tickmarkr 2.6.2 → 2.6.3

package/dist/gates/baseline.js

@@ -1,25 +1,81 @@
 import { createHash, randomUUID } from "node:crypto";
 import { execFileSync } from "node:child_process";
-import { existsSync, readFileSync, mkdirSync, readdirSync, statSync, unlinkSync, writeFileSync, rmdirSync } from "node:fs";
+import { EVICTION_TOMBSTONE_LIMIT, EVICTION_TOMBSTONES_FILE, parseEvictionTombstones } from "../run/receipt-resolver.js";
+import { existsSync, readFileSync, mkdirSync, readdirSync, renameSync, statSync, unlinkSync, writeFileSync, rmdirSync } from "node:fs";
 import { availableParallelism, loadavg, tmpdir } from "node:os";
-import { join } from "node:path";
+import { basename, join } from "node:path";
+import { DEFAULT_EVIDENCE_QUOTA_BYTES } from "../config/config.js";
 import { dependencyLinkRefusal, DEFAULT_SHELL_TIMEOUT_MS, describeCapacity, sameCapacity, sh } from "../run/git.js";
 import { executionSignal } from "../run/execution-budget.js";
 import { isVitestTestCommand, manifestFileCount } from "./test-manifest.js";
 export const EVIDENCE_TAIL_BYTES = 16 * 1024;
-export const EVIDENCE_RUN_QUOTA_BYTES = 8 * 1024 * 1024;
-export function redactGateOutput(text, env) {
-    // One pass over the original bytes: replacing an environment value must not break a token
-    // recognizer, and a short environment value must not rewrite the redaction marker itself.
+/** The default per-run quota; `gates.evidenceQuotaBytes` in config overrides it (OBS-1140). */
+export const EVIDENCE_RUN_QUOTA_BYTES = DEFAULT_EVIDENCE_QUOTA_BYTES;
+/** Benign environment locations substituted by name; every other redaction is material. */
+const BENIGN_ENV_KEYS = ["HOME", "TMPDIR"];
+const REDACTION_CATEGORIES = ["token", "assignment", "secretEnv", "benignEnv"];
+/**
+ * OBS-1139: every category scans the ORIGINAL text on its own; overlapping spans then merge into one
+ * span counted once under its highest-priority member — token, assignment, secret environment, benign
+ * HOME/TMPDIR — so precedence never depends on which match starts first. Counts only: never a value.
+ */
+export function classifyGateOutput(text, env) {
+    // HOME/TMPDIR substitute by name at any length; "/" alone would rewrite every path separator.
+    const benign = BENIGN_ENV_KEYS.flatMap(key => { const value = env[key]; return value && value.length > 1 ? [{ value, label: `$${key}` }] : []; });
     // Short operational values (0, true, vi, test) occur throughout ordinary diagnostics.
-    // Only secret-named keys justify redacting short values.
-    const values = [...new Set(Object.entries(env)
-            .filter(([key, value]) => !!value && (value.length >= 8 || /KEY|TOKEN|SECRET|PASSWORD|CREDENTIAL|AUTH/i.test(key)))
-            .map(([, value]) => value))]
-        .sort((a, b) => b.length - a.length).map(v => v.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"));
+    // Only secret-named keys justify redacting short values. A secret-named value that happens to equal
+    // HOME/TMPDIR stays a secret candidate: precedence decides the span, never candidate pruning.
+    const secrets = [...new Set(Object.entries(env)
+            .filter(([key, value]) => !!value && !BENIGN_ENV_KEYS.includes(key)
+            && (value.length >= 8 || /KEY|TOKEN|SECRET|PASSWORD|CREDENTIAL|AUTH/i.test(key)))
+            .map(([, value]) => value))];
+    const literal = (v) => v.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
     const tokens = String.raw `\b(?:sk-[A-Za-z0-9_-]{12,}|gh[pousr]_[A-Za-z0-9_]{16,}|github_pat_[A-Za-z0-9_]{16,}|AKIA[A-Z0-9]{16}|eyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+)\b`;
     const assignments = String.raw `(?:authorization\s*:\s*bearer|(?:api[_-]?key|token|password|secret)\s*[:=])\s*[^\s"']+`;
-    return text.replace(new RegExp([tokens, assignments, ...values].join("|"), "gi"), "[REDACTED]");
+    // Every value scans the original text on its own: an alternation would consume the first match and
+    // skip a longer overlapping value's tail, so precedence is resolved on the collected spans instead.
+    const scan = (source, rank, label) => [...text.matchAll(new RegExp(source, "gi"))].map(m => ({ start: m.index, end: m.index + m[0].length, rank, label }));
+    const spans = [
+        ...scan(tokens, 0), ...scan(assignments, 1),
+        ...secrets.flatMap(v => scan(literal(v), 2)),
+        ...benign.flatMap(b => scan(literal(b.value), 3, b.label)),
+    ].sort((a, b) => a.start - b.start || b.end - a.end);
+    const merged = [];
+    for (const span of spans) {
+        const last = merged.at(-1);
+        if (!last || span.start >= last.end) {
+            merged.push({ ...span });
+            continue;
+        }
+        // C-11 (D-673): a benign value nested in (or identical to) another is one location, named by the containing
+        // span, which sorts first. Only a PARTIAL overlap of two different benign values is ambiguous: withhold it.
+        if (last.rank === 3 && span.rank === 3 && last.label !== span.label && span.end > last.end)
+            last.rank = 2;
+        else
+            last.rank = Math.min(last.rank, span.rank);
+        last.end = Math.max(last.end, span.end);
+    }
+    const counts = { token: 0, assignment: 0, secretEnv: 0, benignEnv: 0 };
+    let out = "", at = 0;
+    for (const { start, end, rank, label } of merged) {
+        counts[REDACTION_CATEGORIES[rank]]++;
+        out += text.slice(at, start) + (rank === 3 ? label : "[REDACTED]");
+        at = end;
+    }
+    return { text: out + text.slice(at), counts };
+}
+export function redactGateOutput(text, env) {
+    return classifyGateOutput(text, env).text;
+}
+/** Material means bytes were withheld; a benign location substitution withholds nothing. */
+export const redactionMaterial = (c) => c.token + c.assignment + c.secretEnv > 0;
+function readEvictionTombstones(root) {
+    try {
+        return parseEvictionTombstones(readFileSync(join(root, EVICTION_TOMBSTONES_FILE)));
+    }
+    catch {
+        return [];
+    }
 }
 /** Resolve a snapshot reference against actual bytes: eviction never fabricates a live artifact. */
 export function resolveEvidenceArtifact(root, ref) {
@@ -75,7 +131,12 @@ export function beginGateEvidence(cwd, gate, command, opts = {}, nonce) {
                 exitCode: observed?.exitCode ?? null, signal: observed?.signal ?? null,
                 timedOut: observed ? observed.outcome === "timed-out" : null,
             };
-            const clean = [redactGateOutput(stdout, env), redactGateOutput(stderr, env)];
+            const classified = [classifyGateOutput(stdout, env), classifyGateOutput(stderr, env)];
+            const clean = classified.map(c => c.text);
+            const counts = { token: 0, assignment: 0, secretEnv: 0, benignEnv: 0 };
+            for (const c of classified)
+                for (const k of Object.keys(counts))
+                    counts[k] += c.counts[k];
             const refs = clean.map((text, i) => {
                 const bytes = Buffer.from(text);
                 const tail = bytes.subarray(-EVIDENCE_TAIL_BYTES);
@@ -94,19 +155,38 @@ export function beginGateEvidence(cwd, gate, command, opts = {}, nonce) {
                     lock = join(dir, ".write-lock");
                     const quota = opts.quotaBytes !== undefined && Number.isFinite(opts.quotaBytes)
                         ? Math.max(0, Math.floor(opts.quotaBytes)) : EVIDENCE_RUN_QUOTA_BYTES;
+                    // An empty artifact costs no quota, so it is never evicted: every eviction frees bytes.
                     const files = readdirSync(dir).filter(f => /^[a-zA-Z0-9-]+-(stdout|stderr)\.log$/.test(f)).map(f => ({ path: join(dir, f), stat: statSync(join(dir, f)) }))
-                        .sort((a, b) => a.stat.mtimeMs - b.stat.mtimeMs || a.path.localeCompare(b.path));
+                        .filter(f => f.stat.size > 0).sort((a, b) => a.stat.mtimeMs - b.stat.mtimeMs || a.path.localeCompare(b.path));
                     let total = files.reduce((sum, f) => sum + f.stat.size, 0);
                     const incoming = refs.reduce((sum, ref) => sum + ref.retainedBytes, 0);
-                    if (incoming > quota)
-                        throw new Error("evidence quota cannot retain this invocation");
-                    while (files.length && total + incoming > quota) {
+                    // OBS-1140: the quota binds what is already retained too, so a resumed run at zero quota keeps
+                    // zero bytes. An invocation larger than the whole quota is expired at birth; older artifacts
+                    // are evicted only as far as the quota itself demands, never for bytes that could not fit.
+                    const fits = incoming <= quota;
+                    const evicted = [];
+                    while (files.length && total + (fits ? incoming : 0) > quota) {
                         const oldest = files.shift();
+                        const bytes = readFileSync(oldest.path);
                         unlinkSync(oldest.path);
                         total -= oldest.stat.size;
+                        evicted.push({ path: `gate-evidence/${basename(oldest.path)}`, sha256: createHash("sha256").update(bytes).digest("hex"), retainedBytes: bytes.length });
+                    }
+                    if (evicted.length) {
+                        // Identity-bound (path + hash + length) and bounded: the newest 256 survive a restart.
+                        const tombstones = join(root, EVICTION_TOMBSTONES_FILE);
+                        writeFileSync(`${tombstones}.tmp`, JSON.stringify([...readEvictionTombstones(root), ...evicted].slice(-EVICTION_TOMBSTONE_LIMIT)));
+                        renameSync(`${tombstones}.tmp`, tombstones);
                     }
-                    for (const [i, ref] of refs.entries()) {
-                        (opts.write ?? writeFileSync)(join(root, ref.path), Buffer.from(clean[i]).subarray(-EVIDENCE_TAIL_BYTES));
+                    if (!fits) {
+                        availability = "expired";
+                        for (const ref of refs)
+                            ref.availability = "expired";
+                    }
+                    else {
+                        for (const [i, ref] of refs.entries()) {
+                            (opts.write ?? writeFileSync)(join(root, ref.path), Buffer.from(clean[i]).subarray(-EVIDENCE_TAIL_BYTES));
+                        }
                     }
                 }
                 catch {
@@ -131,7 +211,7 @@ export function beginGateEvidence(cwd, gate, command, opts = {}, nonce) {
             }
             return { invocationId, ...(nonce ? { nonce } : {}),
                 subject: { runId: opts.runId ?? "standalone", taskId: opts.taskId ?? null, attempt: opts.attempt ?? null, gate, subjectCommit },
-                termination, availability, redaction: { material: clean[0] !== stdout || clean[1] !== stderr }, stdout: refs[0], stderr: refs[1] };
+                termination, availability, redaction: { material: redactionMaterial(counts), counts }, stdout: refs[0], stderr: refs[1] };
         },
     };
 }

package/dist/gates/cache.d.ts

@@ -9,8 +9,15 @@ export declare function lockfileHash(worktree: string): string;
 export declare function canonicalJson(obj: unknown): string;
 export declare function baselineIdentity(baseline?: Baseline): string;
 export declare function getWorktreeTree(worktree: string): Promise<string>;
+/** OBS-635: environment inputs a runner child reads that capacity and lifecycle do not already bind.
+ * ponytail: a named list, not the whole env (pids and terminal vars would defeat every reuse); extend
+ * it when another variable is shown to change what a runner executes. */
+export declare const RUNNER_ENV_KEYS: readonly ["PATH", "NODE_OPTIONS", "NODE_PATH", "NODE_ENV", "TZ"];
+export declare function runnerInputsHash(env?: NodeJS.ProcessEnv): string;
 export interface GateEnvironmentInput {
     nodeRuntime?: string;
+    /** the runner inputs hash; defaults to this process's (runnerInputsHash) */
+    runner?: string;
     lockfile?: string;
     worktree?: string;
     capacity?: RunCapacity;
@@ -21,6 +28,7 @@ export interface GateEnvironmentInput {
 }
 export interface EnvironmentParts {
     nodeRuntime: string;
+    runner: string;
     lockfile: string;
     capacity: RunCapacity;
     selectedSet?: readonly string[];

package/dist/gates/cache.js

@@ -77,8 +77,17 @@ export async function getWorktreeTree(worktree) {
         rmSync(scratch, { recursive: true, force: true });
     }
 }
+/** OBS-635: environment inputs a runner child reads that capacity and lifecycle do not already bind.
+ * ponytail: a named list, not the whole env (pids and terminal vars would defeat every reuse); extend
+ * it when another variable is shown to change what a runner executes. */
+export const RUNNER_ENV_KEYS = ["PATH", "NODE_OPTIONS", "NODE_PATH", "NODE_ENV", "TZ"];
+export function runnerInputsHash(env = process.env) {
+    return createHash("sha256").update(canonicalJson(Object.fromEntries(RUNNER_ENV_KEYS.map((k) => [k, env[k] ?? null]))))
+        .digest("hex").slice(0, 16);
+}
 export function environmentFingerprint(env) {
     const nodeRuntime = env.nodeRuntime ?? process.version;
+    const runner = env.runner ?? runnerInputsHash();
     const lockfile = env.lockfile ?? (env.worktree ? lockfileHash(env.worktree) : "no-lockfile");
     const cap = env.capacity ?? resolvedCapacity();
     const capacity = { forkCap: cap.forkCap, cores: cap.cores };
@@ -97,6 +106,7 @@ export function environmentFingerprint(env) {
     // explicit `false` and an npmrc `false` are the same policy for the child that ran.
     const payload = canonicalJson({
         nodeRuntime,
+        runner,
         lockfile,
         capacity,
         resolution,
@@ -107,7 +117,7 @@ export function environmentFingerprint(env) {
     const fingerprint = createHash("sha256").update(payload).digest("hex").slice(0, 16);
     return {
         fingerprint,
-        parts: { nodeRuntime, lockfile, capacity, selectedSet, verification },
+        parts: { nodeRuntime, runner, lockfile, capacity, selectedSet, verification },
     };
 }
 export async function computeVerificationIdentity(params) {
@@ -151,7 +161,7 @@ export function verificationIdentityKey(id) {
 export function formatReusedDetails(originalDetails, id) {
     const unadorned = originalDetails.replace(/^reused verdict \(identity: [^)]+\):\s*/, "");
     const envDesc = id.envParts
-        ? ` [node=${id.envParts.nodeRuntime}, lockfile=${id.envParts.lockfile}, capacity=${describeCapacity(id.envParts.capacity)}${id.envParts.selectedSet ? `, selected=${id.envParts.selectedSet.join(",")}` : ""}, protocol=${id.envParts.verification.protocol}, lifecycle=${id.envParts.verification.lifecycle} (${id.envParts.verification.source})]`
+        ? ` [node=${id.envParts.nodeRuntime}, runner=${id.envParts.runner}, lockfile=${id.envParts.lockfile}, capacity=${describeCapacity(id.envParts.capacity)}${id.envParts.selectedSet ? `, selected=${id.envParts.selectedSet.join(",")}` : ""}, protocol=${id.envParts.verification.protocol}, lifecycle=${id.envParts.verification.lifecycle} (${id.envParts.verification.source})]`
         : "";
     const gate = id.gate ?? "gate";
     const prefix = `reused ${id.scope === "tip" ? "tip " : ""}verdict (identity: gate=${gate} tree=${id.tree}${id.worktree ? ` worktree=${id.worktree}` : ""} command=${id.command} baseline=${id.baseline} env=${id.environment}${envDesc})`;

package/dist/gates/llm.d.ts

@@ -69,6 +69,12 @@ export interface LlmRunResult {
 }
 export declare const PROMPT_GLYPHS: readonly ["➜", "❯", "$", "%", ">>", ">"];
 export declare function reviewSeatOutput(raw: string, nonce: string, adapterBannerRows?: readonly string[]): string;
+/** OBS-1168(b): a judge/review seat whose pane could not be created or launched. Typed so the gates can
+ * contain it as seat recovery (another seat) or an infra park — never as a worker failure. */
+export declare class SeatLaunchError extends Error {
+    readonly seat: string;
+    constructor(seat: string, cause: unknown);
+}
 export declare const REVIEW_FIRST_LIVENESS_MS = 30000;
 export declare const REVIEW_SILENT_BYTE_FLOOR = 64;
 export declare function runHeadless(adapter: WorkerAdapter, model: string, prompt: string, cwd: string, timeoutMs?: number, effort?: Effort): Promise<string>;

package/dist/gates/llm.js

@@ -81,19 +81,22 @@ export function gatePaneName(role, taskId, suffix = "") {
 // T2 ownership contract: a canonical owned fallback (the daemon's nameFor now emits one) passes
 // through untouched; run-gates' "-r1" judge-retry suffix becomes attempt+1 so the retry pane's name
 // stays contract-parseable (tickmarkr:judge:<task>:1:<runId>) instead of a corrupted-runId shape.
+// C-3 (D-669): the hop suffix is -r<N> — the judge-flake retry is -r1 and the adjudicator -r2, so the two
+// never share an owned name (a retained retry pane must not shadow the adjudicator's slot).
 export function rolePaneNameFromPrompt(prompt, fallback) {
-    const retry = fallback.endsWith("-r1");
-    const base = retry ? fallback.slice(0, -3) : fallback;
+    const hop = /-r([1-9])$/.exec(fallback);
+    const suffix = hop ? hop[0] : "";
+    const base = hop ? fallback.slice(0, -suffix.length) : fallback;
     const owned = parseOwnedName(base);
     if (owned)
-        return retry ? formatOwnedName({ ...owned, attempt: owned.attempt + 1 }) : base;
+        return hop ? formatOwnedName({ ...owned, attempt: owned.attempt + Number(hop[1]) }) : base;
     const id = prompt.match(/## Task ([^\n:]+):/)?.[1];
     if (!id)
         return fallback;
     if (prompt.startsWith("TICKMARKR-JUDGE"))
-        return gatePaneName("judge", id, retry ? "-r1" : "");
+        return gatePaneName("judge", id, suffix);
     if (prompt.startsWith("TICKMARKR-REVIEW"))
-        return gatePaneName("review", id, retry ? "-r1" : "");
+        return gatePaneName("review", id, suffix);
     return fallback;
 }
 const llmOutputCapture = new AsyncLocalStorage();
@@ -299,6 +302,16 @@ export function reviewSeatOutput(raw, nonce, adapterBannerRows = []) {
     // often the seat's first byte than the harness's exit marker, and the next read completes either.
     return trailer ? seat.slice(0, trailer.index) : seat;
 }
+/** OBS-1168(b): a judge/review seat whose pane could not be created or launched. Typed so the gates can
+ * contain it as seat recovery (another seat) or an infra park — never as a worker failure. */
+export class SeatLaunchError extends Error {
+    seat;
+    constructor(seat, cause) {
+        super(`seat ${seat} failed to launch: ${cause instanceof Error ? cause.message : String(cause)}`);
+        this.seat = seat;
+        this.name = "SeatLaunchError";
+    }
+}
 export const REVIEW_FIRST_LIVENESS_MS = 30_000;
 // OBS-1039: a seat that wrote ten bytes and went quiet escaped the zero-byte beat and sat to the
 // ceiling. Below this many seat-authored bytes at the first beat the seat is `silent` — demoted and
@@ -341,9 +354,15 @@ async function runViaDriverDetailed(adapter, model, prompt, cwd, via, timeoutMs
             adapter.headlessCommand(pf, model, effort),
             gateExitTrailer(nonce),
         ].join("\n"));
-        slot = await via.driver.slot(cwd, rolePaneNameFromPrompt(prompt, via.name), via.label ? { label: via.label } : undefined);
-        via.onSlot?.(slot);
-        await via.driver.run(slot, paneDispatchCommand(scriptPath));
+        try {
+            slot = await via.driver.slot(cwd, rolePaneNameFromPrompt(prompt, via.name), via.label ? { label: via.label } : undefined);
+            via.onSlot?.(slot);
+            await via.driver.run(slot, paneDispatchCommand(scriptPath));
+        }
+        catch (error) {
+            forceClose = true; // a half-launched pane is closed, never kept
+            throw new SeatLaunchError(`${adapter.id}:${model}`, error);
+        }
         if (via.driver.sendKey) {
             try {
                 if (matchesTrustDialog(await via.driver.read(slot, 400), adapter.trustDialog)) {

package/dist/gates/review.d.ts

@@ -51,6 +51,11 @@ export declare function fetchTaskDiff(worktree: string, baseRef: string, files?:
 export declare function checkDiffCap(gate: string, measured: number, cap: number, prefix?: string): GateResult | null;
 /** Apply the strict reviewable-logic cap and the finite, larger capture cap independently. */
 export declare function checkTaskDiffCaps(gate: string, measured: Pick<TaskDiffMeasurement, "logicBytes" | "captureBytes">, logicCap: number, prefix?: string): GateResult | null;
+export declare function isGarbageReview(result: GateResult): result is GateResult & {
+    meta: {
+        reviewer: string;
+    };
+};
 export declare function isDiffCapPark(result: GateResult): boolean;
 export declare function diffCapParkReason(results: GateResult[]): string | null;
 export declare function modelId(model: string): string;
@@ -111,7 +116,7 @@ prefer?: string[], // v1.53 T2: review.prefer — reorders eligible channels, ne
 floor?: Tier, // task/config/prior floor from the caller; the author's own tier is ALWAYS applied here (RF-1)
 history?: string[], // run-scoped picks, oldest to newest; empty preserves the established ranking
 onSeat?: (seat: number, count: number) => void, demoted?: ReadonlySet<string>, excludeVendors?: ReadonlySet<string>, authors?: readonly string[]): BillingChannel | null;
-export type ReviewUnparseableCause = VerdictUnparseableCause | "launch-never-started" | "truncated" | "silent" | "closure-mismatch";
+export type ReviewUnparseableCause = VerdictUnparseableCause | "launch-never-started" | "truncated" | "silent" | "closure-mismatch" | "seat-launch-failed";
 /**
  * This shows the reviewer what the task DECLARED, never what the diff may actually reach. The diff
  * remains a separate stated input, so whether its touched paths fit the declaration stays a reviewer

package/dist/gates/review.js

@@ -6,12 +6,12 @@ import { filesGlob } from "../graph/files-glob.js";
 import { renderAcceptanceItem, TIERS } from "../graph/schema.js";
 import { getAdapter } from "../adapters/registry.js";
 import { shOk } from "../run/git.js";
-import { carryReviewFindings, observedReviewFingerprints, reviewFingerprintMatches, structuredFindings, UNIDENTIFIED } from "../run/journal.js";
+import { carryReviewFindings, isDeferredFinding, observedReviewFingerprints, reviewFingerprintMatches, structuredFindings, UNIDENTIFIED } from "../run/journal.js";
 import { redactSecrets } from "../run/redact.js";
 import { rankPreferredChannels, reviewPreferenceTieBreak } from "../route/role-pick.js";
 import { modelProvider } from "../route/preference.js";
 import { resolveStateDir } from "./cache.js";
-import { appendAnchoredReview, COMPLETION_FAKING_CHECKLIST, dewrapPaneVerdict, extractVerdictJson, generateVerdictNonce, parseAnchoredComments, runLlmDetailed, verdictNonceLine } from "./llm.js";
+import { appendAnchoredReview, COMPLETION_FAKING_CHECKLIST, dewrapPaneVerdict, extractVerdictJson, generateVerdictNonce, parseAnchoredComments, runLlmDetailed, SeatLaunchError, verdictNonceLine } from "./llm.js";
 import { classifyVerdictCause } from "./verdict-cause.js";
 import { captureDiffCapFor, measureArtifactDiff, reviewableLogicDiff, } from "./artifact-manifest.js";
 export { isProtectedEvidence, PROTECTED_EVIDENCE_PREFIXES, REGENERABLE_CAPTURE_PATHS, setAsideReceiptPath, setAsideRegenerableCaptures, } from "./artifact-manifest.js";
@@ -164,6 +164,12 @@ export function checkTaskDiffCaps(gate, measured, logicCap, prefix = "") {
         },
     };
 }
+// C-12: a reviewer is excluded for garbage only on the typed malformed-verdict flag. A material finding whose
+// prose mentions the word "unparseable" is a delivered verdict; matching the prose dropped T7's only reviewer.
+export function isGarbageReview(result) {
+    return result.gate === "review" && !result.pass && result.meta?.unparseable === true
+        && typeof result.meta?.reviewer === "string";
+}
 export function isDiffCapPark(result) {
     return result.pass === false
         && result.meta?.parkKind === "diff-cap"
@@ -400,6 +406,23 @@ function withoutExampleEcho(raw, nonce) {
     const chars = [...reviewResponseExample(nonce).replace(/\s+/g, "")];
     return raw.replace(new RegExp(chars.map((c) => c.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")).join("[\\s│|]*"), "g"), "");
 }
+/**
+ * OBS-1195: the parent an anchored comment names in its own optional `finding` field — a 1-based
+ * index into this verdict's findings, or a carried prior's id. A deferred entry or a resolved prior
+ * settles the anchor; any other valid parent binds it open; anything else leaves it unbound.
+ */
+function anchorParent(binding, findings, priors, resolved) {
+    if (typeof binding === "number") {
+        const entry = Array.isArray(findings) && Number.isInteger(binding) && binding >= 1 ? findings[binding - 1] : undefined;
+        if (!entry || typeof entry !== "object" || typeof entry.note !== "string")
+            return {};
+        return entry.defer === true ? { settled: "deferred" } : { entry: entry.note };
+    }
+    const prior = matchClosureId(binding, priors);
+    if (prior === undefined)
+        return {};
+    return resolved.includes(prior) ? { settled: "resolved", parent: prior } : { parent: prior };
+}
 export async function reviewGate(task, worktree, baseRef, author, channels, adapters, cfg, via, excludeReviewers, 
 // OBS-196: run dir for raw-output persistence on an unparseable verdict; absent (older callers,
 // direct tests) skips persistence and changes nothing else.
@@ -550,13 +573,21 @@ block the merge) or "minor" (style, naming, or preference that should not block)
 block approval. For a minor concern you have decided not to block on, set "defer": true and give a
 one-line "rationale" — it is recorded in the review, never dropped.
 A fix you prescribe that would break suites outside the task's declared write scope (files[]) is a scope finding, never a material one.
+Each material finding names the class of defect it belongs to and binds that class to the goal clause or
+acceptance criterion it violates (or to the regression this diff introduces), states its input → consequence,
+and marks its evidence executed (you ran the reproducer), static (you traced it by reading) or blocked (it
+could not run here). Blocked evidence never turns a finding into a pass, and a worker's own case table or
+enumeration never resolves a finding: judge the diff itself.
 
 Respond with ONLY this JSON:
-{"nonce": "${nonce}", "approve": true|false, "resolved": [], "reraised": [], "findings": [{"note": "...", "severity": "material"|"minor", "defer": false, "rationale": ""}], "comments": [{"path": "path/to/file", "line": 42, "body": "actionable feedback"}]}
+{"nonce": "${nonce}", "approve": true|false, "resolved": [], "reraised": [], "findings": [{"note": "...", "severity": "material"|"minor", "defer": false, "rationale": ""}], "comments": [{"path": "path/to/file", "line": 42, "body": "actionable feedback", "finding": 1}]}
 For every prior material, put its fingerprint in exactly one of resolved (verified fixed) or reraised
 (still a blocking defect). Use only the listed fingerprints; never omit one or put it in both lists.
 Approve iff no material finding remains and every prior material is resolved.
 The top-level comments array is optional. Use it only for actionable line-anchored feedback.
+A comment's optional "finding" names its parent: the 1-based index of its entry in findings, or a prior
+fingerprint copied from above. A comment anchored to a deferred entry or a resolved prior does not block;
+a comment naming no parent stays open.
 
 ${responseRequirement}
 `;
@@ -590,18 +621,33 @@ ${responseRequirement}
             savedBrief = undefined;
         }
     }
-    const llm = await runLlmDetailed(getAdapter(reviewer.adapter, adapters), reviewer.model, prompt, worktree, via ? {
-        driver: via.driver,
-        keep: via.keep,
-        onSlot: via.onSlot,
-        name: via.nameFor("review", reviewer.adapter),
-        label: via.labelFor("review"),
-    } : undefined, 
-    // frontier reviewers routinely need >5min on a configured-cap-sized diff, and `claude -p` buffers all
-    // output until completion — runLlm's 300s default killed reviews mid-flight, returning empty
-    // stdout that read as "unparseable" and escalated to re-implementation of green code
-    // (run-20260709-104447 P87-09). The configured ceiling defaults to that measured 15 minutes.
-    cfg.review.timeoutMs, reviewer.effort);
+    const provider = modelProvider(reviewer.model, reviewer.vendor);
+    let llm;
+    try {
+        llm = await runLlmDetailed(getAdapter(reviewer.adapter, adapters), reviewer.model, prompt, worktree, via ? {
+            driver: via.driver,
+            keep: via.keep,
+            onSlot: via.onSlot,
+            name: via.nameFor("review", reviewer.adapter),
+            label: via.labelFor("review"),
+        } : undefined, 
+        // frontier reviewers routinely need >5min on a configured-cap-sized diff, and `claude -p` buffers all
+        // output until completion — runLlm's 300s default killed reviews mid-flight, returning empty
+        // stdout that read as "unparseable" and escalated to re-implementation of green code
+        // (run-20260709-104447 P87-09). The configured ceiling defaults to that measured 15 minutes.
+        cfg.review.timeoutMs, reviewer.effort);
+    }
+    catch (error) {
+        if (!(error instanceof SeatLaunchError))
+            throw error;
+        // OBS-1168(b): the seat never launched, so there is no verdict and nothing about the WORK. A typed
+        // no-verdict re-routes to another seat in run-gates; an exhausted pool is an infra park.
+        return { gate: "review", pass: false,
+            details: `review dispatch failed — ${error.message} (reviewer ${reviewer.adapter}:${reviewer.model}; vendor: ${reviewer.vendor}; provider: ${provider}; cause: seat-launch-failed) — failing closed`,
+            meta: { ...policyMeta, ...rotationMeta, ...floorMeta, reviewer: channelKey(reviewer), reviewerTier: reviewer.tier,
+                vendor: reviewer.vendor, provider, noVerdict: true, classification: "infra", infra: true,
+                cause: "seat-launch-failed", ...(savedBrief ? { briefPath: savedBrief } : {}) } };
+    }
     const raw = llm.output;
     let saved;
     if (artifactDir) {
@@ -613,7 +659,6 @@ ${responseRequirement}
             saved = undefined; // persistence is evidence, not a gate input — never fail the gate on it
         }
     }
-    const provider = modelProvider(reviewer.model, reviewer.vendor);
     // A pane's own dewrap stops at the first parseable nonce-bound object; once the example's echo is gone
     // a genuinely wrapped verdict behind it is reconstructed here, exactly as llm.ts would have.
     const echoFree = withoutExampleEcho(raw, nonce);
@@ -653,6 +698,8 @@ ${responseRequirement}
                 ...(cause === "malformed-verdict" ? { unparseable: true } : { noVerdict: true, classification: "infra", infra: true }),
                 cause,
                 ...(closureMismatch ? { resolved: v?.resolved, reraised: v?.reraised, carriedFingerprints: priorIds.flatMap(observedReviewFingerprints) } : {}),
+                // OBS-1196: a whole verdict that breaks the closure protocol was DELIVERED; only undelivered bytes are re-asked.
+                ...(closureInvalid ? { closureInvalid: true } : {}),
                 bytes, seatAuthoredBytes: bytes,
                 ...(saved ? { rawPath: saved } : {}),
                 ...(savedBrief ? { briefPath: savedBrief } : {}),
@@ -669,25 +716,24 @@ ${responseRequirement}
         if (decided.pass)
             decided.headline = "requested changes";
         decided.pass = false;
-        // A reviewer may also restate a re-raised material in findings. Preserve the original
-        // prose once so an unchanged defect keeps the same failure brief across repair rounds.
-        for (const finding of reraised) {
-            const line = `- [material] ${finding.note}`;
-            if (!decided.lines.includes(line))
-                decided.lines.push(line);
-        }
     }
-    const prose = `reviewer ${reviewer.adapter}:${reviewer.model} (vendor: ${reviewer.vendor}; provider: ${provider}): ${decided.headline}${decided.lines.length ? "\n" + decided.lines.join("\n") : ""}`;
-    const details = appendAnchoredReview(prose, v);
+    // OBS-1195: an anchor leaves the blocking set only through an explicit, unambiguous parent
+    // disposition: its own `finding` names a deferred entry of this verdict or a prior it resolved.
+    // Path coincidence binds nothing; an unbound or ambiguous anchor stays open exactly as before.
+    const comments = parseAnchoredComments(v);
+    const rawComments = (comments.length ? v.comments : []);
+    const resolvedIds = (v.resolved ?? []).map((id) => matchClosureId(id, priorIds));
+    const anchors = comments.map((comment, i) => ({ ...comment, ...anchorParent(rawComments[i]?.finding, v.findings, priorIds, resolvedIds) }));
+    const openAnchors = { comments: anchors.filter((a) => !a.settled) };
+    const settledAnchors = anchors.flatMap((a, i) => a.settled
+        ? [{ path: a.path, line: a.line, body: a.body, disposition: a.settled, finding: rawComments[i].finding }] : []);
     // Only the verdict's anchors may supply missing evidence, and only when unambiguous.
     // Reuse the journal's path normalization without changing legacy details-only parsing.
-    const anchoredPaths = new Set(parseAnchoredComments(v).map((comment) => {
-        const anchor = structuredFindings("review", `- ${comment.path}:${comment.line} — anchor`)
-            .find((finding) => finding.class === "review:anchored");
-        return anchor?.path ?? comment.path;
-    }));
+    const anchorRow = (comment) => structuredFindings("review", `- ${comment.path}:${comment.line} — ${comment.body}`)
+        .find((finding) => finding.class === "review:anchored");
+    const anchoredPaths = new Set(comments.map((comment) => anchorRow({ ...comment, body: "anchor" })?.path ?? comment.path));
     const anchoredPath = anchoredPaths.size === 1 ? [...anchoredPaths][0] : undefined;
-    const ownDetails = appendAnchoredReview(ownLines.join("\n"), v);
+    const ownDetails = appendAnchoredReview(ownLines.join("\n"), openAnchors);
     const currentRows = (ownDetails.trim() ? structuredFindings("review", ownDetails) : [])
         .map((finding) => finding.path === UNIDENTIFIED && anchoredPath
         ? { ...finding, path: anchoredPath, fingerprint: `${finding.class}|${anchoredPath}|${finding.symbol}` }
@@ -708,7 +754,50 @@ ${responseRequirement}
     const unambiguousRows = linkedRows.map((finding) => finding.reraisedFrom
         && linkedRows.filter((row) => row.reraisedFrom === finding.reraisedFrom).length > 1
         ? { ...finding, reraisedFrom: undefined } : finding);
-    const carriedRows = carryReviewFindings(reraised, unambiguousRows);
+    // OBS-1195: a deferred entry echoing a prior's id names that prior as the concern it defers. The
+    // link rides the row for the journal's anchor fold only; a deferral never re-seats a material chain.
+    // A prior any material row restates — even ambiguously, its link cleared above — is claimed, not
+    // deferred: the deferral gets no link, so its bound anchors stay open (fail closed).
+    const claimed = new Set(linkedRows.map((row) => row.reraisedFrom).filter((id) => id !== undefined));
+    const rows = unambiguousRows.map((finding) => {
+        if (!isDeferredFinding(finding))
+            return finding;
+        const entry = v.findings?.find((entry) => entry && typeof entry === "object" && entry.note === finding.note);
+        const id = matchClosureId(entry?.reraised, reraised);
+        return id && !claimed.has(id) ? { ...finding, reraisedFrom: id } : finding;
+    });
+    // OBS-1195: current prose is rendered once. A prior's original wording is echoed only when no
+    // current finding positively binds it; a bound chain keeps that spelling as lineage instead.
+    for (const finding of reraised) {
+        const line = `- [material] ${finding.note}`;
+        const bound = unambiguousRows.some((row) => row.reraisedFrom === finding.fingerprint);
+        if (!bound && !decided.lines.includes(line))
+            decided.lines.push(line);
+    }
+    const prose = `reviewer ${reviewer.adapter}:${reviewer.model} (vendor: ${reviewer.vendor}; provider: ${provider}): ${decided.headline}${decided.lines.length ? "\n" + decided.lines.join("\n") : ""}`;
+    const details = appendAnchoredReview(prose, openAnchors);
+    // OBS-1195: a prior's `reraisedFrom` is the link an EARLIER verdict drew. Carried forward unrestated,
+    // it would read as this verdict's own material claim on that parent and block the anchor fold from
+    // honouring an explicit deferral of it. Lineage lives in observedFingerprints; drop the stale link.
+    const carried = carryReviewFindings(reraised.map(({ reraisedFrom: _stale, ...prior }) => prior), rows);
+    // An open anchor bound to a current entry names that entry's carried chain, only when unique.
+    const parentOf = (anchor) => {
+        if (anchor.entry === undefined)
+            return anchor.parent;
+        const owners = carried.filter((f) => f.class !== "review:anchored" && f.note === anchor.entry);
+        return owners.length === 1 ? owners[0].fingerprint : undefined;
+    };
+    const carriedRows = carried.map((finding) => {
+        if (finding.class !== "review:anchored")
+            return finding;
+        // Every comment spelling this row must name the same parent, else the anchor stays unbound.
+        const parents = new Set(anchors.filter((a) => {
+            const row = a.settled ? undefined : anchorRow(a);
+            return row?.path === finding.path && row.note === finding.note;
+        }).map(parentOf));
+        const [parent] = parents;
+        return parents.size === 1 && parent ? { ...finding, boundTo: parent } : finding;
+    });
     return {
         gate: "review",
         pass: decided.pass,
@@ -723,6 +812,7 @@ ${responseRequirement}
                 reraisedMatches: (v.reraised ?? []).map((id) => matchClosureId(id, priorIds)).filter((id) => id !== undefined),
             } : {}),
             ...(!decided.pass ? { findings: carriedRows } : {}),
+            ...(settledAnchors.length ? { settledAnchors } : {}),
             ...(saved ? { rawPath: saved } : {}),
             ...(savedBrief ? { briefPath: savedBrief } : {}),
         },