tickmarkr 1.83.0 → 1.85.0

package/dist/run/daemon.js

@@ -1,4 +1,4 @@
-import { randomBytes } from "node:crypto";
+import { createHash, randomBytes } from "node:crypto";
 import { shq } from "../adapters/types.js";
 import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
@@ -8,7 +8,7 @@ import { classifyDeadChannel, NO_TRAILER_SUMMARY, trailerPattern, UNPARSEABLE_TR
 import { allAdapters, discoverChannels, getAdapter, probeAll, readDoctor } from "../adapters/registry.js";
 import { addUsage, channelKey, matchesTrustDialog, QUOTA_RE } from "../adapters/types.js";
 import { bannerShell, paneDispatchCommand } from "../brand.js";
-import { globalConfigDir, loadConfigWithMode, readOverlayFile, repoOverlayPath, } from "../config/config.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";
@@ -20,13 +20,13 @@ import { augmentRetryBrief, consult, renderRetryGuidance } from "./consult.js";
 import { runEnvironment } from "./environment.js";
 import { cleanupRunWorktrees, gitHead, linkNodeModules, npmDependencyInstallCommand, npmDependencyManifestChanged, sh, shGit, WORKTREE_LAYOUT_CONTRACT, worktreePath } from "./git.js";
 import { runInteractiveSeed } from "./interactive-seed.js";
-import { classifyTaskFailure, classifyWorkerResultCause, engagementComparable, Journal, loadRoutingProfile, newRunId, phaseForGate, recordedTaskFailureKind, reviewRoundsSinceApproval } from "./journal.js";
+import { activeRetryBan, classifyTaskFailure, classifyWorkerResultCause, engagementComparable, GATE_FINGERPRINT_CAP, identicalGateFailures, journaledFailureBrief, Journal, loadRoutingProfile, newRunId, normalizeGateFailure, pendingRepairFindings, phaseForGate, recordedTaskFailureKind, repairsSinceApproval, reviewRoundsSinceApproval, structuredFindings, upheldFeedbackByTask } from "./journal.js";
 import { isDiffCapPark } from "../gates/review.js";
 import { 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 { StallProgressTracker } from "./stall.js";
+import { NUDGEABLE_ADAPTERS, PANE_READ_ROWS, StallProgressTracker, stallSnapshotBannerRows } 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
@@ -69,6 +69,107 @@ export function formatSummary(s) {
     return `done: ${s.done.length}, failed: ${s.failed.length}, human: ${s.human.length}, blocked: ${s.blocked.length}, pending: ${s.pending.length}\nintegration branch: ${s.branch}${tip}`;
 }
 const MAX_ATTEMPTS = 10; // ponytail: hard cap so a pathological ladder can never loop forever
+// v1.85 T3 (retry economics): two repairs per engagement, then the fresh ladder. A repair re-uses the
+// findings and the landed diff instead of re-buying onboarding; when two of them have not closed the
+// battery, the cheaper next move is the ladder's channel change, not a third fix-only pass.
+const MAX_REPAIRS = 2;
+/** A named oracle decided this acceptance failure — deterministic, unlike an LLM judge verdict. */
+const isOracleFailure = (g) => g.details.startsWith("oracle failed:");
+/**
+ * R3 (OBS-186): a gate that DECLINED to run is not a gate that failed. The review gate's skip branch
+ * no longer forges `pass: true` to buy passage, so the merge decision has to read the same predicate
+ * the run surfaces already read (src/run/activity.ts): pass, or an honest declared skip. Without this
+ * the honesty change would silently park every judge-only task at merge — an unrun gate blocking work
+ * it was never asked to review. `skipped` is set only by a gate that says so about ITSELF; a red
+ * verdict from a review that actually ran still fails here, exactly as before.
+ *
+ * ONE pair of predicates, every fold. `!g.pass` was correct only while the sole `pass:false` producer
+ * was a gate that actually failed; the moment a decline can be recorded red, every `!g.pass` in this
+ * file — the retry feedback brief, the review-fix eligibility test, the failing-battery list the
+ * ladder and the fingerprint cap are scored on, the structured findings attached to a blocking
+ * verdict — reads an unrun gate as a defect. `gateFailed` is the seam they now share, and the journal
+ * write below is the seam every OUT-of-file fold shares.
+ */
+const gateSatisfied = (g) => g.pass || g.meta?.skipped === true;
+const gateFailed = (g) => !gateSatisfied(g);
+// v1.85 T3: the gates whose failure IS a deterministic measurement — a machine re-ran a command over a
+// tree and printed the same bytes. Those are the failures the fingerprint cap governs (the ruling names
+// it a "deterministic-gate" cap): a third identical answer to a question already answered twice is the
+// ~663m-across-5-runs loop, whatever rung the ladder happens to stand on. An LLM verdict is a different
+// object — two reviewers, or a judge asked twice, can restate one another without the question being
+// closed — and each already carries a tighter bound of its own: REVIEW_ROUND_CAP parks review at the
+// OPERATOR in two rounds, and a judge verdict rides the ladder and the attempt cap. The boundary is a
+// property of the GATE, never of the ladder rung or of the move that would follow the failure.
+const DETERMINISTIC_GATES = new Set(["build", "test", "lint", "evidence", "scope"]);
+const isDeterministicFailure = (g) => DETERMINISTIC_GATES.has(g.gate) || (g.gate === "acceptance" && isOracleFailure(g));
+/**
+ * Is the failing battery narrow enough that a fix-only pass can close it? The ruling's three cases:
+ * review-only, a single deterministic test/lint gate, or acceptance decided by a named oracle.
+ * Unparseable verdicts and diff-cap trips are excluded exactly as they are from the review-fix retry —
+ * neither names anything a worker can fix.
+ */
+function narrowRepairBattery(failing) {
+    if (failing.length !== 1)
+        return false;
+    const g = failing[0];
+    if (g.gate === "review")
+        return g.meta?.unparseable !== true && !isDiffCapPark(g);
+    if (g.gate === "test" || g.gate === "lint")
+        return true;
+    if (g.gate === "acceptance")
+        return isOracleFailure(g) && g.meta?.unparseable !== true;
+    return false;
+}
+/**
+ * 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
+ * never have asked for — it returned at the judge. Those verdicts stay on the record (they are real,
+ * and the retry brief carries them), but they must not spend the OPERATOR-facing review round budget:
+ * otherwise concurrency alone parks a task rounds early for objections the old pipeline never bought.
+ * A round is the gate-result span opened by each `gates` phase-start, per task.
+ */
+export function decisiveReviewRounds(events) {
+    const open = new Map();
+    const spent = new Set();
+    const close = (taskId) => {
+        const round = open.get(taskId) ?? [];
+        if (round.some((e) => e.data.gate !== "review" && e.data.pass === false)) {
+            for (const e of round)
+                if (e.data.gate === "review" && e.data.pass === false)
+                    spent.add(e);
+        }
+        open.delete(taskId);
+    };
+    for (const e of events) {
+        if (!e.taskId)
+            continue;
+        if (e.event === "phase-start" && e.data.phase === "gates")
+            close(e.taskId);
+        else if (e.event === "gate-result")
+            open.set(e.taskId, [...(open.get(e.taskId) ?? []), e]);
+    }
+    for (const taskId of open.keys())
+        close(taskId); // Map iteration tolerates deleting the current key
+    return events.filter((e) => !spent.has(e));
+}
+/** The fix-only contract: the findings verbatim, then the diff content of the work already landed. */
+function repairBrief(findings, diff, baseRef) {
+    const repair = { findings, diff };
+    return [
+        "## Repair attempt — fix ONLY what these findings name",
+        "The commits from your prior attempt are already in this worktree and their diff is reproduced"
+            + " below. Do NOT re-implement that work, do not start over, and do not revert it: make the"
+            + " smallest change that resolves every finding, then commit.",
+        "",
+        "### Failing gate findings (verbatim)",
+        repair.findings,
+        "",
+        `### The work under review (git diff ${baseRef.slice(0, 12)}..HEAD)`,
+        "```diff",
+        repair.diff,
+        "```",
+    ].join("\n");
+}
 // v1.70 T5: request-changes review rounds a single task may draw before it parks for a human decision
 // instead of cycling. Well below MAX_ATTEMPTS so review non-convergence is caught long before the
 // global cap. ponytail: literal constant; lift to cfg.review.roundCap only if a second knob-turner appears.
@@ -92,19 +193,23 @@ export function setEarlyLaunchLivenessMsForTests(ms) {
 export function resetEarlyLaunchLivenessMsForTests() {
     earlyLaunchLivenessMs = EARLY_LAUNCH_LIVENESS_MS;
 }
-// OBS-201: the liveness nudge — the daemon's ACTIVE response to an idle worker holding no trailer,
-// replacing page-a-human-then-burn-the-window (289 of 692 worker-minutes in one measured day).
-// Gate: herdr classifies the pane idle AND the monotonic tracker has seen nothing for
-// NUDGE_AFTER_SILENT_MS (a worker grinding inside a tool run is `working` and never reaches the
-// gate). One nudge per attempt; if the grace passes still idle with no progress, the wait concludes
-// as a stall NOW and the consult sees the un-answered nudge instead of an hour of silence.
-// Allowlist: claude-code first (steering path proven, OBS-122); widen per adapter only with a
-// captured occupied-frame fixture (OBS-181 scar). The message builder takes no nonce — the
+// OBS-201 + T1 (OBS-262): the liveness nudge — the daemon's ACTIVE response to a worker holding no
+// trailer, replacing page-a-human-then-burn-the-window (289 of 692 worker-minutes in one measured
+// day). Gate: NUDGE_AFTER_SILENT_MS of monotonic-tracker silence, regardless of the herdr status
+// reading (unknown/working no longer suppress it — a wedged TUI often scrapes as either); a
+// `blocked` pane still pages instead, since nudging a dialog prompt can't help. One nudge per
+// attempt; if the grace passes with no progress, the wait concludes as a stall NOW and the consult
+// sees the un-answered nudge instead of an hour of silence. Scope allowlist lives in stall.ts
+// (claude-code only; widening is a fixture-capture chore). The message builder takes no nonce — the
 // self-reference guard holds by construction, an echoed bare token can never match the wait regex.
-export const NUDGEABLE_ADAPTERS = new Set(["claude-code"]);
+export { NUDGEABLE_ADAPTERS } from "./stall.js";
 export const WORKER_NUDGE_MESSAGE = "tickmarkr liveness check: if the task is complete, print your TICKMARKR_RESULT completion trailer exactly as specified in your prompt now. If not, state your next concrete action and continue working.";
-const NUDGE_AFTER_SILENT_MS = 3 * 60_000;
+const NUDGE_AFTER_SILENT_MS = 10 * 60_000; // T1 (OBS-262): >=10m tracker silence — was 3m behind an unreachable status gate
 const WORKER_NUDGE_GRACE_MS = 4 * 60_000;
+// T1 review: a false return from driver.nudge is a DELIVERY outcome (missing pin, readiness
+// stable-frame timeout, read-back hiccup), not proof of an unreachable channel — so one failure
+// is retried once in-slice after this settle, and only a failed retry latches nudgeFailed.
+const NUDGE_REDELIVER_MS = 2_000;
 let nudgeAfterSilentMs = NUDGE_AFTER_SILENT_MS;
 let workerNudgeGraceMs = WORKER_NUDGE_GRACE_MS;
 /** Test seam — shrink the nudge gate and grace without minute-long sleeps. */
@@ -116,6 +221,375 @@ export function resetNudgeTimingForTests() {
     nudgeAfterSilentMs = NUDGE_AFTER_SILENT_MS;
     workerNudgeGraceMs = WORKER_NUDGE_GRACE_MS;
 }
+// T1 (OBS-263): in-loop quota-banner classification — the banner IS output, so the empty-output
+// rules can never catch it and the post-loop QUOTA_RE check only runs after the full window. Two
+// consecutive matching slices plus this much monotonic-tracker silence classify (a worker whose
+// diff merely quotes "rate limit" keeps working undisturbed).
+const QUOTA_BANNER_SILENT_MS = 3 * 60_000;
+let quotaBannerSilentMs = QUOTA_BANNER_SILENT_MS;
+/** Test seam — shrink the quota-banner silence gate without minute-long sleeps. */
+export function setQuotaBannerSilentMsForTests(ms) {
+    quotaBannerSilentMs = ms;
+}
+export function resetQuotaBannerSilentMsForTests() {
+    quotaBannerSilentMs = QUOTA_BANNER_SILENT_MS;
+}
+// T1 (OBS-262): the operator page is UNLATCHED — every eligible slice journals a page, and the
+// notification is delivered again on a status change or once this cadence elapses. The cadence is
+// an operator-spam guard only; it can no longer turn a stall into a single page forever. It sits
+// BELOW the dead-channel fast-kill window on purpose (T1 review): at 5m == 5m the second delivery
+// raced the kill on the same slice boundary, so an idle non-nudgeable pane holding no delta — the
+// exact class the repeat exists for — got exactly one page in production.
+const PAGE_REPEAT_MS = 2 * 60_000;
+let pageRepeatMs = PAGE_REPEAT_MS;
+/** Test seam — shrink the repeat-page cadence without minute-long sleeps. */
+export function setPageRepeatMsForTests(ms) {
+    pageRepeatMs = ms;
+}
+export function resetPageRepeatMsForTests() {
+    pageRepeatMs = PAGE_REPEAT_MS;
+}
+// T1 (R1 dead-channel fast-kill): a worker with no trailer, no worktree delta, and no output
+// growth for this long is dead — conclude immediately instead of burning the rolling window.
+// Per-task timeoutMinutes stays the escape valve for slow-but-live workers.
+const DEAD_CHANNEL_FAST_KILL_MS = 5 * 60_000;
+let deadChannelFastKillMs = DEAD_CHANNEL_FAST_KILL_MS;
+/** Test seam — shrink the fast-kill window without minute-long sleeps. */
+export function setDeadChannelFastKillMsForTests(ms) {
+    deadChannelFastKillMs = ms;
+}
+export function resetDeadChannelFastKillMsForTests() {
+    deadChannelFastKillMs = DEAD_CHANNEL_FAST_KILL_MS;
+}
+// T2 (OBS-264): finished work is harvested, never redone. 18 of 18 observed stalls carried 2-33
+// commits, and the redispatch then re-bought verification of work that had already landed. The
+// liveness triad — commits made by this attempt, a FLAT worker-tree CPU delta, and this much
+// monotonic-tracker silence — CONCLUDES the wait. Conclude, never kill: the pane is harvested by
+// the same tail a window expiry uses, and the carried worktree goes straight to gates. Set at the
+// fast-kill's window on purpose (the OBS-264 arithmetic is "a ~36m stall + ~15m redo becomes a
+// ~5m gate pass"); a worker that is merely thinking still burns CPU and is never concluded here.
+const HARVEST_SILENT_MS = 5 * 60_000;
+let harvestSilentMs = HARVEST_SILENT_MS;
+/** Test seam — shrink the harvest silence gate without minute-long sleeps. */
+export function setHarvestSilentMsForTests(ms) {
+    harvestSilentMs = 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.
+const HARVEST_POLL_MS = 2_000;
+function harvestSliceMs(silentMs) {
+    return Math.max(100, silentMs >= harvestSilentMs ? HARVEST_POLL_MS : harvestSilentMs - silentMs);
+}
+// The synthesized result a carried no-trailer harvest hands to the gates. Distinct from anything a
+// worker can claim: it never comes from adapter.parse, and it is journaled under its own event.
+export const HARVESTED_RESULT_SUMMARY = "harvested: the worktree carries committed work; the worker emitted no TICKMARKR_RESULT trailer";
+/** T4 (OBS-266): identity of the command SET a tip verify ran — a changed command is a different verify. */
+export function commandsHash(commands) {
+    return createHash("sha256").update(JSON.stringify(Object.entries(commands).sort())).digest("hex").slice(0, 12);
+}
+/**
+ * T4 (OBS-266): the journal's LAST verification cycle — the (tip, cmdHash) pair the most recent run
+ * of the verify commands spoke for, the gates it got a pass from, and whether anything failed in it.
+ *
+ * The LAST one, never a history of every pair ever green. "The last GREEN verified SHA" is what the
+ * spec licenses a skip against, and only the last cycle is a statement about the state the run is in
+ * now: after A→B→A the tip really moved, and after commands A→B→A the last thing that ran on this
+ * SHA was command set B — both re-verify. A cycle is the contiguous run of events sharing one pair,
+ * so a cycle cut short by a killed process is missing gates and can never satisfy the caller. A
+ * legacy event (no tip/cmdHash) is unattributable and breaks the chain outright.
+ */
+function lastVerifyCycle(events) {
+    let cur;
+    let afterRunEnd = false;
+    for (const e of events) {
+        // New journals delimit every attempt explicitly. run-end is the legacy delimiter: it starts a
+        // new cycle only when another verify event follows, while preserving the just-closed cycle as
+        // the cache candidate for an otherwise unmoved next run-end.
+        if (e.event === "tip-verify-start") {
+            const { tip, cmdHash } = e.data;
+            cur = typeof tip === "string" && typeof cmdHash === "string"
+                ? { tip, cmdHash, gates: new Set(), failed: false }
+                : undefined;
+            afterRunEnd = false;
+            continue;
+        }
+        if (e.event === "run-end") {
+            afterRunEnd = true;
+            continue;
+        }
+        if (e.event !== "tip-verify" && e.event !== "tip-verify-failed")
+            continue;
+        const { tip, gate, cmdHash } = e.data;
+        if (typeof tip !== "string" || typeof gate !== "string" || typeof cmdHash !== "string") {
+            cur = undefined;
+            afterRunEnd = false;
+            continue;
+        }
+        if (!cur || afterRunEnd || cur.tip !== tip || cur.cmdHash !== cmdHash) {
+            cur = { tip, cmdHash, gates: new Set(), failed: false };
+        }
+        afterRunEnd = false;
+        if (e.event === "tip-verify-failed")
+            cur.failed = true;
+        else
+            cur.gates.add(gate);
+    }
+    return cur;
+}
+/**
+ * OBS-34's strict tip verify, but it stops re-paying for an unmoved tip (~334m corpus-wide; 69.5m in
+ * one park-heavy run whose 18 resume cycles merged nothing new). The verify journals the SHA it
+ * verified and the hash of the command set, so a later run-end can recognize the same verified state:
+ * head equals the LAST green verified SHA, commands unchanged, tree clean → journal
+ * `tip-verify-cached` and skip. ANY doubt — moved head, changed commands, dirty tree, a gate missing
+ * from that cycle's green set, a failure recorded in it, a cycle older than the last one — runs the
+ * full verify. The tip-verify-before-green law is untouched: a cached green is a verified green OF
+ * THAT EXACT COMMIT, established by the most recent real run of the same commands.
+ * Returns whether the tip is failing.
+ */
+export async function verifyIntegrationTipCached(intWt, commands, journal, opts = {}) {
+    const cmdHash = commandsHash(commands);
+    const tip = await gitHead(intWt);
+    const porcelain = await shGit("git status --porcelain", intWt);
+    const clean = porcelain.code === 0 && porcelain.stdout.trim() === "";
+    const last = lastVerifyCycle(journal.read());
+    const cached = last !== undefined && !last.failed && last.tip === tip && last.cmdHash === cmdHash
+        && Object.keys(commands).every((g) => last.gates.has(g));
+    // A pair can be verified red and then green without either SHA or command hash changing (for
+    // example, an external service or ignored fixture recovers). Delimit attempts explicitly so that
+    // the earlier red cannot remain latched into the later complete green cycle.
+    journal.append("tip-verify-start", undefined, { tip, cmdHash, gates: Object.keys(commands), cached: clean && cached });
+    if (clean && cached) {
+        journal.append("tip-verify-cached", undefined, { tip, cmdHash, gates: Object.keys(commands) });
+        // The skip must not read as a red. Every surface derives the tip's verdict from this cycle's
+        // `tip-verify` events (cockpit derive.ts tipVerificationPassed: a run-end claiming "passed" with
+        // ZERO events is fail-closed to FALSE), so a carried-forward green still journals its per-gate
+        // pass — `cached: true` keeps it honest about not having re-run the command.
+        for (const gate of Object.keys(commands)) {
+            journal.append("tip-verify", undefined, { gate, cmd: commands[gate], pass: true, exitCode: 0, cached: true, tip, cmdHash });
+        }
+        return false;
+    }
+    let tipFailed = false;
+    for (const r of await verifyIntegrationTip(intWt, commands, journal.dir)) {
+        if (r.pass) {
+            journal.append("tip-verify", undefined, { gate: r.gate, cmd: r.cmd, pass: true, exitCode: r.exitCode, details: r.details, tip, cmdHash });
+        }
+        else {
+            journal.append("tip-verify-failed", undefined, {
+                gate: r.gate,
+                cmd: r.cmd,
+                exitCode: r.exitCode,
+                fingerprints: r.fingerprints,
+                artifact: r.artifact,
+                lastMergedTask: opts.lastMergedTask,
+                tip,
+                cmdHash,
+            });
+            tipFailed = true;
+        }
+    }
+    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)
@@ -125,6 +599,26 @@ async function commitsAheadOf(base, wt) {
         return [];
     return r.stdout.trim().split("\n").filter(Boolean);
 }
+// T1 (R1 fast-kill): worktree delta = commits ahead of the task base OR any uncommitted change.
+// The fast-kill may only run once the ABSENCE of work has been positively established, so every
+// probe fails OPEN toward "delta": a throwing rev-parse, a non-zero `git log`, and a non-zero
+// `git status` all report delta. commitsAheadOf() is deliberately NOT reused here — it converts a
+// failed log into an empty list, which reads as "no commits" and would kill a live worker.
+async function worktreeHasDelta(wt, base) {
+    try {
+        const head = await gitHead(wt);
+        if (head !== base) {
+            const log = await shGit(`git log --reverse --format=%H ${shq(base)}..${shq(head)}`, wt);
+            if (log.code !== 0 || log.stdout.trim().length > 0)
+                return true;
+        }
+        const r = await shGit("GIT_OPTIONAL_LOCKS=0 git status --porcelain", wt);
+        return r.code !== 0 || r.stdout.trim().length > 0;
+    }
+    catch {
+        return true; // an unreadable worktree is never evidence that the worker did nothing
+    }
+}
 async function cherryPickCommits(wt, commits) {
     const carried = [];
     for (const hash of commits) {
@@ -295,7 +789,12 @@ export async function runDaemon(repoRoot, opts = {}) {
             for (const [id, st] of journal.replayStatuses()) {
                 if (opts.retryFailed && st === "failed" && recordedTaskFailureKind(replayEvents, id) === "dispatch") {
                     graph = setStatus(graph, id, "pending");
-                    resume.delete(id);
+                    // OBS-254: clear ATTEMPT AND CHANNEL STATE ONLY. Deleting the whole entry also deleted
+                    // upheldFeedback — the operator's funded brief — and the next dispatch advertised an empty
+                    // "fix these specifically" heading. A dispatch that died before worker-result then re-ran the
+                    // worker with the uphold's findings silently gone.
+                    const prior = resume.get(id);
+                    resume.set(id, { attempts: 0, tried: [], ...(prior?.upheldFeedback ? { upheldFeedback: prior.upheldFeedback } : {}) });
                     continue;
                 }
                 // operator release: a graph.json edit back to "pending" beats a replayed human/failed park (locked decision 12)
@@ -485,7 +984,7 @@ export async function runDaemon(repoRoot, opts = {}) {
             // reviewer-exclusion list (badReviewers), never a second parallel counter. OBS-189: scoped to the
             // current engagement — an operator approval (uphold or accept) resets the round budget, so an upheld
             // task can dispatch its funded attempt instead of re-parking against the whole journal's history.
-            const reviewRoundsDrawn = () => reviewRoundsSinceApproval(journal.read(), t.id);
+            const reviewRoundsDrawn = () => reviewRoundsSinceApproval(decisiveReviewRounds(journal.read()), t.id);
             // OBS-193: journal the in-gate review retry (mirrors judge-retry) and exclude the flaked seat from
             // later attempts' reviewer picks. One helper, called from both onGate sites (satisfied-gate + main).
             const noteReviewRetry = (g) => {
@@ -498,9 +997,85 @@ export async function runDaemon(repoRoot, opts = {}) {
                     badReviewers.push(rr.flaked);
                 }
             };
+            // v1.85 T3 (ruling R4): every BLOCKING review/judge result lands its findings in the journal
+            // structured — class + canonical path + stable symbol — so a retry, a consult or an auto-uphold
+            // decision reads identity instead of re-parsing prose, and line-number churn is not a new finding.
+            // One helper, both onGate sites (satisfied-gate resume + main attempt loop).
+            const journalGateResult = (g) => {
+                const blocking = gateFailed(g) && (g.gate === "review" || g.gate === "acceptance");
+                // R3 (OBS-186): a gate that DECLINED has no verdict to state, and this row is the ONE seam every
+                // fold outside this file shares. Writing `pass: false` for a decline is what turned a skip into
+                // a failure at all of them at once — the engagement round budget (reviewRoundsSinceApproval,
+                // journal.ts), the operator's failed-gate list (cli/commands/approve.ts), the record's
+                // gate-failure total (cli/commands/report.ts), the cockpit's gate rows (tui/cockpit/derive.ts).
+                // Each keys on `pass === false`; none of them is reachable from this task's file scope, and
+                // patching five copies of the same question would be the wrong fix even if they were. So the
+                // ledger simply does not claim a verdict it does not have.
+                // The legacy baseline declines (a build command the repo never configured) have always written
+                // `pass: true` beside `skipped: true` and every consumer already reads them right, so their row
+                // is untouched: only a decline that would otherwise be recorded RED changes shape here.
+                const unverdicted = g.meta?.skipped === true && !g.pass;
+                journal.append("gate-result", t.id, {
+                    gate: g.gate, ...(unverdicted ? {} : { pass: g.pass }), details: g.details,
+                    ...(g.meta?.skipped === true ? { skipped: true } : {}),
+                    // R3 (OBS-186): a declined review is journal truth, not an absence. `skipped: true` alone
+                    // says a gate did not run; these say WHICH policy declined it and WHY, so a reader of the
+                    // ledger never has to infer participation from a details string. The green-skip branch that
+                    // made this row indistinguishable from a pass is gone (src/gates/review.ts).
+                    ...(g.meta?.verdict === "skipped"
+                        ? { verdict: "skipped", policy: g.meta.policy, reason: g.meta.reason }
+                        : {}),
+                    // T4 (OBS-265): a test verdict says WHICH suite spoke. A round runs the selected subset as a
+                    // screen and the full suite as the verdict, so without these two the journal would carry a
+                    // `test` row whose scope no consumer could recover.
+                    ...(Array.isArray(g.meta?.selectedTests) ? { selectedTests: g.meta.selectedTests } : {}),
+                    ...(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) } : {}),
+                });
+            };
+            // R3 (OBS-186): judge ‖ review are launched together and publish in COMPLETION order
+            // (run-gates.ts) — a race. Three oracles assert the opposite: a scripted run's journal is
+            // byte-identical run to run (tests/run/narration.test.ts, tests/run/notify-identity.test.ts), and
+            // a round's gate-result order matches its phase-start order (tests/run/daemon.test.ts). Retiring
+            // complexityThreshold is what REACHES this, not what introduces it: those fixtures used to skip
+            // review and journal ONE verdict row per round, so the pair's order was never exercised — and the
+            // operator's config has run `complexityThreshold: 0` since 2026-07-31, so production rounds have
+            // journaled both siblings all along.
+            //
+            // Ordering is the LEDGER's job, not the pipeline's. run-gates still reports each completion the
+            // instant it happens; the daemon writes its ledger in GATE_NAMES order. Only the LATER gate is
+            // ever held, and only while an earlier sibling is still in flight — a review that finishes first
+            // waits for acceptance, never the reverse. That keeps T4's durability where it pays (the first
+            // verdict to land is still published immediately) and bounds the exposure to one row for the
+            // remainder of one already-running gate. A gate that THROWS kills the round before merge, so a
+            // row held behind it is lost with the round it belonged to — not a verdict that could have merged.
+            const parallelPending = new Set();
+            let heldParallel;
+            const notePhaseStart = (e) => {
+                if (e.parentAt !== undefined)
+                    parallelPending.add(e.gate);
+            };
+            const inParallelOrder = (gate, publish) => {
+                parallelPending.delete(gate);
+                const rank = GATE_NAMES.indexOf(gate);
+                if ([...parallelPending].some((p) => GATE_NAMES.indexOf(p) < rank)) {
+                    heldParallel = publish;
+                    return;
+                }
+                publish();
+                const held = heldParallel;
+                heldParallel = undefined;
+                held?.();
+            };
             // OBS-189: the operator upheld the reviewer — the findings ARE the brief for this funded attempt.
-            let feedback = rs?.upheldFeedback
-                ? `The operator UPHELD the reviewer's findings — address them without discarding landed work.\nreview: ${rs.upheldFeedback}`
+            // OBS-254: RE-DERIVED from the journal here, at prompt-build time, rather than trusted to survive
+            // 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
+                ? `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
@@ -509,6 +1084,12 @@ export async function runDaemon(repoRoot, opts = {}) {
             let tokens; // SPEND-02: accumulated across attempts — parked spend is still spend
             let metered = 0; // SPEND-02: attempts that returned a usage record; distinguishes unmetered from measured-zero
             let tipMoves = 0; // OBS-15: one re-gate allowance per task, never reset by a worker retry
+            // T4 (OBS-265): a task whose test gate has failed once stops selecting tests for good — the
+            // selector already proved it cannot speak for this diff, so every later round runs full. SEEDED
+            // FROM THE JOURNAL, not from process memory: a park, a `resume`, or a daemon restart must not
+            // hand the selector a clean slate it did not earn, and the journal is the only state that
+            // survives all three (same read as upheldFeedbackByTask above).
+            let testGateFailed = journal.read().some((e) => e.event === "gate-result" && e.taskId === t.id && e.data.gate === "test" && e.data.pass === false);
             let retryMode = "fresh";
             let lastContextTokens; // v1.23 reset signal, including stalled/quota attempts
             // v1.29: only a gate-failed attempt can seed same-session retry. The next attempt consumes this
@@ -568,7 +1149,14 @@ export async function runDaemon(repoRoot, opts = {}) {
                 });
                 await driver.notify(`tickmarkr ${runId}: ${t.id} consult verdict: ${v.action}`, { tier: "attention" });
                 if (v.action === "retry") {
-                    feedback = renderRetryGuidance(v) || feedback;
+                    // The guidance is ADDED to the brief, never swapped for it: the failure bytes the journal
+                    // already holds are the one thing the next attempt cannot rediscover for free. The
+                    // fingerprint cap's ban on an identical retry is NOT enforced here — a verdict is only one
+                    // of several ways this task reaches a re-dispatch, so the ban is enforced at the dispatch
+                    // seam every one of them passes through (see enforceRetryBan).
+                    const guidance = renderRetryGuidance(v);
+                    if (guidance && !feedback.includes(guidance))
+                        feedback = feedback ? `${feedback}\n\n${guidance}` : guidance;
                     return true;
                 }
                 if (v.action === "reroute") {
@@ -657,7 +1245,36 @@ export async function runDaemon(repoRoot, opts = {}) {
                 };
                 const gateAuthor = rs?.lastAssignment ?? assignment;
                 const satisfiedIndex = GATE_NAMES.indexOf(satisfiedGate);
-                const remainingGates = t.gates.filter((gate) => GATE_NAMES.indexOf(gate) > satisfiedIndex);
+                // The serial pipeline could have at most one blocking result, so "everything after the
+                // approved gate" was enough. v1.85 can record both verdict siblings red in one round, and a
+                // selected test screen can be green without a complete suite. Approval waives exactly its
+                // named gate: every other red from that round is re-run, and test is forced unless the prior
+                // journal proves a full suite completed.
+                const priorEvents = journal.read();
+                let priorRoundStart = -1;
+                for (let i = priorEvents.length - 1; i >= 0; i--) {
+                    const e = priorEvents[i];
+                    if (e.event === "phase-start" && e.taskId === t.id && e.data.phase === "gates") {
+                        priorRoundStart = i;
+                        break;
+                    }
+                }
+                const priorResults = new Map();
+                for (const e of priorEvents.slice(priorRoundStart + 1)) {
+                    if (e.event !== "gate-result" || e.taskId !== t.id || typeof e.data.gate !== "string"
+                        || !GATE_NAMES.includes(e.data.gate))
+                        continue;
+                    priorResults.set(e.data.gate, e);
+                }
+                const remainingGates = t.gates.filter((gate) => {
+                    if (gate === satisfiedGate)
+                        return false;
+                    const prior = priorResults.get(gate)?.data;
+                    const followsApproved = GATE_NAMES.indexOf(gate) > satisfiedIndex;
+                    const otherRed = prior?.pass === false;
+                    const needsFullSuite = gate === "test" && prior?.fullSuite !== true;
+                    return followsApproved || otherRed || needsFullSuite;
+                });
                 const resumedTask = { ...t, gates: remainingGates };
                 gateLoop: while (true) {
                     const gated = await gitHead(wt);
@@ -665,6 +1282,8 @@ export async function runDaemon(repoRoot, opts = {}) {
                     const { results } = await runGates(resumedTask, {
                         worktree: wt, baseRef: taskBase, result: priorResult, author: gateAuthor,
                         commands, baseline, channels, adapters, cfg, artifactDir: journal.dir,
+                        // a recheck re-verifies a human's release: it never selects tests down, it runs the suite.
+                        pipeline: "v185",
                         via: cfg.visibility.llm === "pane"
                             ? {
                                 driver: trackedDriver,
@@ -677,25 +1296,25 @@ export async function runDaemon(repoRoot, opts = {}) {
                         excludeReviewers: badReviewers,
                         onGate: async (e) => {
                             if (e.phase === "start") {
-                                journal.phaseStart(t.id, phaseForGate(e.gate), { gate: e.gate, index: e.index, total: e.total });
+                                notePhaseStart(e);
+                                journal.phaseStart(t.id, phaseForGate(e.gate), { gate: e.gate, index: e.index, total: e.total, ...(e.parentAt === undefined ? {} : { parallel: true }) });
                                 return;
                             }
                             const g = e.result;
-                            journal.append("gate-result", t.id, {
-                                gate: g.gate, pass: g.pass, details: g.details,
-                                ...(g.meta?.skipped === true ? { skipped: true } : {}),
+                            inParallelOrder(g.gate, () => {
+                                journalGateResult(g);
+                                noteReviewRetry(g);
+                                if (g.gate === "review" && !g.pass && /unparseable/.test(g.details)
+                                    && typeof g.meta?.reviewer === "string") {
+                                    badReviewers.push(g.meta.reviewer);
+                                }
                             });
-                            noteReviewRetry(g);
-                            if (g.gate === "review" && !g.pass && /unparseable/.test(g.details)
-                                && typeof g.meta?.reviewer === "string") {
-                                badReviewers.push(g.meta.reviewer);
-                            }
                         },
                     });
                     const approvedCommits = await commitsAheadOf(taskBase, wt);
                     graph = addEvidence(graph, t.id, { commits: approvedCommits, gateResults: results });
                     saveGraph(repoRoot, graph);
-                    if (!results.every((g) => g.pass)) {
+                    if (!results.every(gateSatisfied)) {
                         gateFails++;
                         await park(t, "post-approval gate failed", "gate-fail", gateAuthor, rs?.attempts ?? 0, startMs, gateFails, consults, tokens, metered, retryMode);
                         return;
@@ -772,15 +1391,58 @@ export async function runDaemon(repoRoot, opts = {}) {
                 // over-threshold context still forces fresh, and the escalation ladder bounds the chain.
                 const priorSession = retrySession;
                 retrySession = undefined;
+                // v1.85 T3: the fingerprint cap banned an IDENTICAL retry — this task+gate already produced the
+                // same failure bytes twice, so re-running the same channel on the same brief is a paid
+                // re-measurement of an answer the journal already holds. Enforced HERE, at the one seam every
+                // re-dispatch passes through, and NOT at the consult verdict that set it: a terminal verdict
+                // falling through to the cap's own `retry` rung, a review-fix round and the ladder itself all
+                // reach a dispatch without ever consulting again, and worker-launch below would then expire a
+                // ban nothing had honoured. Bound to the channel the cap fired on, so a move that already went
+                // elsewhere is not refused for a ban that was never about it; with no untried channel left the
+                // task parks naming the ban rather than buying a third round.
+                const banned = activeRetryBan(journal.read(), t.id, channelKey(assignment));
+                if (banned) {
+                    const next = failover("escalate");
+                    journal.append("retry-same-banned", t.id, {
+                        gate: banned, from: channelKey(assignment), to: next ? channelKey(next) : null,
+                    });
+                    if (!next) {
+                        await park(t, `identical ${banned} failure twice this engagement — an identical retry is banned and no untried channel is left`, "gate-fail", assignment, attempt, startMs, gateFails, consults, tokens, metered, retryMode);
+                        return;
+                    }
+                    assignment = next;
+                    if (!tried.includes(channelKey(next)))
+                        tried.push(channelKey(next));
+                }
                 const retryAdapter = adapters.find((a) => a.id === assignment.adapter);
-                retryMode = priorSession
-                    && priorSession.channel === channelKey(assignment)
-                    && (priorSession.contextTokens !== undefined
-                        ? priorSession.contextTokens < cfg.contextWarnTokens
-                        : retryAdapter?.resumeUnknownContext === true)
-                    && retryAdapter?.resumeCommand
-                    ? "resume"
-                    : "fresh";
+                // v1.85 T3: a repair attempt is dispatched FRESH by construction — the whole point is that the
+                // brief, not a surviving session, carries the findings and the diff. It therefore outranks the
+                // session-resume choice, and the mode is journaled so the ledger can price repairs against
+                // fresh re-dispatches. Read from the JOURNAL, before this dispatch's own event lands: a run that
+                // stopped between funding the repair and sending it resumes still carrying the findings.
+                const journaledSoFar = journal.read(); // read BEFORE this dispatch's own event lands
+                let repairFindings = pendingRepairFindings(journaledSoFar, t.id);
+                // OBS-254, one layer below the upheld brief: the ordinary gate-fail brief was loop-local, so any
+                // path that rebuilt this task's state (a resume, `--retry-failed`, a fresh daemon) dispatched a
+                // retry that had forgotten why it was retrying. Re-derived from the journal here, at prompt-build
+                // time, and MERGED rather than substituted — a retry never discards what the journal already
+                // holds. Row-wise, because the live brief may already quote one of them (a delivery-readiness
+                // failure the loop just wrote) and repeating it helps no worker.
+                const journaledRows = journaledFailureBrief(journaledSoFar, t.id).filter((row) => !feedback.includes(row));
+                if (journaledRows.length > 0) {
+                    const brief = journaledRows.join("\n\n");
+                    feedback = feedback ? `${brief}\n\n${feedback}` : brief;
+                }
+                retryMode = repairFindings
+                    ? "repair"
+                    : priorSession
+                        && priorSession.channel === channelKey(assignment)
+                        && (priorSession.contextTokens !== undefined
+                            ? priorSession.contextTokens < cfg.contextWarnTokens
+                            : retryAdapter?.resumeUnknownContext === true)
+                        && retryAdapter?.resumeCommand
+                        ? "resume"
+                        : "fresh";
                 // v1.23 T3: over-threshold context still forces fresh at the retry boundary; never interrupt a
                 // running attempt. Unknown/below emits no reset event.
                 if (attempt > 0 && lastContextTokens !== undefined && lastContextTokens >= cfg.contextWarnTokens) {
@@ -808,6 +1470,15 @@ export async function runDaemon(repoRoot, opts = {}) {
                     carriedCommits = await cherryPickCommits(wt, commitsToCarry);
                     journal.append("worktree-recreation", t.id, { attempted: commitsToCarry, carried: carriedCommits });
                 }
+                // T2 review (material): harvest eligibility is "does this WORKTREE carry unverified work",
+                // measured against taskBase — the same base the fast-kill's delta probe and the gates
+                // themselves use. It was measured against this attempt's post-carry HEAD, which excluded
+                // every commit cherry-picked forward: attempt 0 commits and walls, attempt 1 receives that
+                // commit and goes silent, and the silent retry — whose worktree already held the whole
+                // deliverable — was NOT harvested, took the stall consult, and could be redispatched to
+                // re-produce it. The routing branches that "starting HEAD" was protecting no longer need it:
+                // quota, dead-channel and provider-death all classify the PRE-HARVEST outcome below and fire
+                // BEFORE the synthesis, so carried-only work reaches gates without bypassing any failover.
                 const priorNamed = [...new Set([...commitsToCarry, ...carriedCommits])];
                 const presentCommits = new Set(carriedCommits);
                 for (const h of commitsToCarry) {
@@ -836,6 +1507,36 @@ export async function runDaemon(repoRoot, opts = {}) {
                     });
                     await driver.notify(`tickmarkr ${runId}: ${t.id} lost ${lostCommits.length} of ${commitsToCarry.length} landed commit(s) recreating its worktree — it will re-do that work`, { tier: "attention" });
                 }
+                // v1.85 T3: "fully carried commits" is a repair PRECONDITION, and it is re-validated here —
+                // the eligibility test ran one attempt ago, but the carry that decides it happens above, on
+                // this dispatch. A tree that lost part of the prior attempt's work cannot be repaired: a
+                // fix-only contract ("do NOT re-implement that work") over a diff whose implementation is
+                // missing would have the worker patch an incomplete tree and forbid it from rebuilding the
+                // rest. The fresh ladder owns this dispatch instead. `retryMode` is corrected before
+                // worker-launch records it, so the ledger's launch event names what the worker actually got.
+                if (repairFindings !== undefined && lostCommits.length > 0) {
+                    journal.append("repair-cancelled", t.id, { reason: "carry incomplete", attempted: commitsToCarry, lost: lostCommits });
+                    repairFindings = undefined;
+                    retryMode = "fresh";
+                }
+                // v1.85 T3: a repair dispatch replaces the bare gate-fail brief with a fix-only contract that
+                // carries the failing findings VERBATIM and the diff CONTENT of the work already in this
+                // worktree. The measured loss it removes: 62 of 68 re-dispatches were fresh, each re-buying
+                // ~20m of onboarding to rediscover a diff and a finding the journal already held.
+                // The diff is measured HERE, from the worktree the worker will actually open, after the carry —
+                // never from the pre-recreation tree, so what the brief quotes is what the worker has.
+                if (repairFindings) {
+                    const raw = await shGit(`git diff ${shq(taskBase)}..HEAD`, wt);
+                    const cap = cfg.gates.diffCap ?? DEFAULT_DIFF_CAP; // same fallback the measuring gates use
+                    const diff = raw.stdout.length > cap
+                        ? `${raw.stdout.slice(0, cap)}\n… diff truncated at gates.diffCap (${cap} bytes)`
+                        : raw.stdout;
+                    const brief = repairBrief(repairFindings, diff, taskBase);
+                    // anything the live brief holds beyond the journaled findings (a consult's guidance) is kept:
+                    // a repair adds the diff and the fix-only contract, it never subtracts what was already known.
+                    feedback = feedback && !brief.includes(feedback) ? `${brief}\n\n${feedback}` : brief;
+                    journal.append("repair-dispatch", t.id, { diffBytes: diff.length, capped: raw.stdout.length > cap });
+                }
                 if (feedback || priorNamed.length > 0) {
                     feedback = augmentRetryBrief(feedback, { attempted: commitsToCarry, carried: carriedCommits, present: presentCommits });
                 }
@@ -892,6 +1593,90 @@ export async function runDaemon(repoRoot, opts = {}) {
                     workerCmd,
                     exitMarkerCmd,
                 ].join("\n"));
+                // T2 (OBS-264): the liveness triad, shared by BOTH wait loops (a headless worker stalls on
+                // finished work exactly as a visible one does — and rode the whole window before this). It
+                // sits ABOVE the fast-kill and the nudge because the population it governs is the opposite
+                // one: the kill condemns a pane holding NOTHING, while every one of the 18 observed stalls
+                // held 2-33 commits that the redispatch then re-bought. Concluding is not killing — the
+                // post-loop tail harvests this attempt exactly as a window expiry does, and the carried
+                // worktree goes to gates. The probe runs only once the tracker is ALREADY silent, so a
+                // working worker never pays for it, and an unreadable snapshot RESETS the observation:
+                // unmeasurable CPU is never evidence a worker stopped.
+                // v1.85 T3: the prompt has now actually reached a worker. The journal-derived retry decisions (a
+                // funded repair's findings, an identical-retry ban) expire HERE and nowhere earlier: everything
+                // between task-dispatch and this line — worktree recreation, setup, prompt write, slot allocation,
+                // the launch itself — can still die with no worker having seen the brief, and `--retry-failed`
+                // must then re-send that same brief rather than a fresh prompt on a possibly banned channel.
+                const noteLaunched = () => journal.append("worker-launch", t.id, { attempt, retryMode });
+                let cpuFlat;
+                let cpuAccountant;
+                let cpuGapCount = 0;
+                let unmeasurableNoted = false;
+                // One line per attempt, whichever way the CPU leg turns out to be unmeasurable. A triad that
+                // can never conclude is this feature silently ABSENT — on a host whose `ps` the probe cannot
+                // read, every stall would ride its whole window out again with nothing saying why. Named once
+                // per attempt, not per slice: the condition is structural, and a per-slice line would bury it.
+                const noteUnmeasurable = (reason) => {
+                    if (unmeasurableNoted)
+                        return;
+                    unmeasurableNoted = true;
+                    journal.append("worker-harvest-unmeasurable", t.id, { slot: slot.name, attempt, reason });
+                };
+                const harvestConcludes = async (silentMs) => {
+                    if (silentMs < harvestSilentMs) {
+                        await cpuAccountant?.stop();
+                        cpuAccountant = undefined;
+                        cpuFlat = undefined;
+                        cpuGapCount = 0;
+                        return false;
+                    }
+                    // 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) {
+                        noteUnmeasurable("interactive-seed launch is not in the probed process tree");
+                        return false;
+                    }
+                    if (cpuAccountant === undefined) {
+                        cpuAccountant = new WorkerTreeCpuAccountant(dispatchScript, wt);
+                        await cpuAccountant.start();
+                    }
+                    const observation = cpuAccountant.read();
+                    if (observation.gaps !== cpuGapCount) {
+                        cpuGapCount = observation.gaps;
+                        cpuFlat = undefined;
+                        noteUnmeasurable("one or more worker process snapshots could not be read");
+                    }
+                    const cpu = observation.cpu;
+                    if (cpu === undefined) {
+                        // Unmeasurable CPU is never evidence a worker stopped: RESET the observation rather than
+                        // conclude on it, and name the gap — a probe whose snapshot never parses is the same
+                        // 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;
+                    }
+                    const now = Date.now();
+                    if (cpu.ms !== cpuFlat?.ms) {
+                        cpuFlat = { ms: cpu.ms, since: now };
+                        return false;
+                    }
+                    if (now - cpuFlat.since < harvestCpuFlatWindowMs(cpu.resolutionMs))
+                        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,
+                    });
+                    return true;
+                };
                 // SPEND-01: this attempt's dispatch wall-clock — the usage collect cursor. Captured once here, the
                 // single site, so a test can reason about it; keep Date.now() out of profile.ts (still pure) and
                 // out of adapter module scope (the cursor is a parameter, threaded from the daemon).
@@ -931,6 +1716,10 @@ export async function runDaemon(repoRoot, opts = {}) {
                 let output;
                 let exitCode;
                 let timedOut = 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.
+                let processExited = false;
                 let earlyLaunchDead = false;
                 let settleParsed;
                 let seedResult;
@@ -966,26 +1755,369 @@ export async function runDaemon(repoRoot, opts = {}) {
                     await park(t, "escalation ladder exhausted", "ladder-exhausted", assignment, attempt + 1, startMs, gateFails, consults, tokens, metered, retryMode);
                     return false;
                 };
-                if (interactive) {
-                    // v1.2 interactive: the TUI doesn't exit on completion — the trailer is the finish line.
-                    // The exit wrapper still fires if the TUI dies (crash/quit): fast-fail instead of burning the timeout.
-                    finished = false;
-                    exitCode = null;
-                    if (adapter.interactiveSeed) {
-                        // v1.69 T6: launch the real TUI without a prompt, wait for readiness, inject one seed turn,
-                        // then fall through to the normal trailer harvest. A failed seed is recorded as a finished
-                        // failure rather than allowed to race the trailer wait.
-                        try {
-                            seedResult = await runInteractiveSeed({ driver, slot, adapter, assignment, promptFile, taskTimeoutMinutes });
+                try {
+                    if (interactive) {
+                        // v1.2 interactive: the TUI doesn't exit on completion — the trailer is the finish line.
+                        // The exit wrapper still fires if the TUI dies (crash/quit): fast-fail instead of burning the timeout.
+                        finished = false;
+                        exitCode = null;
+                        if (adapter.interactiveSeed) {
+                            // v1.69 T6: launch the real TUI without a prompt, wait for readiness, inject one seed turn,
+                            // then fall through to the normal trailer harvest. A failed seed is recorded as a finished
+                            // failure rather than allowed to race the trailer wait.
+                            try {
+                                seedResult = await runInteractiveSeed({ driver, slot, adapter, assignment, promptFile, taskTimeoutMinutes });
+                            }
+                            catch (error) {
+                                if (!(error instanceof DeliveryReadinessError))
+                                    throw error;
+                                if (await handleDeliveryReadiness(error))
+                                    continue attempts;
+                                return;
+                            }
+                            noteLaunched();
+                            output = seedResult.output;
                         }
-                        catch (error) {
-                            if (!(error instanceof DeliveryReadinessError))
-                                throw error;
-                            if (await handleDeliveryReadiness(error))
-                                continue attempts;
-                            return;
+                        else {
+                            try {
+                                await driver.run(slot, paneDispatchCommand(dispatchScript));
+                            }
+                            catch (error) {
+                                if (!(error instanceof DeliveryReadinessError))
+                                    throw error;
+                                if (await handleDeliveryReadiness(error))
+                                    continue attempts;
+                                return;
+                            }
+                            noteLaunched();
+                            output = await driver.read(slot, PANE_READ_ROWS);
+                        }
+                        if (seedResult?.seedFailed) {
+                            finished = false;
+                        }
+                        else {
+                            // v1.22 T5 / OBS-19: auto-answer a fingerprint-matched trust dialog exactly once per slot.
+                            // Any other blocked/idle dialog pages the operator (unlatched since T1 — see below).
+                            let trustAnswered = false;
+                            // OBS-201: one liveness nudge per attempt; the grace deadline is its OWN timer, never the
+                            // stall window (the nudge's pane echo is absorbed before it starts, or the echo itself
+                            // would reset the window and make the early conclusion unreachable).
+                            let nudged = false;
+                            let nudgeFailed = false; // T1: an undeliverable nudge — the operator, not the daemon, is the actor
+                            let nudgeDeadline;
+                            // T1: page cadence, NOT a latch — a second page fires on a status change or once
+                            // pageRepeatMs elapses, so an operator who missed the first one is paged again.
+                            let lastPagedStatus;
+                            let lastPagedAt = 0;
+                            // T1 review: worker-status is journaled on CHANGE, not per slice — every journal append
+                            // is narrated to the run's live surface (cli/commands/run.ts) and feeds activity.ts's
+                            // `now:` cell, so a per-slice append wrote a status line per worker per ~30s slice and
+                            // pinned `now` to worker-status. On-change keeps post-hoc analysis at a fraction of the
+                            // volume while still recording which gate held.
+                            let lastStatus;
+                            finished = false;
+                            exitCode = null;
+                            // OBS-54: reaping keys on new pane output, not dispatch wall clock. Poll at least twice per
+                            // stall window (and at the existing 30s cadence for normal windows) so an active worker resets it.
+                            const stallWindowMs = taskTimeoutMinutes * 60_000;
+                            // v1.76: only monotonic work (seed submission, transcript growth, or context growth) resets
+                            // the stall clock. Raw pane differences are terminal chrome until proven otherwise.
+                            let everHadOutput = output.length > 0;
+                            const stallProgress = new StallProgressTracker();
+                            stallProgress.observe({ paneText: output, seedSubmitted: true, contextTokens });
+                            let lastProgressAt = Date.now();
+                            // T1: in-loop detector state — consecutive quota-banner slices. The dead-channel fast-kill
+                            // reads the tracker's RAW row-growth clock (lastRowGrowthAt), not the re-arm-suppressed
+                            // lastProgressAt — see the kill below.
+                            let quotaStreak = 0;
+                            let rowSaturationHeld = false; // journaled once per attempt when the kill stands down
+                            while (Date.now() - lastProgressAt < stallWindowMs) {
+                                const sliceStart = Date.now();
+                                const remaining = stallWindowMs - (sliceStart - lastProgressAt);
+                                let slice = Math.min(BLOCKED_POLL_MS, Math.max(100, Math.min(stallWindowMs / 2, remaining)));
+                                if (!everHadOutput) {
+                                    const earlyLeft = earlyLaunchLivenessMs - (sliceStart - attemptStart);
+                                    if (earlyLeft > 0)
+                                        slice = Math.min(slice, earlyLeft);
+                                }
+                                // T2 (OBS-264): never sleep PAST the instant the harvest probe becomes eligible, and past
+                                // it let the probe own the cadence (HARVEST_POLL_MS) — a 30s trailer slice would
+                                // otherwise cost a minute per pair of CPU samples. At the shipped 5m gate this leaves
+                                // every slice before the gate exactly as long as it already was.
+                                slice = Math.min(slice, harvestSliceMs(sliceStart - lastProgressAt));
+                                if (await driver.waitOutput(slot, `(${trailerPattern(nonce)})|TICKMARKR_EXIT_${nonce}:\\d`, slice, { regex: true })) {
+                                    // verify before accepting: a worker that merely DISPLAYS a marker (e.g. editing tickmarkr's
+                                    // own source, where "TICKMARKR_EXIT:" is a string literal) must not end the wait. Only a
+                                    // parseable trailer or a digit-suffixed exit marker in the harvest is completion.
+                                    output = await driver.read(slot, PANE_READ_ROWS); // TUI transcripts carry chrome — read deeper than print's 500
+                                    finished = new RegExp(trailerPattern(nonce)).test(output);
+                                    const exit = exitRe.exec(output);
+                                    if (finished || exit) {
+                                        exitCode = exit ? Number(exit[1]) : null; // null ⇔ the TUI is still alive
+                                        await sampleContext(); // final poll-seam sample before leaving the wait
+                                        break;
+                                    }
+                                }
+                                const paneText = await driver.read(slot, PANE_READ_ROWS);
+                                if (paneText.length > 0)
+                                    everHadOutput = true;
+                                // OBS-117 (v1.71 T6): zero raw output by the early-launch deadline is a dead channel now.
+                                if (!everHadOutput && Date.now() - attemptStart >= earlyLaunchLivenessMs) {
+                                    earlyLaunchDead = true;
+                                    output = paneText;
+                                    break;
+                                }
+                                // v1.23 T2: piggyback on this poll slice — same cadence as blocked/idle checks, no new timer.
+                                await sampleContext();
+                                if (stallProgress.observe({ paneText, contextTokens })) {
+                                    lastProgressAt = Date.now();
+                                    // T1 review fix: progress AFTER a delivered nudge means the worker answered it —
+                                    // disarm the grace deadline. Without this the expiry below fires at the next
+                                    // quiet patch ≥ the grace (4m) measured from the rolling lastProgressAt, so the
+                                    // exact population the nudge rescued (workers prone to long silences, e.g. a 6m
+                                    // test run) was force-concluded as if it had ignored the nudge. The answered
+                                    // worker returns to the full rolling window, and the consult sees the truth.
+                                    if (nudged && nudgeDeadline !== undefined) {
+                                        nudgeDeadline = undefined;
+                                        journal.append("worker-nudge-answered", t.id, { slot: slot.name, attempt });
+                                    }
+                                }
+                                const sliceNow = Date.now();
+                                // T1 (OBS-263): quota banners are classified IN-LOOP — two consecutive matching slices
+                                // plus >=3m tracker silence — then the post-loop quota failover runs NOW, not after the
+                                // window. The match reads the chrome-filtered tail: the bottom of a rendered TUI frame
+                                // is fixed composer/welcome chrome (codex pins a "usage limit resets available" line
+                                // there — it matched every frame of the wedged-MCP fixture), so the known chrome is
+                                // filtered by identity, never by novelty — a banner already on screen at launch
+                                // classifies exactly like one printed mid-attempt (T1 review: a novelty baseline
+                                // exculpated the launch-throttle case forever).
+                                if (QUOTA_RE.test(stallSnapshotBannerRows(paneText)))
+                                    quotaStreak++;
+                                else
+                                    quotaStreak = 0;
+                                if (quotaStreak >= 2 && sliceNow - lastProgressAt >= quotaBannerSilentMs) {
+                                    // no `output =` here: the post-loop no-trailer tail re-reads the pane anyway, so an
+                                    // assignment would only split the classification read from the verdict read.
+                                    journal.append("quota-banner", t.id, { slot: slot.name, attempt, silentMs: sliceNow - lastProgressAt });
+                                    break;
+                                }
+                                // T1 (OBS-262): the `paged` latch is deleted — status is sampled EVERY slice (and
+                                // journaled on change — see lastStatus above), so post-hoc analysis can see which
+                                // gate held. page on "idle" too: herdr's blocked-scrape is strict and proved flaky
+                                // for TUI dialogs (live check: cursor's trust dialog scraped as idle).
+                                // "unknown"/"working" never page.
+                                const st = await driver.status(slot);
+                                if (st !== lastStatus) {
+                                    lastStatus = st;
+                                    journal.append("worker-status", t.id, { slot: slot.name, status: st, attempt });
+                                }
+                                if (st === "blocked" || st === "idle") {
+                                    // T5: once-per-slot auto-answer when the adapter declares a trust dialog and the pane
+                                    // text matches. tickmarkr created the worktree from the operator's own repo — safe by construction.
+                                    if (!trustAnswered && adapter.trustDialog && driver.sendKey) {
+                                        try {
+                                            const paneText = await driver.read(slot, 80);
+                                            if (matchesTrustDialog(paneText, adapter.trustDialog)) {
+                                                trustAnswered = true;
+                                                // v1.25 T1: audit trail for live runs — prove the dialog appeared and was answered.
+                                                // Latch + sendKey + no-page continue stay byte-identical; this append is additive only.
+                                                journal.append("trust-auto-answer", t.id, { slot: slot.name, adapter: adapter.id });
+                                                await driver.sendKey(slot, adapter.trustDialog.key);
+                                                const spent = Date.now() - sliceStart;
+                                                if (spent < slice)
+                                                    await new Promise((r) => setTimeout(r, Math.min(slice - spent, 1_000)));
+                                                continue; // do not page — keep waiting for the trailer
+                                            }
+                                        }
+                                        catch {
+                                            /* read/send failed — fall through to page the operator */
+                                        }
+                                    }
+                                }
+                                // T1 (OBS-262): the daemon ACTS on a silent worker before paging anyone. Gate: monotonic
+                                // tracker silent ≥ the nudge threshold — the herdr status reading (idle/unknown/working)
+                                // no longer holds the gate hostage; only `blocked` stays page-only (nudging a dialog
+                                // prompt can't help). One nudge per attempt, then the grace timer owns the conclusion.
+                                const nudgeable = st !== "blocked" && !!driver.nudge && NUDGEABLE_ADAPTERS.has(adapter.id);
+                                // T1 review (answer-then-die): `nudgeable` is a per-slice property of the adapter and
+                                // status — it is NOT "a daemon action is pending". An action is pending only while the
+                                // nudge can still fire (un-nudged) or its grace window is armed; once the worker
+                                // ANSWERS, the disarm above clears nudgeDeadline while `nudged` stays latched, and the
+                                // 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);
+                                // 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
+                                // of monotonic tracker silence (defined once, above, and run identically by the print loop).
+                                // T2 review (material): it carries the SAME nudge hold as the fast-kill below, by the
+                                // same clause and for the same reason. The CPU leg cannot tell "idle because finished"
+                                // from "idle because holding an unsubmitted turn in its input box" — both read flat CPU
+                                // under a silent tracker — and the nudge is the one signal that can. Under the shipped
+                                // constants (harvest 5m < nudge 10m) an unheld harvest concluded every committed
+                                // 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))
+                                    break;
+                                // T1 (R1 dead-channel fast-kill): no trailer, no worktree delta, and no output growth
+                                // for the fast-kill window — the channel is dead, so conclude NOW
+                                // (journaled) and let the existing no-trailer tail classify and route the attempt.
+                                // The tracker is the growth signal on purpose: raw pane bytes grow on cosmetic repaint
+                                // (an elapsed "9s"→"10s" lengthens the read and would hide a frozen pane). But the
+                                // tracker SHARES the read window's ceiling: its row signal saturates once a sample
+                                // FILLS a PANE_READ_ROWS read on raw lines (blanks/chrome included — see stall.ts),
+                                // and past that point a flat tracker means "unmeasurable", not "dead". For unmetered
+                                // adapters (codex, cursor-agent, grok, opencode — no contextUsage, so the token signal
+                                // never fires) rows are the ONLY
+                                // signal, and those are exactly the non-nudgeable adapters this kill governs — a live
+                                // worker past the ceiling would be concluded dead mid-work. So the kill STANDS DOWN on
+                                // a saturated row signal (journaled once per attempt); the rolling window still owns
+                                // that pane, exactly as pre-T1. Token growth counts as life either way, so a metered
+                                // worker thinking through a long tool run survives.
+                                // The triad has NO status exemption: a pane that herdr reports as blocked, idle, working
+                                // or unknown dies alike once it holds no trailer, no delta and no growth — waiting the
+                                // rolling window out on a status reading is exactly the blindness T1 removes. A matched
+                                // trust dialog is auto-answered above and `continue`s before ever reaching here.
+                                // The NUDGE gets first crack at a nudgeable pane: the fast-kill holds while the daemon
+                                // still has an action of its own pending (un-nudged, or inside the grace window) —
+                                // under the shipped constants (kill 5m < nudge 10m) a delta-less pane would otherwise
+                                // die before the rescue could ever fire. An ANSWERED nudge leaves nothing pending, so
+                                // the hold lifts and the triad governs again. Once a nudge has failed to deliver TWICE
+                                // (one in-slice retry filters a driver flake — see below), the delta clause drops too:
+                                // a delta is past work, and a pane no signal can reach must not buy the rest of the
+                                // window with it (a `working` pane can't be paged either, so this is the only bound
+                                // that path has).
+                                if (stallProgress.rowSignalSaturated && !rowSaturationHeld) {
+                                    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
+                                    && (nudgeFailed || !(await worktreeHasDelta(wt, taskBase)))) {
+                                    journal.append("worker-dead", t.id, { slot: slot.name, attempt, silentMs: sliceNow - lastOutputGrowthAt });
+                                    break;
+                                }
+                                if (nudgeable && !nudged && sliceNow - lastProgressAt >= nudgeAfterSilentMs) {
+                                    nudged = true;
+                                    // T1 review: a false return is a driver-delivery outcome (missing pin, readiness
+                                    // stable-frame timeout, read-back hiccup), not proof of an unreachable channel — so
+                                    // one failure is a flake class, retried once in-slice after a short settle. Only a
+                                    // failed RETRY condemns the channel. Both failures happen inside this slice, so the
+                                    // latch stays immediate and exactly one failure is journaled per attempt.
+                                    let delivered = await driver.nudge(slot, WORKER_NUDGE_MESSAGE);
+                                    if (!delivered) {
+                                        await new Promise((r) => setTimeout(r, NUDGE_REDELIVER_MS));
+                                        delivered = await driver.nudge(slot, WORKER_NUDGE_MESSAGE);
+                                    }
+                                    if (delivered) {
+                                        // absorb the nudge's own echo BEFORE arming the grace timer — post-nudge progress
+                                        // is measured against this baseline, not against the echo.
+                                        const echo = await driver.read(slot, PANE_READ_ROWS);
+                                        stallProgress.observe({ paneText: echo, contextTokens });
+                                        nudgeDeadline = Date.now() + workerNudgeGraceMs;
+                                        journal.append("worker-nudge", t.id, { slot: slot.name, attempt });
+                                    }
+                                    else {
+                                        // T1: an undeliverable nudge is itself a condemnation — the channel is unreachable,
+                                        // so the fast-kill above stops requiring a clean worktree for THIS pane (see there).
+                                        // A blocked/idle pane is still paged below; a `working`/`unknown` one cannot be, and
+                                        // for it the conclusion's own escalation is what notifies. Nothing here waits the
+                                        // rolling window out.
+                                        nudgeFailed = true;
+                                        journal.append("worker-nudge-failed", t.id, { slot: slot.name, attempt });
+                                    }
+                                }
+                                if (nudged && nudgeDeadline !== undefined && sliceNow >= nudgeDeadline && sliceNow - lastProgressAt >= workerNudgeGraceMs) {
+                                    // grace spent, still no post-nudge progress: re-harvest once (the trailer may have
+                                    // landed between polls), then conclude the wait as a stall NOW — the consult sees the
+                                    // un-answered nudge instead of the remainder of the window.
+                                    nudgeDeadline = undefined;
+                                    output = await driver.read(slot, PANE_READ_ROWS);
+                                    finished = new RegExp(trailerPattern(nonce)).test(output);
+                                    const exit = exitRe.exec(output);
+                                    if (finished || exit) {
+                                        exitCode = exit ? Number(exit[1]) : null;
+                                        await sampleContext();
+                                        break;
+                                    }
+                                    journal.append("worker-nudge-expired", t.id, { slot: slot.name, attempt, graceMs: workerNudgeGraceMs });
+                                    lastProgressAt = Date.now() - stallWindowMs; // the existing harvest/classify tail runs unmodified
+                                    continue; // conclude via the loop condition — never page over an acted-on nudge
+                                }
+                                // Unlatched page (T1): the page DECISION fires and is journaled every slice the
+                                // operator is the right actor — i.e. the nudge path doesn't own it (non-allowlisted
+                                // adapter, no nudge surface, or a blocked dialog) or the nudge was attempted and
+                                // FAILED. A nudgeable worker below the silence threshold waits for its nudge; a
+                                // DELIVERED nudge's pending grace suppresses the page — the daemon already acted. An
+                                // ANSWERED nudge has no action pending (the disarm cleared the deadline), so a pane
+                                // that then reads blocked/idle is the operator's again.
+                                // Delivery is unlatched too: the operator is notified again on a status change or
+                                // once pageRepeatMs elapses, so a missed first page is not the last one.
+                                const pageable = (st === "blocked" || st === "idle")
+                                    && (!nudgePending || nudgeFailed);
+                                if (pageable) {
+                                    journal.append("operator-page", t.id, { slot: slot.name, attempt, status: st });
+                                    if (st !== lastPagedStatus || sliceNow - lastPagedAt >= pageRepeatMs) {
+                                        lastPagedStatus = st;
+                                        lastPagedAt = sliceNow;
+                                        const why = st === "blocked" ? "is blocked on a prompt — approve in its pane" : "looks idle without finishing — check its pane";
+                                        await driver.notify(`tickmarkr ${runId}: ${slot.name} ${why}`, { tier: "attention" });
+                                    }
+                                }
+                                // a dead pane or a false-positive marker display returns fast — sleep the unspent slice, never hot-spin
+                                const spent = Date.now() - sliceStart;
+                                if (spent < slice)
+                                    await new Promise((r) => setTimeout(r, Math.min(slice - spent, 1_000)));
+                            }
+                            if (!finished && exitCode === null) {
+                                // timed out (or only ever saw false positives): harvest whatever the pane holds now
+                                timedOut = Date.now() - lastProgressAt >= stallWindowMs;
+                                output = await driver.read(slot, PANE_READ_ROWS);
+                                finished = new RegExp(trailerPattern(nonce)).test(output);
+                                const exit = exitRe.exec(output);
+                                exitCode = exit ? Number(exit[1]) : null;
+                            }
+                            if (finished) {
+                                await driver.waitAgentStatus(slot, "idle", 5_000); // settle, then re-harvest the final render
+                                output = await driver.read(slot, PANE_READ_ROWS);
+                            }
+                            // T5 / OBS-111: an interactive harvest can race the TUI's final paint. When the pane
+                            // contains the nonce token but the JSON hasn't balanced yet, settle and re-read through
+                            // the existing pane-read seam once or twice before recording a malformed-trailer cause.
+                            if (interactive) {
+                                const stallWindowMs = taskTimeoutMinutes * 60_000;
+                                const settleDeadline = attemptStart + stallWindowMs;
+                                const settleDelayMs = 1_000;
+                                const maxSettleRetries = 2;
+                                let settleTries = 0;
+                                settleParsed = adapter.parse(output, nonce);
+                                while (settleParsed.summary === UNPARSEABLE_TRAILER_SUMMARY && settleTries < maxSettleRetries) {
+                                    const remaining = settleDeadline - Date.now();
+                                    if (remaining <= 0)
+                                        break;
+                                    await new Promise((r) => setTimeout(r, Math.min(settleDelayMs, remaining)));
+                                    output = await driver.read(slot, PANE_READ_ROWS);
+                                    settleParsed = adapter.parse(output, nonce);
+                                    settleTries++;
+                                }
+                                if (settleParsed.summary !== UNPARSEABLE_TRAILER_SUMMARY) {
+                                    finished = settleParsed.summary !== NO_TRAILER_SUMMARY;
+                                }
+                            }
                         }
-                        output = seedResult.output;
                     }
                     else {
                         try {
@@ -998,229 +2130,68 @@ export async function runDaemon(repoRoot, opts = {}) {
                                 continue attempts;
                             return;
                         }
-                        output = await driver.read(slot, 1000);
-                    }
-                    if (seedResult?.seedFailed) {
-                        finished = false;
-                    }
-                    else {
-                        let paged = false;
-                        // v1.22 T5 / OBS-19: auto-answer a fingerprint-matched trust dialog exactly once per slot.
-                        // Any other blocked/idle dialog still pages the operator (paged latch below).
-                        let trustAnswered = false;
-                        // OBS-201: one liveness nudge per attempt; the grace deadline is its OWN timer, never the
-                        // stall window (the nudge's pane echo is absorbed before it starts, or the echo itself
-                        // would reset the window and make the early conclusion unreachable).
-                        let nudged = false;
-                        let nudgeDeadline;
-                        finished = false;
-                        exitCode = null;
-                        // OBS-54: reaping keys on new pane output, not dispatch wall clock. Poll at least twice per
-                        // stall window (and at the existing 30s cadence for normal windows) so an active worker resets it.
+                        noteLaunched();
+                        // OBS-54: headless workers have the same output-inactivity budget as visible panes.
+                        // v1.76: same monotonic-progress measure as the interactive site; harvest stays raw.
                         const stallWindowMs = taskTimeoutMinutes * 60_000;
-                        // v1.76: only monotonic work (seed submission, transcript growth, or context growth) resets
-                        // the stall clock. Raw pane differences are terminal chrome until proven otherwise.
-                        let everHadOutput = output.length > 0;
+                        const initialPane = await driver.read(slot, 500);
+                        let everHadOutput = initialPane.length > 0;
                         const stallProgress = new StallProgressTracker();
-                        stallProgress.observe({ paneText: output, seedSubmitted: true, contextTokens });
+                        stallProgress.observe({ paneText: initialPane, seedSubmitted: true, contextTokens });
                         let lastProgressAt = Date.now();
+                        finished = false;
+                        // T2 review (material): the exit marker proves the PROCESS EXITED, never that the worker
+                        // emitted a trailer — the two were the same flag here, so a headless worker that committed
+                        // and exited cleanly without one entered the tail as finished:true, skipping the harvest
+                        // synthesis entirely and reaching gates with the worker's own ok:false and no
+                        // worker-result-harvested row. The interactive site has always kept them apart (`finished`
+                        // there is the trailer regex; the exit marker only sets exitCode), and the cause taxonomy
+                        // already names this shape "clean-exit-no-trailer" — unreachable in print mode until now.
                         while (Date.now() - lastProgressAt < stallWindowMs) {
-                            const sliceStart = Date.now();
-                            const remaining = stallWindowMs - (sliceStart - lastProgressAt);
+                            const remaining = stallWindowMs - (Date.now() - lastProgressAt);
                             let slice = Math.min(BLOCKED_POLL_MS, Math.max(100, Math.min(stallWindowMs / 2, remaining)));
                             if (!everHadOutput) {
-                                const earlyLeft = earlyLaunchLivenessMs - (sliceStart - attemptStart);
+                                const earlyLeft = earlyLaunchLivenessMs - (Date.now() - attemptStart);
                                 if (earlyLeft > 0)
                                     slice = Math.min(slice, earlyLeft);
                             }
-                            if (await driver.waitOutput(slot, `(${trailerPattern(nonce)})|TICKMARKR_EXIT_${nonce}:\\d`, slice, { regex: true })) {
-                                // verify before accepting: a worker that merely DISPLAYS a marker (e.g. editing tickmarkr's
-                                // own source, where "TICKMARKR_EXIT:" is a string literal) must not end the wait. Only a
-                                // parseable trailer or a digit-suffixed exit marker in the harvest is completion.
-                                output = await driver.read(slot, 1000); // TUI transcripts carry chrome — read deeper than print's 500
-                                finished = new RegExp(trailerPattern(nonce)).test(output);
-                                const exit = exitRe.exec(output);
-                                if (finished || exit) {
-                                    exitCode = exit ? Number(exit[1]) : null; // null ⇔ the TUI is still alive
-                                    await sampleContext(); // final poll-seam sample before leaving the wait
-                                    break;
-                                }
+                            // T2 (OBS-264): same probe cadence the interactive loop uses — see harvestSliceMs.
+                            slice = Math.min(slice, harvestSliceMs(Date.now() - lastProgressAt));
+                            if (await driver.waitOutput(slot, `TICKMARKR_EXIT_${nonce}:\\d`, slice, { regex: true })) {
+                                processExited = true;
+                                break;
                             }
-                            const paneText = await driver.read(slot, 1000);
+                            const paneText = await driver.read(slot, 500);
                             if (paneText.length > 0)
                                 everHadOutput = true;
-                            // OBS-117 (v1.71 T6): zero raw output by the early-launch deadline is a dead channel now.
                             if (!everHadOutput && Date.now() - attemptStart >= earlyLaunchLivenessMs) {
                                 earlyLaunchDead = true;
-                                output = paneText;
                                 break;
                             }
-                            // v1.23 T2: piggyback on this poll slice — same cadence as blocked/idle checks, no new timer.
                             await sampleContext();
                             if (stallProgress.observe({ paneText, contextTokens }))
                                 lastProgressAt = Date.now();
-                            // page on "idle" too: herdr's blocked-scrape is strict and proved flaky for TUI dialogs
-                            // (live check: cursor's trust dialog scraped as idle). "unknown" never pages — that's just
-                            // a pane the scraper can't read (subprocess, dead pane); the task timeout covers those.
-                            const st = paged ? "" : await driver.status(slot);
-                            if (!paged && (st === "blocked" || st === "idle")) {
-                                // T5: once-per-slot auto-answer when the adapter declares a trust dialog and the pane
-                                // text matches. tickmarkr created the worktree from the operator's own repo — safe by construction.
-                                if (!trustAnswered && adapter.trustDialog && driver.sendKey) {
-                                    try {
-                                        const paneText = await driver.read(slot, 80);
-                                        if (matchesTrustDialog(paneText, adapter.trustDialog)) {
-                                            trustAnswered = true;
-                                            // v1.25 T1: audit trail for live runs — prove the dialog appeared and was answered.
-                                            // Latch + sendKey + no-page continue stay byte-identical; this append is additive only.
-                                            journal.append("trust-auto-answer", t.id, { slot: slot.name, adapter: adapter.id });
-                                            await driver.sendKey(slot, adapter.trustDialog.key);
-                                            const spent = Date.now() - sliceStart;
-                                            if (spent < slice)
-                                                await new Promise((r) => setTimeout(r, Math.min(slice - spent, 1_000)));
-                                            continue; // do not page — keep waiting for the trailer
-                                        }
-                                    }
-                                    catch {
-                                        /* read/send failed — fall through to page the operator */
-                                    }
-                                }
-                                // OBS-201: the daemon ACTS on an idle pane before paging anyone. Gate: idle AND the
-                                // monotonic tracker silent ≥ the nudge threshold (a worker inside a tool run reads as
-                                // `working` and never gets here; a briefly-settling TUI hasn't been silent long enough).
-                                const nudgeable = st === "idle" && !!driver.nudge && NUDGEABLE_ADAPTERS.has(adapter.id);
-                                if (nudgeable && !nudged) {
-                                    if (Date.now() - lastProgressAt >= nudgeAfterSilentMs) {
-                                        nudged = true;
-                                        if (await driver.nudge(slot, WORKER_NUDGE_MESSAGE)) {
-                                            // absorb the nudge's own echo BEFORE arming the grace timer — post-nudge progress
-                                            // is measured against this baseline, not against the echo.
-                                            const echo = await driver.read(slot, 1000);
-                                            stallProgress.observe({ paneText: echo, contextTokens });
-                                            nudgeDeadline = Date.now() + workerNudgeGraceMs;
-                                            journal.append("worker-nudge", t.id, { slot: slot.name, attempt });
-                                        }
-                                        else {
-                                            journal.append("worker-nudge-failed", t.id, { slot: slot.name, attempt });
-                                            // nudge undeliverable — fall back to today's behavior: page once, window backstop
-                                            paged = true;
-                                            await driver.notify(`tickmarkr ${runId}: ${slot.name} looks idle without finishing — check its pane`, { tier: "attention" });
-                                        }
-                                    }
-                                    // idle but not yet silent past the threshold: neither nudge nor page this slice
-                                }
-                                else if (nudgeable && nudged && nudgeDeadline !== undefined) {
-                                    if (Date.now() >= nudgeDeadline && Date.now() - lastProgressAt >= workerNudgeGraceMs) {
-                                        // grace spent, still idle, no post-nudge progress: re-harvest once (the trailer may
-                                        // have landed between polls), then conclude the wait as a stall NOW — the consult
-                                        // sees the un-answered nudge instead of the remainder of the window.
-                                        nudgeDeadline = undefined;
-                                        output = await driver.read(slot, 1000);
-                                        finished = new RegExp(trailerPattern(nonce)).test(output);
-                                        const exit = exitRe.exec(output);
-                                        if (finished || exit) {
-                                            exitCode = exit ? Number(exit[1]) : null;
-                                            await sampleContext();
-                                            break;
-                                        }
-                                        journal.append("worker-nudge-expired", t.id, { slot: slot.name, attempt, graceMs: workerNudgeGraceMs });
-                                        lastProgressAt = Date.now() - stallWindowMs; // the existing harvest/classify tail runs unmodified
-                                    }
-                                }
-                                else {
-                                    paged = true; // page once — the visible pane is the operator's to unblock; task timeout is the backstop
-                                    const why = st === "blocked" ? "is blocked on a prompt — approve in its pane" : "looks idle without finishing — check its pane";
-                                    await driver.notify(`tickmarkr ${runId}: ${slot.name} ${why}`, { tier: "attention" });
-                                }
-                            }
-                            // a dead pane or a false-positive marker display returns fast — sleep the unspent slice, never hot-spin
-                            const spent = Date.now() - sliceStart;
-                            if (spent < slice)
-                                await new Promise((r) => setTimeout(r, Math.min(slice - spent, 1_000)));
-                        }
-                        if (!finished && exitCode === null) {
-                            // timed out (or only ever saw false positives): harvest whatever the pane holds now
-                            timedOut = Date.now() - lastProgressAt >= stallWindowMs;
-                            output = await driver.read(slot, 1000);
-                            finished = new RegExp(trailerPattern(nonce)).test(output);
-                            const exit = exitRe.exec(output);
-                            exitCode = exit ? Number(exit[1]) : null;
-                        }
-                        if (finished) {
-                            await driver.waitAgentStatus(slot, "idle", 5_000); // settle, then re-harvest the final render
-                            output = await driver.read(slot, 1000);
-                        }
-                        // T5 / OBS-111: an interactive harvest can race the TUI's final paint. When the pane
-                        // contains the nonce token but the JSON hasn't balanced yet, settle and re-read through
-                        // the existing pane-read seam once or twice before recording a malformed-trailer cause.
-                        if (interactive) {
-                            const stallWindowMs = taskTimeoutMinutes * 60_000;
-                            const settleDeadline = attemptStart + stallWindowMs;
-                            const settleDelayMs = 1_000;
-                            const maxSettleRetries = 2;
-                            let settleTries = 0;
-                            settleParsed = adapter.parse(output, nonce);
-                            while (settleParsed.summary === UNPARSEABLE_TRAILER_SUMMARY && settleTries < maxSettleRetries) {
-                                const remaining = settleDeadline - Date.now();
-                                if (remaining <= 0)
-                                    break;
-                                await new Promise((r) => setTimeout(r, Math.min(settleDelayMs, remaining)));
-                                output = await driver.read(slot, 1000);
-                                settleParsed = adapter.parse(output, nonce);
-                                settleTries++;
-                            }
-                            if (settleParsed.summary !== UNPARSEABLE_TRAILER_SUMMARY) {
-                                finished = settleParsed.summary !== NO_TRAILER_SUMMARY;
-                            }
+                            // T2 (OBS-264): the same liveness triad the interactive loop runs. A headless worker that
+                            // committed and went quiet is finished work too, and before this it rode the entire
+                            // window out before anything looked at its commits.
+                            // No nudge hold here, deliberately: print mode has no nudge surface at all (no pane to
+                            // 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.
+                            if (await harvestConcludes(Date.now() - lastProgressAt))
+                                break;
                         }
+                        output = await driver.read(slot, 500);
+                        exitCode = Number(exitRe.exec(output)?.[1] ?? 1);
+                        // Completion is the trailer, exactly as in the interactive loop. A process that exited
+                        // without one is finished:false with a non-null exitCode — the harvest synthesis then owns
+                        // it when the worktree carries work, and classifyWorkerResultCause names it otherwise.
+                        finished = new RegExp(trailerPattern(nonce)).test(output);
+                        timedOut = !processExited && !finished && Date.now() - lastProgressAt >= stallWindowMs;
                     }
                 }
-                else {
-                    try {
-                        await driver.run(slot, paneDispatchCommand(dispatchScript));
-                    }
-                    catch (error) {
-                        if (!(error instanceof DeliveryReadinessError))
-                            throw error;
-                        if (await handleDeliveryReadiness(error))
-                            continue attempts;
-                        return;
-                    }
-                    // OBS-54: headless workers have the same output-inactivity budget as visible panes.
-                    // v1.76: same monotonic-progress measure as the interactive site; harvest stays raw.
-                    const stallWindowMs = taskTimeoutMinutes * 60_000;
-                    const initialPane = await driver.read(slot, 500);
-                    let everHadOutput = initialPane.length > 0;
-                    const stallProgress = new StallProgressTracker();
-                    stallProgress.observe({ paneText: initialPane, seedSubmitted: true, contextTokens });
-                    let lastProgressAt = Date.now();
-                    finished = false;
-                    while (Date.now() - lastProgressAt < stallWindowMs) {
-                        const remaining = stallWindowMs - (Date.now() - lastProgressAt);
-                        let slice = Math.min(BLOCKED_POLL_MS, Math.max(100, Math.min(stallWindowMs / 2, remaining)));
-                        if (!everHadOutput) {
-                            const earlyLeft = earlyLaunchLivenessMs - (Date.now() - attemptStart);
-                            if (earlyLeft > 0)
-                                slice = Math.min(slice, earlyLeft);
-                        }
-                        if (await driver.waitOutput(slot, `TICKMARKR_EXIT_${nonce}:\\d`, slice, { regex: true })) {
-                            finished = true;
-                            break;
-                        }
-                        const paneText = await driver.read(slot, 500);
-                        if (paneText.length > 0)
-                            everHadOutput = true;
-                        if (!everHadOutput && Date.now() - attemptStart >= earlyLaunchLivenessMs) {
-                            earlyLaunchDead = true;
-                            break;
-                        }
-                        await sampleContext();
-                        if (stallProgress.observe({ paneText, contextTokens }))
-                            lastProgressAt = Date.now();
-                    }
-                    output = await driver.read(slot, 500);
-                    exitCode = Number(exitRe.exec(output)?.[1] ?? 1);
-                    timedOut = !finished && Date.now() - lastProgressAt >= stallWindowMs;
+                finally {
+                    await cpuAccountant?.stop();
                 }
                 // SPEND-01 interactive metering race: the harvest loop breaks on the trailer, but the worker
                 // shell may still be running post-trailer bookkeeping (session-store flush, fake usage stamp,
@@ -1232,7 +2203,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                 }
                 // keepPanes retains visible context, not a timed-out subprocess tree. Close before consult/retry
                 // can recreate the worktree; Herdr and subprocesses that reached their exit marker stay unchanged.
-                if (keepOpen && (finished || driver.id !== "subprocess"))
+                if (keepOpen && (finished || processExited || driver.id !== "subprocess"))
                     keptSlots.push(slot);
                 else
                     await closeSlot(slot);
@@ -1247,15 +2218,51 @@ export async function runDaemon(repoRoot, opts = {}) {
                     tokens = addUsage(tokens, attemptUsage);
                     metered++;
                 }
-                const result = settleParsed ?? adapter.parse(output, nonce);
-                const cause = classifyWorkerResultCause({ output, ok: result.ok, finished, exitCode, summary: result.summary, timedOut });
+                let result = settleParsed ?? adapter.parse(output, nonce);
+                const workerFinished = finished;
+                const workerCause = classifyWorkerResultCause({ output, ok: result.ok, finished, exitCode, summary: result.summary, timedOut });
                 journal.append("worker-result", t.id, {
-                    ok: result.ok, summary: result.summary, deviations: result.deviations, finished, exitCode,
-                    mode: interactive ? "interactive" : "print", ...(cause ? { cause } : {}),
+                    ok: result.ok, summary: result.summary, deviations: result.deviations, finished: workerFinished, exitCode,
+                    mode: interactive ? "interactive" : "print", ...(workerCause ? { cause: workerCause } : {}),
                 });
-                if (result.ok && finished)
+                // T2 review (routing precedence): provider-death, quota and dead-channel classification
+                // derive from ONE rule — the worker's OWN outcome, workerFinished and this PRE-HARVEST
+                // parse — never from the synthesized result below. A worker that committed and then walled
+                // (provider outage, quota banner, dead CLI) must still route; the harvest synthesis only
+                // decides whether THIS attempt's worktree goes to gates, and when routing wins instead, the
+                // commits survive via the existing commitsToCarry/cherryPickCommits carry-forward.
+                const preHarvestResult = result;
+                // T2 (OBS-264): recognize committed no-trailer work BEFORE any no-trailer streak, provider,
+                // quota or dead-channel routing. Gates never trusted the trailer, so this successful synthesis
+                // must enter exactly where a worker-claimed ok enters. Preserve the parsed worker truth in the
+                // worker-result row above and name the synthesized gate input in its own harvested event.
+                let harvestedCommits = [];
+                if (!workerFinished) {
+                    harvestedCommits = await commitsAheadOf(taskBase, wt);
+                    if (harvestedCommits.length > 0) {
+                        result = { ok: true, summary: HARVESTED_RESULT_SUMMARY, deviations: [], raw: output };
+                        finished = true;
+                        journal.append("worker-result-harvested", t.id, {
+                            attempt, commits: harvestedCommits, summary: HARVESTED_RESULT_SUMMARY, source: "harvest",
+                        });
+                    }
+                }
+                // T2 review: provider-death is the THIRD routing branch that must read the pre-harvest
+                // outcome — the synthesis must not null it. A worker that committed and then printed the
+                // outage banner without exiting (workerFinished false, so the harvest fires) still takes
+                // the capped same-channel requeue below; nulling the cause here would skip that branch and
+                // let classifyDeadChannel(preHarvestResult) demote the channel run-wide on a transient blip.
+                const cause = harvestedCommits.length > 0 && workerCause !== "provider-death" ? undefined : workerCause;
+                // T2 review (family): the no-trailer streak is accounted on the SAME ONE rule the routing
+                // branches above use — workerFinished and the PRE-HARVEST parse — never the synthesized
+                // result. A channel that commits but never emits a parseable trailer still burned a
+                // no-trailer window (OBS-57): the synthesis decides whether THIS worktree goes to gates, it
+                // never certifies the channel. Reading the synthesized `finished`/`ok` here reset the streak
+                // on every harvest, so a CLI that produces commits and swallows every trailer was immune to
+                // the two-window demotion and stayed first pick for the rest of the run.
+                if (preHarvestResult.ok && workerFinished)
                     noTrailerStreak.set(channelKey(assignment), 0);
-                else if (!finished && cause !== "provider-death") {
+                else if (!workerFinished && cause !== "provider-death") {
                     const ck = channelKey(assignment);
                     const streak = (noTrailerStreak.get(ck) ?? 0) + 1;
                     noTrailerStreak.set(ck, streak);
@@ -1275,8 +2282,19 @@ export async function runDaemon(repoRoot, opts = {}) {
                 }
                 // quota exhaustion → failover within floor; does NOT consume the ladder (spec §4)
                 // print: guarded on exit code — exit-0 output that merely MENTIONS "rate limit" must not failover
-                // interactive: a harvested trailer beats quota mentions; without one, quota text fails over (spec v1.2 §2)
-                const quotaHit = (interactive ? !finished : exitCode !== 0) && QUOTA_RE.test(output);
+                // interactive: a worker-CLAIMED trailer beats quota mentions; without one, quota text fails over
+                // (spec v1.2 §2) — matched on the chrome-filtered tail, the exact discrimination the in-loop
+                // classifier makes, so the two can never disagree. A TUI harvest is the whole retained pane:
+                // an unscoped match failed a worker over for quoting "rate limit" in its own diff (tail
+                // scoping kills that — the mention sits ABOVE the tail), and a raw-tail match fires on fixed
+                // chrome (codex's welcome line — filtered by identity, so a launch-time banner this backstop
+                // exists to catch is never exculpated). Print output keeps the exit-code guard.
+                // T2 review: the gate is `workerFinished`, not the harvest-synthesized `finished` — a
+                // committed-but-quota-walled attempt routes here FIRST (its commits ride the carry-forward
+                // into the next attempt's recreated worktree), it never buys a gate run on throttled work.
+                const quotaHit = interactive
+                    ? !workerFinished && QUOTA_RE.test(stallSnapshotBannerRows(output))
+                    : exitCode !== 0 && QUOTA_RE.test(output);
                 if (quotaHit) {
                     const next = failover("quota-failover");
                     journal.append("quota-failover", t.id, { from: channelKey(assignment), to: next ? channelKey(next) : null });
@@ -1316,7 +2334,10 @@ export async function runDaemon(repoRoot, opts = {}) {
                 // cap above is spent, so a transient blip still recovers in place.
                 // OBS-117 (v1.71 T6): a silent launch failure has no CLI signature to parse — the same
                 // setup-required typed dead-channel path a late-harvest "command not found" would take.
-                const dead = classifyDeadChannel(result) ?? (earlyLaunchDead ? "setup-required" : undefined);
+                // T2 review: classify the PRE-HARVEST parse — classifyDeadChannel bails on any ok:true
+                // result, so reading the synthesized harvest result would swallow auth-required /
+                // setup-required / provider-outage for every committed-but-walled attempt (in both modes).
+                const dead = classifyDeadChannel(preHarvestResult) ?? (earlyLaunchDead ? "setup-required" : undefined);
                 if (dead) {
                     const from = channelKey(assignment);
                     demotedChannels.add(from); // excluded for later attempts AND later tasks in this run
@@ -1382,33 +2403,36 @@ export async function runDaemon(repoRoot, opts = {}) {
                 }
                 const onGate = async (e) => {
                     if (e.phase === "start") {
-                        journal.phaseStart(t.id, phaseForGate(e.gate), { gate: e.gate, index: e.index, total: e.total });
+                        notePhaseStart(e);
+                        journal.phaseStart(t.id, phaseForGate(e.gate), { gate: e.gate, index: e.index, total: e.total, ...(e.parentAt === undefined ? {} : { parallel: true }) });
                         return;
                     }
                     const g = e.result;
-                    // GATE-09 (ROADMAP SC-4): journal every judge retry as an attributable event — which gate flaked,
-                    // which channel flaked, which channel retried — so `tickmarkr journal`/report can distinguish "judge
-                    // flaked, retried" from "worker failed" (run-20260711-185020 P43-03 L70-72 billed a judge flake as
-                    // a worker attempt; 47-01 fixed WHO retries, this closes the audit-trail half). The condition is
-                    // META-ONLY (D-03): gate === "acceptance" + typeof-shape guards on meta.judgeRetry — never a
-                    // details-regex. The v1.1 review regex below is grandfathered, not precedent. Appended BEFORE the
-                    // gate-result so attribution precedes the verdict in the stream. secondUnparseable is derived from
-                    // the final result's meta.unparseable (set by run-gates when the retry ALSO flaked — double-garbage).
-                    if (g.gate === "acceptance" && typeof g.meta?.judgeRetry === "object" && g.meta.judgeRetry !== null) {
-                        const jr = g.meta.judgeRetry;
-                        if (typeof jr.flaked === "string" && typeof jr.retried === "string") {
-                            journal.append("judge-retry", t.id, {
-                                gate: "acceptance", flaked: jr.flaked, retried: jr.retried,
-                                ...(g.meta.unparseable === true ? { secondUnparseable: true } : {}),
-                            });
+                    inParallelOrder(g.gate, () => {
+                        // GATE-09 (ROADMAP SC-4): journal every judge retry as an attributable event — which gate flaked,
+                        // which channel flaked, which channel retried — so `tickmarkr journal`/report can distinguish "judge
+                        // flaked, retried" from "worker failed" (run-20260711-185020 P43-03 L70-72 billed a judge flake as
+                        // a worker attempt; 47-01 fixed WHO retries, this closes the audit-trail half). The condition is
+                        // META-ONLY (D-03): gate === "acceptance" + typeof-shape guards on meta.judgeRetry — never a
+                        // details-regex. The v1.1 review regex below is grandfathered, not precedent. Appended BEFORE the
+                        // gate-result so attribution precedes the verdict in the stream. secondUnparseable is derived from
+                        // the final result's meta.unparseable (set by run-gates when the retry ALSO flaked — double-garbage).
+                        if (g.gate === "acceptance" && typeof g.meta?.judgeRetry === "object" && g.meta.judgeRetry !== null) {
+                            const jr = g.meta.judgeRetry;
+                            if (typeof jr.flaked === "string" && typeof jr.retried === "string") {
+                                journal.append("judge-retry", t.id, {
+                                    gate: "acceptance", flaked: jr.flaked, retried: jr.retried,
+                                    ...(g.meta.unparseable === true ? { secondUnparseable: true } : {}),
+                                });
+                            }
                         }
-                    }
-                    journal.append("gate-result", t.id, { gate: g.gate, pass: g.pass, details: g.details, ...(g.meta?.skipped === true ? { skipped: true } : {}) });
-                    noteReviewRetry(g);
-                    // v1.1 failover: never re-ask a reviewer channel that produced garbage for this task
-                    if (g.gate === "review" && !g.pass && /unparseable/.test(g.details) && typeof g.meta?.reviewer === "string") {
-                        badReviewers.push(g.meta.reviewer);
-                    }
+                        journalGateResult(g);
+                        noteReviewRetry(g);
+                        // v1.1 failover: never re-ask a reviewer channel that produced garbage for this task
+                        if (g.gate === "review" && !g.pass && /unparseable/.test(g.details) && typeof g.meta?.reviewer === "string") {
+                            badReviewers.push(g.meta.reviewer);
+                        }
+                    });
                 };
                 let results = [];
                 let commits = [];
@@ -1418,6 +2442,7 @@ export async function runDaemon(repoRoot, opts = {}) {
                     ({ results, commits } = await runGates(t, {
                         worktree: wt, baseRef: taskBase, result, author: assignment,
                         commands, baseline, channels, adapters, cfg, artifactDir: journal.dir,
+                        pipeline: "v185", selectTests: !testGateFailed,
                         via: cfg.visibility.llm === "pane"
                             ? {
                                 driver: trackedDriver,
@@ -1439,7 +2464,9 @@ export async function runDaemon(repoRoot, opts = {}) {
                     }));
                     graph = addEvidence(graph, t.id, { commits, gateResults: results, artifacts: [promptFile] });
                     saveGraph(repoRoot, graph);
-                    if (results.every((g) => g.pass)) {
+                    if (results.some((g) => g.gate === "test" && !g.pass))
+                        testGateFailed = true;
+                    if (results.every(gateSatisfied)) {
                         const m = await mergeSerial(taskBranch, t, gated);
                         if (m.tipMoved) {
                             journal.append("tip-moved", t.id, m.tipMoved);
@@ -1484,21 +2511,113 @@ export async function runDaemon(repoRoot, opts = {}) {
                 // 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((g) => !g.pass).map((g) => `${g.gate}: ${g.details}`).join("\n\n");
+                feedback = results.filter(gateFailed).map((g) => `${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.
                 // Unparseable verdicts (already retried in-gate, OBS-193) and diff-cap trips (the diff cannot
                 // shrink by retrying, OBS-48) fall through to the ladder unchanged. Review runs last, so a
                 // failed review with every other gate green is exactly "the work landed, the reviewer objects".
-                const reviewFail = results.find((g) => g.gate === "review" && !g.pass);
+                const reviewFail = results.find((g) => g.gate === "review" && gateFailed(g));
                 const reviewFixRetry = reviewFail !== undefined
                     && reviewFail.meta?.unparseable !== true
                     && !isDiffCapPark(reviewFail)
-                    && results.every((g) => g.pass || g.gate === "review");
-                const step = reviewFixRetry ? "retry" : r.ladder[Math.min(ladderIdx++, r.ladder.length - 1)];
-                journal.append("escalation", t.id, { step, attempt: attempt + 1, ...(reviewFixRetry ? { reviewFix: true } : {}) });
-                await driver.notify(`tickmarkr ${runId}: ${t.id} escalation: ${step}`, { tier: "attention" });
+                    && results.every((g) => gateSatisfied(g) || g.gate === "review");
+                const failing = results.filter(gateFailed);
+                // Decided before the cap, because the cap's question is whether the NEXT move would re-buy a
+                // measurement already made — and a funded repair is one of the moves that would.
+                const landed = await commitsAheadOf(taskBase, wt);
+                // T4 (OBS-265): judge and review are now ONE round, so a round can report both failing where the
+                // serial walk returned at the judge and review never ran. Eligibility is scored on the battery
+                // the serial contract would have surfaced — review only speaks for a round nothing else failed —
+                // so removing the waiting does not re-price the ladder. The journal below still names every
+                // failing gate, and `feedback` still carries every one of them to the next attempt.
+                const repairBattery = failing.some((g) => g.gate !== "review") ? failing.filter((g) => g.gate !== "review") : failing;
+                const repairable = narrowRepairBattery(repairBattery) && lostCommits.length === 0 && landed.length > 0;
+                const repairsDrawn = repairsSinceApproval(journal.read(), t.id);
+                const repair = repairable && repairsDrawn < MAX_REPAIRS;
+                // v1.85 T3: the fingerprint cap. Two normalized-identical failures of one DETERMINISTIC gate on
+                // one task (volatile tokens — worktree prefixes, line refs, durations, run ids — are not
+                // information) mean the round about to be bought is a re-measurement: ~663m across 5 runs went to
+                // exactly this loop. The threshold is a property of the FAILURE, never of the move that would
+                // follow it, so it is evaluated on EVERY such gate at EVERY ladder position — including a rung
+                // that would change channel, and including a review-fix round. Conditioning it on the next rung
+                // was the first shape of this and let a second identical failure buy an escalate/consult round
+                // the criterion says it may not buy. What the cap does NOT reach is an LLM verdict, which is a
+                // different object with its own tighter bound (see isDeterministicFailure).
+                //
+                // It fires on the CROSSING, not as a latch: the consult it forces and the ban it sets govern the
+                // next move, so re-firing on the third identical failure would only re-buy the round it just paid
+                // for — and the ladder, whose rung this failure still spends, bounds the rest.
+                const repeated = failing.find((g) => isDeterministicFailure(g)
+                    && identicalGateFailures(journal.read(), t.id, g.gate, normalizeGateFailure(g.details)) === GATE_FINGERPRINT_CAP);
+                // The rung the cap spent, when it fired — the move below executes THIS instead of drawing a
+                // second one, so a cap costs exactly the rung the failure would have cost anyway.
+                let capStep;
+                if (repeated) {
+                    const normalized = normalizeGateFailure(repeated.details);
+                    journal.append("gate-fingerprint-cap", t.id, {
+                        gate: repeated.gate,
+                        occurrences: GATE_FINGERPRINT_CAP,
+                        fingerprint: normalized.slice(0, 500),
+                        retrySameBanned: true,
+                        channel: channelKey(assignment), // the ban is bound to the channel that produced the repeat
+                        attempt: attempt + 1,
+                    });
+                    await driver.notify(`tickmarkr ${runId}: ${t.id} ${repeated.gate} failed identically twice — consulting, identical retry banned`, { tier: "attention" });
+                    // The cap takes the ladder's MOVE, never its accounting: this failure still spends the rung it
+                    // would have spent, so a task that cannot converge still reaches ladder exhaustion on exactly
+                    // the budget it always had and the cap can never hand a stuck task extra rounds.
+                    capStep = r.ladder[Math.min(ladderIdx++, r.ladder.length - 1)];
+                    journal.append("escalation", t.id, { step: capStep, attempt: attempt + 1, fingerprintCap: true });
+                    await driver.notify(`tickmarkr ${runId}: ${t.id} escalation: ${capStep}`, { tier: "attention" });
+                    const v = await runConsult("gate-fail-repeat", output, feedback, results);
+                    // Rule: a terminal cap consult vetoes a same-channel retry, but cannot veto an `escalate`
+                    // rung that already satisfies retry-same-banned by changing channel. Thus terminal+retry
+                    // parks through the shared verdict boundary, while terminal+escalate records the consult as
+                    // advisory and executes the already-spent rung below. Recoverable verdicts still control the
+                    // move directly. The paired fixture asserts both directions of this boundary.
+                    const recoverable = v.action === "retry" || v.action === "reroute";
+                    if (recoverable || capStep !== "escalate") {
+                        if (await applyVerdict(v, attempt + 1, "gate-fail"))
+                            continue;
+                        return;
+                    }
+                    journal.append("consult-verdict", t.id, { action: v.action, notes: v.notes, capAdvisory: true });
+                }
+                // v1.85 T3: a narrow battery over fully carried commits earns a REPAIR (decided above) — the
+                // next dispatch carries the findings verbatim and the diff content instead of re-onboarding a
+                // fresh worker. Budget is engagement-scoped and journal-derived, so a resume inherits it rather
+                // than refunding it; the third repair-eligible failure falls back to the fresh ladder.
+                //
+                // `landed` is measured from the worktree rather than read from runGates: runGates returns at
+                // the first failure, so a red test or lint gate never reaches the evidence stage and its
+                // `commits` come back empty — reading them would make the ruling's test/lint case unreachable.
+                //
+                // Budget spent means the FRESH LADDER owns this failure — including a review-only one, whose
+                // same-channel fix retry is exactly the round the budget just declared too expensive to repeat.
+                const repairExhausted = repairable && !repair;
+                if (repair && !capStep) {
+                    journal.append("repair-attempt", t.id, {
+                        repair: repairsDrawn + 1, of: MAX_REPAIRS, gates: failing.map((g) => g.gate),
+                        commits: landed.length,
+                        findings: feedback, // the failure bytes this repair must carry, replayable across a resume
+                    });
+                }
+                else if (repairExhausted && !capStep) {
+                    journal.append("repair-exhausted", t.id, { repairs: repairsDrawn, of: MAX_REPAIRS, gates: failing.map((g) => g.gate) });
+                }
+                const step = capStep ?? (repair || (reviewFixRetry && !repairExhausted)
+                    ? "retry"
+                    : r.ladder[Math.min(ladderIdx++, r.ladder.length - 1)]);
+                if (!capStep) { // a capped failure already journaled and announced the rung it spent
+                    journal.append("escalation", t.id, {
+                        step, attempt: attempt + 1,
+                        ...(reviewFixRetry && !repairExhausted ? { reviewFix: true } : {}),
+                        ...(repair ? { repair: repairsDrawn + 1 } : {}),
+                    });
+                    await driver.notify(`tickmarkr ${runId}: ${t.id} escalation: ${step}`, { tier: "attention" });
+                }
                 if (step === "retry")
                     continue;
                 if (step === "escalate") {
@@ -1574,24 +2693,7 @@ export async function runDaemon(repoRoot, opts = {}) {
         // OBS-34: post-merge integration-tip verify — strict exit codes, no baseline forgiveness.
         const lastMergedTask = [...journal.read()].reverse().find((e) => e.event === "merge" && e.taskId)?.taskId;
         if (summary.done.length > 0 && Object.keys(commands).length > 0) {
-            const tipResults = await verifyIntegrationTip(intWt, commands, journal.dir);
-            let tipFailed = false;
-            for (const r of tipResults) {
-                if (r.pass) {
-                    journal.append("tip-verify", undefined, { gate: r.gate, cmd: r.cmd, pass: true, exitCode: r.exitCode, details: r.details });
-                }
-                else {
-                    journal.append("tip-verify-failed", undefined, {
-                        gate: r.gate,
-                        cmd: r.cmd,
-                        exitCode: r.exitCode,
-                        fingerprints: r.fingerprints,
-                        artifact: r.artifact,
-                        lastMergedTask,
-                    });
-                    tipFailed = true;
-                }
-            }
+            const tipFailed = await verifyIntegrationTipCached(intWt, commands, journal, { lastMergedTask });
             summary.tipVerify = tipFailed ? "failed" : "passed";
             if (tipFailed && lastMergedTask)
                 summary.lastMergedTask = lastMergedTask;