tree.add(p.pid);
-                grew = true;
-            }
-        }
-    }
-    const precise = new Map();
-    let preciseResolutionMs;
-    for (const p of rows) {
-        if (!tree.has(p.pid))
-            continue;
-        const cpu = await linuxProcessCpuMs(p.pid, cwd);
-        precise.set(p.pid, cpu?.ms ?? p.cpuMs);
-        if (cpu !== undefined)
-            preciseResolutionMs = cpu.resolutionMs;
-    }
-    // Even an empty worker tree needs the host's actual measurement quantum: on Linux the /proc
-    // jiffy clock remains available after the worker exits, while `ps time` only prints whole seconds.
-    if (preciseResolutionMs === undefined && existsSync("/proc/self/stat")) {
-        preciseResolutionMs = (await linuxProcessCpuMs(String(process.pid), cwd))?.resolutionMs;
-    }
-    return {
-        processes: precise,
-        resolutionMs: preciseResolutionMs ?? (rows.some((p) => p.frac) ? 10 : 1_000),
-    };
-}
-export async function workerTreeCpuMs(marker, cwd) {
-    const snapshot = await workerTreeCpuSnapshot(marker, cwd);
-    if (snapshot === undefined)
-        return undefined;
-    return {
-        ms: [...snapshot.processes.values()].reduce((sum, cpuMs) => sum + cpuMs, 0),
-        resolutionMs: snapshot.resolutionMs,
-    };
-}
-// Sparse live-tree totals forget a tool's CPU as soon as that tool exits. This attempt-local
-// accountant instead adds each observed process's CPU DELTA to a monotonic total and replaces only
-// the live-PID cursor on each sample. When a PID disappears, its contribution stays in `totalMs`;
-// if that PID is later reused, its fresh total is added from zero because it left `live` in between.
-class WorkerTreeCpuAccountant {
-    marker;
-    cwd;
-    active = false;
-    loop;
-    live = new Map();
-    totalMs = 0;
-    gaps = 0;
-    consecutiveGaps = 0;
-    latest;
-    constructor(marker, cwd) {
-        this.marker = marker;
-        this.cwd = cwd;
-    }
-    async sample() {
-        const snapshot = await workerTreeCpuSnapshot(this.marker, this.cwd);
-        if (snapshot === undefined) {
-            this.gaps++;
-            this.live.clear();
-            this.latest = undefined;
-            // Stop forking `ps` at 10Hz once the host has proved it cannot answer — see the cap's comment.
-            if (++this.consecutiveGaps >= HARVEST_CPU_UNMEASURABLE_SAMPLE_CAP)
-                this.active = false;
-            return;
-        }
-        this.consecutiveGaps = 0;
-        for (const [pid, cpuMs] of snapshot.processes) {
-            const prior = this.live.get(pid);
-            this.totalMs += prior === undefined || cpuMs < prior ? cpuMs : cpuMs - prior;
-        }
-        this.live = snapshot.processes;
-        this.latest = { ms: this.totalMs, resolutionMs: snapshot.resolutionMs };
-    }
-    async start() {
-        if (this.active)
-            return;
-        this.active = true;
-        await this.sample();
-        this.loop = (async () => {
-            while (this.active) {
-                await new Promise((resolve) => setTimeout(resolve, HARVEST_CPU_ACCOUNTING_POLL_MS));
-                if (this.active)
-                    await this.sample();
-            }
-        })();
-    }
-    read() {
-        return { cpu: this.latest, gaps: this.gaps };
-    }
-    async stop() {
-        this.active = false;
-        await this.loop;
-    }
-}
 async function commitsAheadOf(base, wt) {
     const head = await gitHead(wt);
     if (head === base)
@@ -925,6 +738,11 @@ export async function runDaemon(repoRoot, opts = {}) {
     let releaseApprovalSerialization;
     try {
         let graph = loadGraph(repoRoot);
+        // One bounded snapshot supplies every dispatch. On resume the current journal still participates
+        // in the fold so its task-done/ordinary-approval events can retire older evidence, but findings it
+        // produced are suppressed: same-run feedback already replays those bytes and must not deliver them
+        // twice. Evidence from genuinely earlier runs remains available to a resumed fresh dispatch.
+        const priorRunEvidence = readPriorRunEvidence(repoRoot, graph.tasks, { suppressRunId: runId });
         // GATE-FIX-4 defect 4 (no-op run refusal): a fresh run on a graph with nothing dispatchable used
         // to journal {run-start, run-end} with zero dispatches — and downstream readers (greenness exit,
         // status, notify) treat that run-end as completion, so an all-terminal graph "went green" having
@@ -1159,6 +977,19 @@ export async function runDaemon(repoRoot, opts = {}) {
                 for (const w of await detectVacuousOracles(repoRoot, graph.tasks))
                     journal.append("baseline-warning", w.taskId, { ...w });
             }
+            // OBS-547: ONE full per-task collateral prediction for the whole run, computed here — uncapped, and
+            // handed whole to every scope gate below (never recomputed at a red, never the plan's 20-item view).
+            // Persisted beside baseline.json and RELOADED on resume (same reason baseline is): the prediction is
+            // pre-dispatch state, and rescanning a repository the run has since edited would let an offender flip
+            // between predicted and missed across stop/amend/resume. Missing file (pre-OBS-547 journal) ⇒ scan and
+            // pin it now, so every later resume of this run agrees with this one.
+            const collateralPath = join(journal.dir, "collateral.json");
+            const pinnedCollateral = opts.resume && existsSync(collateralPath)
+                ? new Map(Object.entries(JSON.parse(readFileSync(collateralPath, "utf8"))))
+                : null;
+            const collateral = pinnedCollateral ?? collateralHits(graph.tasks, repoRoot);
+            if (!pinnedCollateral)
+                writeFileSync(collateralPath, JSON.stringify(Object.fromEntries(collateral), null, 2));
             // T6: open the narrator AFTER run-start/run-resume is journaled so the watch surface has a run to
             // show. driver.narrator is undefined on subprocess → no-op (subprocess spawns nothing). Swallowed:
             // a failed-to-open or later-dead watch pane never affects the run.
@@ -1249,11 +1080,60 @@ export async function runDaemon(repoRoot, opts = {}) {
                 saveGraph(repoRoot, graph);
                 journal.append("task-human", t.id, { reason, kind });
                 if (assignment) {
-                    journal.telemetry({ taskId: t.id, shape: t.shape, adapter: assignment.adapter, model: assignment.model, channel: assignment.channel, attempts, outcome: "human", durationMs: Date.now() - startMs, parkKind: kind, gateFails, consults, tokens, meteredAttempts: tokens ? metered : undefined, retryMode });
+                    // OBS-547: `metered` counts CHARGEABLE metered attempts, so an unchargeable dispatch passes 0 and
+                    // the count is omitted rather than written as 0 or as `1` beside `attempts: 0` — a row claiming
+                    // more metered attempts than it charges reports as "floor: 1/0 attempts metered". tokens stay
+                    // (the spend was real); tokens-without-a-count is the already-modelled degraded ⇒ floor row.
+                    journal.telemetry({ taskId: t.id, shape: t.shape, adapter: assignment.adapter, model: assignment.model, channel: assignment.channel, attempts, outcome: "human", durationMs: Date.now() - startMs, parkKind: kind, gateFails, consults, tokens, meteredAttempts: tokens && metered ? metered : undefined, retryMode });
                 }
                 await reconcile({ spareLiveLlm: true }); // task-human is a terminal event — sweep, sparing sibling tasks' live LLM panes
                 await driver.notify(`tickmarkr ${runId}: ${t.id} needs a human — ${reason}`, { tier: "attention" });
             };
+            // OBS-547: cross-reference a scope red against the prediction this run already computed. Every hard
+            // offender predicted ⇒ an AUTHORING defect whose repair is pre-written: journal the classification
+            // with the verbatim files[] lines and park unchargeable. It owes nothing to the quality machinery —
+            // no chargeable attempt, no gateFails, no escalation, no ladder rung — and `scope-authoring` is what
+            // makes a resume replay agree (journal.ts replayResumeState). One unpredicted offender keeps today's
+            // chargeable behaviour and records the miss, so the lint's blind spots accumulate as evidence rather
+            // than folklore. ONE disposition for BOTH paths that observe a red — the ordinary attempt and the
+            // resume gate replay: which code path noticed the red must never decide who pays for it. Returns
+            // true when it parked; the caller then returns without charging anything.
+            // `attempts` is the count of CHARGEABLE attempts before this dispatch — never this dispatch's own
+            // count. The two callers arrive at it differently (the attempt loop's index already excludes the
+            // dispatch in flight; a gate replay's rs.attempts already includes it), and the ordinal journaled
+            // below is derived from it here so both paths number the same dispatch identically.
+            const dispositionScopeRed = async (t, results, assignment, attempts, startMs, gateFails, consults, tokens, metered, retryMode) => {
+                const scopeRed = results.find((g) => g.gate === "scope" && gateFailed(g));
+                const verdict = scopeRed?.meta?.collateral;
+                if (!verdict)
+                    return false;
+                if (verdict.authoring) {
+                    journal.append("scope-authoring", t.id, {
+                        gate: "scope",
+                        predicted: verdict.predicted,
+                        repair: verdict.repair,
+                        attempt: attempts + 1,
+                        chargeable: false,
+                        // The dispatch was PHYSICALLY metered even though nobody is charged for it. Physical metering
+                        // is recorded here, separately, so the telemetry row can stay chargeable-consistent: passing
+                        // this count on would print `meteredAttempts: 1` beside `attempts: 0`.
+                        ...(tokens ? { tokens, meteredDispatches: metered } : {}),
+                        ...(assignment ? { channel: channelKey(assignment) } : {}),
+                    });
+                    await driver.notify(`tickmarkr ${runId}: ${t.id} scope red was PREDICTED — authoring defect, no attempt charged`, { tier: "attention" });
+                    await park(t, `scope: every out-of-scope path was predicted by the collateral lint before dispatch — authoring defect, not a worker failure. Repair:\n${verdict.repair}`, "authoring", assignment, attempts, startMs, gateFails, consults, tokens, 0, retryMode);
+                    return true;
+                }
+                if (verdict.missed.length) {
+                    journal.append("collateral-miss", t.id, {
+                        gate: "scope",
+                        unpredicted: verdict.missed,
+                        predicted: verdict.predicted,
+                        attempt: attempts + 1,
+                    });
+                }
+                return false;
+            };
             const execTask = async (t) => {
                 const startMs = Date.now();
                 const taskTimeoutMinutes = t.timeoutMinutes ?? cfg.taskTimeoutMinutes;
@@ -1283,6 +1163,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                 // fresh-budget release, prefer nextChannel over the surviving tried-list so burned channels are
                 // not re-tried first (consult bans / prior failovers survive the release).
                 const rs = resume.get(t.id);
+                const contentDigest = taskContentDigest(t);
                 if (rs?.lastAssignment && rs.attempts > 0
                     && channels.some((c) => channelKey(c) === channelKey(rs.lastAssignment))
                     && !demotedChannels.has(channelKey(rs.lastAssignment))) {
@@ -1376,7 +1257,19 @@ export async function runDaemon(repoRoot, opts = {}) {
                         ...(g.meta?.fullSuite === true ? { fullSuite: true } : {}),
                         // A finding's path is its own evidence path. Do not pass task scope here: a declaration says
                         // where work is allowed, not where this verdict found the defect.
-                        ...(blocking ? { findings: structuredFindings(g.gate, g.details) } : {}),
+                        ...(blocking ? {
+                            taskContentDigest: contentDigest,
+                            findings: structuredFindings(g.gate, g.details),
+                        } : {}),
+                        // v2.0 T2 (OBS-554): the gate's OWN measurement, lifted verbatim from the meta run-gates
+                        // stamped WHERE THE GATE RAN. Nothing here re-derives a duration by subtracting journal
+                        // timestamps — that would measure this row's queue as well as its work. It rides the
+                        // gate-result row itself because that IS the row a recalibration reads: a measurement kept
+                        // in a side stream is one join away from the verdict it explains, and the two can drift.
+                        // A gate run-gates did not measure contributes NO field rather than a fabricated zero — for
+                        // every recalibration this telemetry funds, a gap is honest and a zero is a lie. The
+                        // seven-gate closed set is asserted end-to-end in tests/run/gate-telemetry.test.ts.
+                        ...gateMeasurement(g.meta),
                     });
                 };
                 // R3 (OBS-186): judge ‖ review are launched together and publish in COMPLETION order
@@ -1418,9 +1311,18 @@ export async function runDaemon(repoRoot, opts = {}) {
                 // in resume state. The journal already holds the upheld review's bytes; no reset of attempt or
                 // channel state can take them away, on any path, including `resume --retry-failed`.
                 const upheldFeedback = upheldFeedbackByTask(journal.read()).get(t.id) ?? rs?.upheldFeedback;
-                let feedback = upheldFeedback
+                const carriedEvidence = priorRunEvidence.findings
+                    .filter((finding) => finding.taskId === t.id)
+                    .map(formatPriorFindingEvidence)
+                    .join("\n\n");
+                const withCarriedEvidence = (brief) => {
+                    if (!carriedEvidence || brief.includes(carriedEvidence))
+                        return brief;
+                    return brief ? `${brief}\n\n${carriedEvidence}` : carriedEvidence;
+                };
+                let feedback = withCarriedEvidence(upheldFeedback
                     ? `The operator UPHELD the reviewer's findings — address them without discarding landed work.\nreview: ${upheldFeedback}`
-                    : "";
+                    : "");
                 let ladderIdx = 0;
                 let modeFallbackNoted = false; // v1.2: journal the interactive→print fallback once per task, not per attempt
                 let gateFails = 0; // TEL-02: incremented ONLY where feedback is built from failing gates — never derived from attempts (quota failovers bump attempts too, Pitfall 6)
@@ -1677,6 +1579,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                         const { results } = await runGates(resumedTask, {
                             worktree: wt, baseRef: taskBase, result: priorResult, author: gateAuthor,
                             commands, baseline, channels: pools.review, judgeChannels: pools.judge, adapters, cfg, artifactDir: journal.dir,
+                            collateral: collateral.get(t.id) ?? [],
                             // a recheck re-verifies a human's release: it never selects tests down, it runs the suite.
                             pipeline: "v185",
                             via: cfg.visibility.llm === "pane"
@@ -1710,6 +1613,20 @@ export async function runDaemon(repoRoot, opts = {}) {
                         graph = addEvidence(graph, t.id, { commits: approvedCommits, gateResults: results });
                         saveGraph(repoRoot, graph);
                         if (!results.every(gateSatisfied)) {
+                            // OBS-547: disposition FIRST, and through the same helper the ordinary attempt path uses. A
+                            // crash between a journaled scope red and its classification must not turn a predicted red
+                            // into a charged one just because a resume is what observed it.
+                            // rs.attempts COUNTS the interrupted dispatch this replay is judging; the chargeable
+                            // attempts behind it are one fewer. Passing the count itself would number the same dispatch
+                            // one higher here than on the ordinary path and bill an attempt the fresh path forgives.
+                            // UNLESS an earlier resume already journaled the classification and died before its park:
+                            // replayResumeState() has then ALREADY taken that dispatch back, so subtracting again would
+                            // erase an EARLIER chargeable attempt and attribute the park to the assignment the rewind
+                            // restored. Ask the journal which state this is, and take the classified dispatch's own
+                            // assignment back with it.
+                            const classified = journal.classifiedDispatch(t.id);
+                            if (await dispositionScopeRed(t, results, classified?.assignment ?? gateAuthor, classified ? (rs?.attempts ?? 0) : Math.max(0, (rs?.attempts ?? 0) - 1), startMs, gateFails, consults, tokens, metered, retryMode))
+                                return;
                             gateFails++;
                             // Observed green gates are only measurements. If the resumed suffix is red, preserve that
                             // result in the journal and return to the ordinary attempt/consult ladder, which rebuilds
@@ -1734,7 +1651,9 @@ export async function runDaemon(repoRoot, opts = {}) {
                         }
                         graph = setStatus(graph, t.id, "done");
                         saveGraph(repoRoot, graph);
-                        journal.append("task-done", t.id, { attempts: rs?.attempts ?? 0, assignment: gateAuthor });
+                        journal.append("task-done", t.id, {
+                            attempts: rs?.attempts ?? 0, assignment: gateAuthor, taskContentDigest: contentDigest,
+                        });
                         journal.append("merge", t.id, { branch: taskBranch, commit: await integrationHead(intWt) });
                         journal.telemetry({
                             taskId: t.id, shape: t.shape, adapter: gateAuthor.adapter, model: gateAuthor.model,
@@ -1956,7 +1875,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                     const promptFile = writePrompt(journal.dir, t, attempt, feedback, nonce);
                     // OBS-56: state the non-interactive, one-pass finish contract and the OBS-54 stall budget in every
                     // worker prompt, not only consult retry guidance. Prepended so prompt.ts's completion trailer stays last.
-                    const workerContract = `## Harness contract\n- This harness is non-interactive: make one continuous pass; do not stop for questions or follow-up input.\n- You have a ${taskTimeoutMinutes} minute stall window. Budget the full suite once, then commit and emit the completion trailer before it expires.\n- Each test: acceptance criterion must exist as a vitest test whose OWN title (the leaf, not counting enclosing describe titles) is the criterion string verbatim — never shortened, never decorated. Nesting under describe() is allowed.`; // OBS-64; OBS-511: leaf-title rule stated where the worker reads it
+                    const workerContract = `## Harness contract\n- This harness is non-interactive: make one continuous pass; do not stop for questions or follow-up input.\n- You have a ${taskTimeoutMinutes} minute stall window. The gates run the full suite for you — never spend this window on one; commit your work and emit the completion trailer inside it.\n- Each test: acceptance criterion must exist as a vitest test whose OWN title (the leaf, not counting enclosing describe titles) is the criterion string verbatim — never shortened, never decorated. Nesting under describe() is allowed.`; // OBS-64; OBS-511: leaf-title rule stated where the worker reads it; OBS-548: the suite is the GATES' job — this repo's own suite outlasts the output-silence windows the daemon polices, so a worker obeying "budget the full suite" was killed by construction
                     // OBS-47: state the worktree layout contract in the worker prompt (cheap-tier workers were
                     // committing/deleting node_modules and tripping the scope gate). The harness re-asserts the link
                     // itself before gates regardless of what the worker does with it.
@@ -2023,30 +1942,34 @@ export async function runDaemon(repoRoot, opts = {}) {
                         unmeasurableNoted = true;
                         journal.append("worker-harvest-unmeasurable", t.id, { slot: slot.name, attempt, reason });
                     };
-                    const harvestConcludes = async (silentMs) => {
-                        if (silentMs < harvestSilentMs) {
+                    // The CPU leg needs a marker in the worker's own argv, and every launch path puts this
+                    // attempt's dispatch script there EXCEPT interactiveSeed: runInteractiveSeed launches the
+                    // TUI directly, by a command the ADAPTER owns (seed.launch(model)) which tickmarkr cannot
+                    // make attempt-unique and must deliver verbatim. A marker that matches nothing reads as
+                    // zero CPU — precisely the false "flat" that would harvest a worker mid-turn — so a seeded
+                    // attempt has no measurable CPU leg and neither reader ever concludes it. The other half of
+                    // OBS-264 is untouched there: when its window does expire with commits on the worktree, the
+                    // no-trailer tail gates them instead of buying a fresh worker to re-produce them.
+                    const armCpuLeg = async (armed) => {
+                        if (!armed) {
                             await cpuAccountant?.stop();
                             cpuAccountant = undefined;
                             cpuFlat = undefined;
                             cpuGapCount = 0;
-                            return false;
+                            return;
                         }
-                        // The CPU leg needs a marker in the worker's own argv, and every launch path puts this
-                        // attempt's dispatch script there EXCEPT interactiveSeed: runInteractiveSeed launches the
-                        // TUI directly, by a command the ADAPTER owns (seed.launch(model)) which tickmarkr cannot
-                        // make attempt-unique and must deliver verbatim. A marker that matches nothing reads as
-                        // zero CPU — precisely the false "flat" that would harvest a worker mid-turn — so a seeded
-                        // attempt has no measurable CPU leg and the triad never concludes it. The other half of
-                        // OBS-264 is untouched there: when its window does expire with commits on the worktree, the
-                        // no-trailer tail gates them instead of buying a fresh worker to re-produce them.
+                        if (hasSeed || cpuAccountant !== undefined)
+                            return;
+                        cpuAccountant = new WorkerTreeCpuAccountant(dispatchScript, wt);
+                        await cpuAccountant.start();
+                    };
+                    const readCpuLeg = () => {
                         if (hasSeed) {
                             noteUnmeasurable("interactive-seed launch is not in the probed process tree");
-                            return false;
-                        }
-                        if (cpuAccountant === undefined) {
-                            cpuAccountant = new WorkerTreeCpuAccountant(dispatchScript, wt);
-                            await cpuAccountant.start();
+                            return { state: "unmeasurable" };
                         }
+                        if (cpuAccountant === undefined)
+                            return { state: "unmeasurable" }; // no window open: nothing measured yet
                         const observation = cpuAccountant.read();
                         if (observation.gaps !== cpuGapCount) {
                             cpuGapCount = observation.gaps;
@@ -2060,21 +1983,29 @@ export async function runDaemon(repoRoot, opts = {}) {
                             // structural hole as the seeded launch, and must not be the one that stays silent.
                             cpuFlat = undefined;
                             noteUnmeasurable("the worker process snapshot could not be read");
-                            return false;
+                            return { state: "unmeasurable" };
                         }
                         const now = Date.now();
                         if (cpu.ms !== cpuFlat?.ms) {
                             cpuFlat = { ms: cpu.ms, since: now };
-                            return false;
+                            return { state: "accruing" };
                         }
-                        if (now - cpuFlat.since < harvestCpuFlatWindowMs(cpu.resolutionMs))
+                        return now - cpuFlat.since < harvestCpuFlatWindowMs(cpu.resolutionMs)
+                            ? { state: "accruing" }
+                            : { state: "flat", cpu };
+                    };
+                    const harvestConcludes = async (silentMs) => {
+                        if (silentMs < harvestSilentMs)
+                            return false;
+                        const leg = readCpuLeg();
+                        if (leg.state !== "flat")
                             return false;
                         const carried = await commitsAheadOf(taskBase, wt);
                         if (carried.length === 0)
                             return false; // nothing landed: not this branch's population
                         journal.append("worker-harvest", t.id, {
                             slot: slot.name, attempt, commits: carried.length,
-                            silentMs, cpuMs: cpu.ms, cpuResolutionMs: cpu.resolutionMs,
+                            silentMs, cpuMs: leg.cpu.ms, cpuResolutionMs: leg.cpu.resolutionMs,
                         });
                         return true;
                     };
@@ -2117,6 +2048,12 @@ export async function runDaemon(repoRoot, opts = {}) {
                     let output;
                     let exitCode;
                     let timedOut = false;
+                    // OBS-548 addendum: the mechanism that fired is what the repair brief and the consult read.
+                    // A fast-kill lands with the rolling window nowhere near expiry (measured: 693 s of silence
+                    // inside a 1,800,000 ms window), so the trailer-less tail's stall-timeout fallback named a
+                    // window that could not have killed anything and briefed the next worker to fix a stall that
+                    // never happened. Fourth conflation in the taxonomy OBS-53 opened.
+                    let deadChannelKilled = false;
                     // T2 review: print mode's "the exit marker appeared". Kept apart from `finished` (the
                     // trailer) but still needed by the keepPanes decision below, whose contract is about a
                     // subprocess tree that REACHED its exit marker, not about what the worker claimed.
@@ -2152,7 +2089,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                             keptSlots.push(slot);
                         else
                             await closeSlot(slot);
-                        feedback = `delivery readiness failed after ${error.waitedMs}ms; pane transcript:\n${error.transcript}`;
+                        feedback = withCarriedEvidence(`delivery readiness failed after ${error.waitedMs}ms; pane transcript:\n${error.transcript}`);
                         const step = r.ladder[Math.min(ladderIdx++, r.ladder.length - 1)];
                         journal.append("escalation", t.id, { step, attempt: attempt + 1 });
                         await driver.notify(`tickmarkr ${runId}: ${t.id} escalation: ${step}`, { tier: "attention" });
@@ -2260,6 +2197,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                                 // lastProgressAt — see the kill below.
                                 let quotaStreak = 0;
                                 let rowSaturationHeld = false; // journaled once per attempt when the kill stands down
+                                let cpuHeld = false; // likewise for the CPU leg's stand-down (OBS-548)
                                 while (Date.now() - lastProgressAt < stallWindowMs) {
                                     const sliceStart = Date.now();
                                     const remaining = stallWindowMs - (sliceStart - lastProgressAt);
@@ -2396,6 +2334,30 @@ export async function runDaemon(repoRoot, opts = {}) {
                                     // daemon has nothing left to do — the pane falls back under the fast-kill and page
                                     // watchdogs like any other, instead of riding the whole rolling window untended.
                                     const nudgePending = nudgeable && (!nudged || nudgeDeadline !== undefined);
+                                    // T1 review fix: the kill's "no output growth" leg clocks off the RAW growth signals,
+                                    // never lastProgressAt alone — the flat-token rule (stall.ts) deliberately suppresses
+                                    // the re-arm report on row growth once tokens stick, and contextTokens is sticky across
+                                    // read misses, so a metered non-nudgeable adapter (pi) streaming rows under a stale
+                                    // counter presented a frozen lastProgressAt and was killed mid-work. lastRowGrowthAt is
+                                    // recorded on every high-water advance, suppressed or not; token growth already rides
+                                    // lastProgressAt. Either one advancing is output growth.
+                                    const lastOutputGrowthAt = Math.max(stallProgress.lastRowGrowthAt ?? 0, lastProgressAt);
+                                    // Everything the fast-kill can decide from the CHANNEL, evaluated before the CPU leg is
+                                    // asked for anything: the kill's own probe costs a `ps` every 100ms, so an ineligible
+                                    // slice must not pay for it. The CPU reading is consulted at the kill itself, below.
+                                    const fastKillEligible = !stallProgress.rowSignalSaturated
+                                        && !nudgePending && !nudgeFailed
+                                        && sliceNow - lastOutputGrowthAt >= deadChannelFastKillMs
+                                        && worktreeSinceLaunch === "unchanged";
+                                    const harvestCpuEligible = !nudgePending && !nudgeFailed
+                                        && sliceNow - lastProgressAt >= harvestSilentMs;
+                                    // OBS-548: ONE accountant serves both readers, so its lifecycle is decided once per
+                                    // slice from BOTH windows. Arming per-reader let the harvest's own "not silent enough
+                                    // yet" branch stop the sampler the fast-kill had just started, and a restarted
+                                    // accountant re-bases its monotonic total — the leg could then never read flat. The
+                                    // accountant also carries the readers' nudge holds: when neither reader may conclude,
+                                    // there is no reason to fork its 10 Hz process snapshots through the remaining window.
+                                    await armCpuLeg(harvestCpuEligible || fastKillEligible);
                                     // T2 (OBS-264): the liveness triad CONCLUDES the wait on finished work — commits ahead
                                     // of the task base (this attempt's own AND any carried forward — see the eligibility
                                     // comment at the dispatch site), a flat worker-tree CPU delta, and >= harvestSilentMs
@@ -2408,10 +2370,15 @@ export async function runDaemon(repoRoot, opts = {}) {
                                     // claude-code worker before the rescue could fire, leaving T1's nudge dead code for
                                     // exactly the committed-and-stalled population OBS-264 is about. Holding concludes at
                                     // ~14m (nudge + grace) rather than ~36m — nearly all of the OBS-264 win, and a worker
-                                    // that only needed a submit answers with a full trailer instead of partial work. An
-                                    // ANSWERED or twice-undeliverable nudge leaves nothing pending, so the triad governs
-                                    // again; the hold is on a pending daemon ACTION, never on the adapter being nudgeable.
-                                    if ((!nudgePending || nudgeFailed) && await harvestConcludes(sliceNow - lastProgressAt))
+                                    // that only needed a submit answers with a full trailer instead of partial work.
+                                    // An ANSWERED nudge leaves nothing pending, so the triad governs again; the hold is on a
+                                    // pending daemon ACTION, never on the adapter being nudgeable.
+                                    // OBS-548: an UNDELIVERABLE nudge holds too. A worker inside one long foreground command
+                                    // has no input box, so it is the population that CANNOT be nudged and is also the one
+                                    // most likely to be legitimately silent — letting its delivery failure lift the hold made
+                                    // the failure itself the trigger. `:357-359` already reads a false driver.nudge return as
+                                    // a delivery outcome, "not proof of an unreachable channel"; unreachable is not at rest.
+                                    if (!nudgePending && !nudgeFailed && await harvestConcludes(sliceNow - lastProgressAt))
                                         break;
                                     // T1 (R1 dead-channel fast-kill): no trailer, an unchanged launch tree, and no output growth
                                     // for the fast-kill window — the channel is dead, so conclude NOW
@@ -2443,20 +2410,30 @@ export async function runDaemon(repoRoot, opts = {}) {
                                         rowSaturationHeld = true;
                                         journal.append("worker-dead-held", t.id, { slot: slot.name, attempt, reason: "row-signal-saturated" });
                                     }
-                                    // T1 review fix: the kill's "no output growth" leg clocks off the RAW growth signals,
-                                    // never lastProgressAt alone — the flat-token rule (stall.ts) deliberately suppresses
-                                    // the re-arm report on row growth once tokens stick, and contextTokens is sticky across
-                                    // read misses, so a metered non-nudgeable adapter (pi) streaming rows under a stale
-                                    // counter presented a frozen lastProgressAt and was killed mid-work. lastRowGrowthAt is
-                                    // recorded on every high-water advance, suppressed or not; token growth already rides
-                                    // lastProgressAt. Either one advancing is output growth.
-                                    const lastOutputGrowthAt = Math.max(stallProgress.lastRowGrowthAt ?? 0, lastProgressAt);
-                                    if (!stallProgress.rowSignalSaturated
-                                        && (!nudgePending || nudgeFailed)
-                                        && sliceNow - lastOutputGrowthAt >= deadChannelFastKillMs
-                                        && worktreeSinceLaunch === "unchanged") {
-                                        journal.append("worker-dead", t.id, { slot: slot.name, attempt, silentMs: sliceNow - lastOutputGrowthAt });
-                                        break;
+                                    // OBS-548: the FOURTH leg, and the one the daemon already measured. The three legs above
+                                    // are all channel-side: they say a pane stopped talking. The tree's CPU says whether
+                                    // anything is still WORKING, and the harvest triad forty lines up refuses exactly this
+                                    // conclusion without it — "a worker that is merely thinking still burns CPU and is never
+                                    // concluded here". A live CPU delta is a live worker, by construction; an UNMEASURABLE
+                                    // reading keeps the fail-open contract the probe already states (undefined is never
+                                    // evidence a worker stopped), so on a host whose `ps` cannot be read the fast-kill stands
+                                    // down and the rolling window owns the pane, exactly as pre-T1.
+                                    if (fastKillEligible) {
+                                        const leg = readCpuLeg();
+                                        if (leg.state === "flat") {
+                                            deadChannelKilled = true;
+                                            journal.append("worker-dead", t.id, {
+                                                slot: slot.name, attempt, silentMs: sliceNow - lastOutputGrowthAt,
+                                                cpuMs: leg.cpu.ms, cpuResolutionMs: leg.cpu.resolutionMs,
+                                            });
+                                            break;
+                                        }
+                                        if (!cpuHeld) {
+                                            cpuHeld = true;
+                                            journal.append("worker-dead-held", t.id, {
+                                                slot: slot.name, attempt, reason: leg.state === "accruing" ? "cpu-accruing" : "cpu-unmeasurable",
+                                            });
+                                        }
                                     }
                                     if (nudgeable && !nudged && sliceNow - lastProgressAt >= nudgeAfterSilentMs) {
                                         nudged = true;
@@ -2623,6 +2600,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                                 // steer, driver.nudge is never consulted on this path), so there is no pending daemon
                                 // action for the triad to preempt — the asymmetry with the interactive call site above is
                                 // the absence of the thing being held for, not an oversight.
+                                await armCpuLeg(Date.now() - lastProgressAt >= harvestSilentMs);
                                 if (await harvestConcludes(Date.now() - lastProgressAt))
                                     break;
                             }
@@ -2665,7 +2643,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                     }
                     let result = settleParsed ?? adapter.parse(output, nonce);
                     const workerFinished = finished;
-                    const workerCause = classifyWorkerResultCause({ output, ok: result.ok, finished, exitCode, summary: result.summary, timedOut });
+                    const workerCause = classifyWorkerResultCause({ output, ok: result.ok, finished, exitCode, summary: result.summary, timedOut, deadChannel: deadChannelKilled });
                     journal.append("worker-result", t.id, {
                         ok: result.ok, summary: result.summary, deviations: result.deviations, finished: workerFinished, exitCode,
                         mode: interactive ? "interactive" : "print", ...(workerCause ? { cause: workerCause } : {}),
@@ -2888,6 +2866,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                         ({ results, commits } = await runGates(t, {
                             worktree: wt, baseRef: taskBase, result, author: assignment,
                             commands, baseline, channels: pools.review, judgeChannels: pools.judge, adapters, cfg, artifactDir: journal.dir,
+                            collateral: collateral.get(t.id) ?? [],
                             pipeline: "v185", selectTests: !testGateFailed,
                             via: cfg.visibility.llm === "pane"
                                 ? {
@@ -2930,7 +2909,9 @@ export async function runDaemon(repoRoot, opts = {}) {
                             }
                             graph = setStatus(graph, t.id, "done");
                             saveGraph(repoRoot, graph);
-                            journal.append("task-done", t.id, { attempts: attempt + 1, assignment });
+                            journal.append("task-done", t.id, {
+                                attempts: attempt + 1, assignment, taskContentDigest: contentDigest,
+                            });
                             journal.append("merge", t.id, { branch: taskBranch, commit: await integrationHead(intWt) });
                             // firstAttemptOk/gateFails/consults are recorded FACTS, not policy — a parkKind:"stall" row is
                             // recorded but NOT quality-negative in v1.6; Phase 12 owns reward policy, so flipping it later needs zero data migration.
@@ -2962,11 +2943,15 @@ export async function runDaemon(repoRoot, opts = {}) {
                         await park(t, `${infraFailure.gate}: ${infraFailure.details}`, "infra", assignment, attempt + 1, startMs, gateFails, consults, tokens, metered, retryMode);
                         return;
                     }
+                    // OBS-547: who pays for this red is decided by the run's collateral prediction (see
+                    // dispositionScopeRed) — an authoring defect parks unchargeable before any accounting below.
+                    if (await dispositionScopeRed(t, results, assignment, attempt, startMs, gateFails, consults, tokens, metered, retryMode))
+                        return;
                     gateFails++; // this attempt's gates failed — the one place quality degradation is verified (never inferred from attempts)
                     // v1.53 T3: prefer the CLI's own session id captured from this attempt's output (kimi's resume
                     // trailer) over the harness slot name; absent hook or no capture keeps today's slot-name id.
                     retrySession = { channel: channelKey(assignment), id: adapter.sessionIdFrom?.(output) ?? sessionId, contextTokens };
-                    feedback = results.filter(gateFailed).map((g) => `${g.gate}: ${g.details}`).join("\n\n");
+                    feedback = withCarriedEvidence(results.filter(gateFailed).map((g) => `${g.gate}: ${g.details}`).join("\n\n"));
                     // OBS-189/G3 (park-economics patch): a request-changes review is a findings brief, not a worker
                     // defect — the fix attempt stays on the same channel with the findings as feedback and consumes
                     // no escalation-ladder rung. Bounded by the engagement round cap at the top of this loop.