fix(takes): kill propose_takes rescan loop, keep every claim, and route cycle/brainstorm models through config (#2804 #2805 )

Takeover of #2804 and #2805 (stacked), rebased onto current master. Rescan loop + dropped claims (#2804): - Zero-claim scans now write a status='empty' sentinel row so the (source_id, page_slug, content_hash, prompt_version) cache hits on every subsequent cycle instead of re-spending the extractor LLM call on the same unchanged page forever. - The idempotency index gains md5(claim_text): multi-claim pages keep every claim (the 4-column unique index silently dropped claims 2..N via ON CONFLICT DO NOTHING). proposals_inserted now counts rows that actually landed (RETURNING id), not insert attempts. - Migration renumbered 123 -> 125 (master shipped v123 configurable_fts_language and v124 after the PR was cut; the duplicate v123 would never have run under the version > current runner). Schema parity across schema.sql, pglite-schema.ts, schema-embedded.ts. Model-tier honoring (#2805), reworked for current master: - propose_takes, grade_takes, calibration_profile resolve their model through resolveModel(models.<phase> > models.default > env > getChatModel()) — per-phase config keys are now honored while the gateway chat model (already tier-resolved by reconfigureGatewayWithEngine) stays the default, preserving master's #2451 semantics (provider-prefixed stored ids, nested prefixes like openrouter:... intact) and the #2997 opts.model ?? getChatModel() idiom. The PR's original bare-tail-everywhere approach was dropped where it conflicted with #2451; grade_takes keeps the bare tail ONLY for its cache key / stored judge_model_id (continuity with historical rows), while the judge call itself gets the full resolved string — fixing the live label-vs-actual mismatch (call rode chat_model, telemetry recorded a hardcoded claude-sonnet-4-6). - brainstorm resolves via models.brainstorm > tier chain (was hardcoded anthropic:claude-sonnet-4-6). - gbrain models lists the four new per-task keys; output caps and context-window entries added for newer Claude models. Tests: test/propose-takes-rescan.test.ts (new, hermetic PGLite); config-key coverage in propose-takes/grade-takes tests; existing #2451 and gateway-chat-model tests pass unchanged. Co-authored-by: p3ob7o <p3ob7o@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-08-16 18:02:30 +00:00 · 2026-07-21 14:33:19 -07:00
34 changed files with 436 additions and 761 deletions
@@ -527,9 +527,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
  process.on('SIGINT',  () => { void shutdown('SIGINT'); });

  let consecutiveErrors = 0;
-  // Parser-probe fixture warning is once-per-process, not once-per-cycle
-  // (compiled-binary installs have no source tree; don't spam the log).
-  let parserProbeFixtureWarned = false;
  // v0.37.7.0 #1162 — counter for consecutive reconnect failures.
  // Reset on every successful health probe or reconnect. Threshold
  // controlled by GBRAIN_AUTOPILOT_MAX_RECONNECT_FAILS env (default 30).
@@ -1076,36 +1073,17 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
    // loop. Probe runs even when cycleOk=false (probe may surface signal
    // explaining why the cycle is failing).
    try {
-      const { resolveProbeEnabled, resolveProbeMaxUsd, runNightlyQualityProbe } = await import('../core/cycle/nightly-quality-probe.ts');
-      // Dual-plane read: `gbrain config set` (what the doctor enable hint
-      // prints) writes the DB plane; ~/.gbrain/config.json is the fallback.
-      let dbEnabled: string | null = null;
-      let dbMaxUsd: string | null = null;
-      try {
-        dbEnabled = await engine.getConfig('autopilot.nightly_quality_probe.enabled');
-        dbMaxUsd = await engine.getConfig('autopilot.nightly_quality_probe.max_usd');
-      } catch { /* DB unavailable → file plane only */ }
-      const probeEnabled = resolveProbeEnabled(dbEnabled, cfg?.autopilot?.nightly_quality_probe?.enabled);
+      const probeEnabled = cfg?.autopilot?.nightly_quality_probe?.enabled === true;
      if (probeEnabled) {
+        const { runNightlyQualityProbe } = await import('../core/cycle/nightly-quality-probe.ts');
        const { runLongMemEvalForProbe, runCrossModalBatchForProbe } = await import('../core/cycle/nightly-probe-adapters.ts');
        const { isAvailable } = await import('../core/ai/gateway.ts');
-        const { existsSync } = await import('node:fs');
-        const { fileURLToPath } = await import('node:url');
-        const { join } = await import('node:path');
-        const maxUsd = resolveProbeMaxUsd(dbMaxUsd, cfg?.autopilot?.nightly_quality_probe?.max_usd);
-        // The committed fixture (test/fixtures/longmemeval-nightly.jsonl)
-        // lives in the gbrain PACKAGE, not the brain repo — repoPath is
-        // sync.repo_path (the user's brain), where the fixture never
-        // exists, so the probe error'd on every real install. Resolve the
-        // package root from the module location; keep repoPath as the
-        // fallback for setups that vendor the fixture into the brain repo.
-        const pkgRoot = fileURLToPath(new URL('../..', import.meta.url));
-        const fixtureAtPkgRoot = existsSync(join(pkgRoot, 'test', 'fixtures', 'longmemeval-nightly.jsonl'));
+        const maxUsd = Number(cfg?.autopilot?.nightly_quality_probe?.max_usd ?? 5);
        await runNightlyQualityProbe({
          isEnabled: () => true, // already gated above; phase re-checks for defense-in-depth
          hasEmbeddingProvider: () => isAvailable('embedding'),
          resolveMaxUsd: () => maxUsd,
-          resolveRepoRoot: () => (fixtureAtPkgRoot ? pkgRoot : repoPath ?? gbrainHomePath('.')),
+          resolveRepoRoot: () => repoPath ?? gbrainHomePath('.'),
          runLongMemEval: runLongMemEvalForProbe,
          runCrossModalBatch: runCrossModalBatchForProbe,
          now: () => new Date(),
@@ -1117,62 +1095,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
      // informational; autopilot loop continues.
    }

-    // 4.6 — Nightly conversation-parser probe (v0.41.16.0 phase module;
-    // the scheduler wire-up was deferred at ship and is added here). Same
-    // posture as 4.5: the phase owns its gates (enabled/mode-gate, LLM
-    // key), the wiring owns invocation + the audit row, and a probe
-    // failure NEVER crashes the autopilot loop. Per D10 the probe is
-    // default-ON for search.mode=tokenmax, opt-in otherwise.
-    try {
-      const { runConversationParserNightlyProbe } = await import('../core/conversation-parser/nightly-probe.ts');
-      const { logParserProbeEvent, parserProbeRanWithin } = await import('../core/audit-parser-probe.ts');
-      const { isAvailable } = await import('../core/ai/gateway.ts');
-      const { existsSync } = await import('node:fs');
-      const { fileURLToPath } = await import('node:url');
-      const { join } = await import('node:path');
-      // Flag reads dual-plane: the DB row (`gbrain config set …`) wins,
-      // ~/.gbrain/config.json is the fallback. search.mode lives on the
-      // DB plane only (mode.ts owns it).
-      let parserDbEnabled: string | null = null;
-      let dbSearchMode: string | null = null;
-      try {
-        parserDbEnabled = await engine.getConfig('autopilot.conversation_parser_probe.enabled');
-        dbSearchMode = await engine.getConfig('search.mode');
-      } catch { /* DB unavailable → file plane only */ }
-      const parserEnabled = parserDbEnabled != null
-        ? parserDbEnabled === 'true'
-        : cfg?.autopilot?.conversation_parser_probe?.enabled === true;
-      const searchMode = dbSearchMode ?? '';
-      // Fixtures are committed in the gbrain package (test/fixtures/…),
-      // NOT the brain repo — resolve from the module location. Compiled
-      // binaries carry no source tree: skip quietly instead of writing
-      // failure rows that would flip doctor to WARN on every binary install.
-      const pkgRoot = fileURLToPath(new URL('../..', import.meta.url));
-      const fixturePath = join(pkgRoot, 'test', 'fixtures', 'conversation-formats', 'all.jsonl');
-      const adversarialPath = join(pkgRoot, 'test', 'fixtures', 'conversation-formats', 'adversarial.jsonl');
-      const shouldInvoke = parserEnabled || searchMode === 'tokenmax';
-      if (shouldInvoke && existsSync(fixturePath) && existsSync(adversarialPath)) {
-        const result = await runConversationParserNightlyProbe({
-          isEnabled: () => parserEnabled,
-          searchMode: () => searchMode,
-          hasLlmKey: () => isAvailable('chat'),
-          resolveFixturePath: () => fixturePath,
-          resolveAdversarialPath: () => adversarialPath,
-          now: () => new Date(),
-          shouldSkipForRateLimit: () => parserProbeRanWithin(24 * 60 * 60 * 1000),
-        });
-        // rate_limited is a non-run: the loop ticks every few minutes, so
-        // logging every skip would flood the audit file with no-signal rows.
-        if (result.outcome !== 'rate_limited') logParserProbeEvent(result);
-      } else if (shouldInvoke && !parserProbeFixtureWarned) {
-        parserProbeFixtureWarned = true;
-        console.error(`[parser-probe] fixtures not found under ${pkgRoot}; skipping (probe needs a source-checkout install)`);
-      }
-    } catch (e) {
-      logError('autopilot.parser_probe', e);
-      // Informational, like 4.5: do NOT bump consecutiveErrors.
-    }
-
    // Wait for next cycle
    await new Promise(r => setTimeout(r, interval * 1000));
  }
@@ -2960,54 +2960,6 @@ function _resolveSyncFreshnessHours(varName: string, fallback: number): number {
 * branch (disabled / enabled-no-events / enabled-all-pass / enabled-with-failures)
 * without spinning up the audit JSONL or a real config file.
 */
-/**
- * Pure function form of the conversation_parser_probe_health check.
- * Mirrors computeNightlyQualityProbeHealthCheck: skip-with-hint when the
- * probe is off and silent, surface the last 7 days of audit events when
- * it has run, WARN on any non-pass outcome.
- *
- * `effectiveEnabled` folds the D10 mode-gate in: explicitly enabled OR
- * search.mode=tokenmax (where the probe is default-on).
- */
-export function computeConversationParserProbeHealthCheck(
-  effectiveEnabled: boolean,
-  events: ReadonlyArray<{ outcome: string; ts: string; reason?: string }>,
-): Check {
-  const name = 'conversation_parser_probe_health';
-  if (!effectiveEnabled && events.length === 0) {
-    return {
-      name,
-      status: 'ok',
-      message:
-        'disabled (opt-in; default-on only for search.mode=tokenmax). Enable with: ' +
-        '`gbrain config set autopilot.conversation_parser_probe.enabled true`',
-    };
-  }
-  if (events.length === 0) {
-    return {
-      name,
-      status: 'ok',
-      message: 'enabled but no probe events in the last 7 days (next run by autopilot; fixtures require a source-checkout install).',
-    };
-  }
-  const bad = events.filter(e => e.outcome !== 'pass');
-  const latest = events[events.length - 1]!;
-  if (bad.length > 0) {
-    return {
-      name,
-      status: 'warn',
-      message:
-        `${bad.length}/${events.length} probe run(s) in the last 7 days did not pass; ` +
-        `latest: ${latest.outcome}${latest.reason ? ` (${latest.reason})` : ''}`,
-    };
-  }
-  return {
-    name,
-    status: 'ok',
-    message: `${events.length} probe run(s) in the last 7 days, all pass (latest ${latest.ts}).`,
-  };
-}
-
 export function computeNightlyQualityProbeHealthCheck(
  probeEnabled: boolean,
  events: ReadonlyArray<{ outcome: string; ts: string; detail?: string }>,
@@ -4891,17 +4843,10 @@ export async function buildChecks(
  try {
    const { readRecentQualityProbeEvents } = await import('../core/audit-quality-probe.ts');
    const { loadConfig } = await import('../core/config.ts');
-    const { resolveProbeEnabled } = await import('../core/cycle/nightly-quality-probe.ts');
    let probeEnabled = false;
    try {
-      // Dual-plane read, matching the autopilot gate: the DB row (what the
-      // enable hint's `gbrain config set` writes) wins; file plane fallback.
-      let dbVal: string | null = null;
-      try {
-        dbVal = engine ? await engine.getConfig('autopilot.nightly_quality_probe.enabled') : null;
-      } catch { /* DB unavailable → file plane only */ }
      const cfg = loadConfig();
-      probeEnabled = resolveProbeEnabled(dbVal, (cfg as any)?.autopilot?.nightly_quality_probe?.enabled);
+      probeEnabled = Boolean((cfg as any)?.autopilot?.nightly_quality_probe?.enabled);
    } catch { /* config unavailable → treat as disabled */ }
    const events = readRecentQualityProbeEvents(7);
    const check = computeNightlyQualityProbeHealthCheck(probeEnabled, events);
@@ -5085,29 +5030,19 @@ export async function buildChecks(

  // 3d.5 v0.41.13.0 — conversation_parser_probe_health. Mode-gated
  // per D10: ON when search.mode=tokenmax, opt-in for other modes.
-  // Surfaces the last 7 days of nightly-probe audit events; warn on any
-  // non-pass outcome (fail / budget_exceeded / adversarial_false_positive).
-  // (Until the autopilot wire-up this was a hardcoded "Skipped" stub.)
-  try {
-    const { readRecentParserProbeEvents } = await import('../core/audit-parser-probe.ts');
-    let parserProbeEnabled = false;
-    try {
-      let dbVal: string | null = null;
-      let dbMode: string | null = null;
-      try {
-        dbVal = engine ? await engine.getConfig('autopilot.conversation_parser_probe.enabled') : null;
-        dbMode = engine ? await engine.getConfig('search.mode') : null;
-      } catch { /* DB unavailable → file plane only */ }
-      const { loadConfig } = await import('../core/config.ts');
-      const fileVal = (loadConfig() as any)?.autopilot?.conversation_parser_probe?.enabled;
-      const flagOn = dbVal != null ? dbVal === 'true' : fileVal === true;
-      parserProbeEnabled = flagOn || dbMode === 'tokenmax';
-    } catch { /* config unavailable → treat as disabled */ }
-    const parserEvents = readRecentParserProbeEvents(7);
-    checks.push(computeConversationParserProbeHealthCheck(parserProbeEnabled, parserEvents));
-  } catch {
-    // Best-effort; audit-log read failure shouldn't stop doctor.
-  }
+  // Surface the last 7 days of nightly-probe events; warn on FAIL /
+  // BUDGET_EXCEEDED / adversarial_false_positive.
+  //
+  // v0.41.13.0 ships the probe as opt-in (autopilot wiring deferred
+  // to T7 in the cathedral plan); this check skips with an enable
+  // hint until the probe has at least one audit event written.
+  checks.push({
+    name: 'conversation_parser_probe_health',
+    status: 'ok',
+    message:
+      'Skipped (nightly probe is opt-in; enable with ' +
+      '`gbrain config set autopilot.conversation_parser_probe.enabled true`)',
+  });

  // 3e. home_dir_in_worktree (v0.35.8.0). Walks up from `gbrainPath()`
  // looking for a `.git` directory OR file. If found, warns: `~/.gbrain/`
@@ -76,7 +76,7 @@ FLAGS:
                           dimensions (goal, depth, sourcing, specificity, useful).
  --cycles N               1-3. Default: 3 in TTY, 1 in non-TTY (T11). Each
                           cycle is 3 model calls; verdict aggregates over them.
-  --slot-a-model <id>      Override default 'openai:gpt-5.2'.
+  --slot-a-model <id>      Override default 'openai:gpt-4o'.
  --slot-b-model <id>      Override default 'anthropic:claude-opus-4-7'.
  --slot-c-model <id>      Override default 'google:gemini-1.5-pro'.
  --receipt-dir <path>     Default: gbrainPath('eval-receipts').
@@ -468,14 +468,6 @@ interface BatchRow {
  question_id: string;
  question: string;
  hypothesis: string;
-  /**
-   * Gold answer from the benchmark dataset, when the upstream eval emits
-   * it (eval-longmemeval does). Folded into the judge task so CORRECTNESS
-   * is verifiable — without it a judge panel that sees only
-   * {question, hypothesis} cannot validate a terse factual answer against
-   * a haystack it never saw.
-   */
-  answer?: string;
 }

 /**
@@ -589,7 +581,6 @@ function readBatchRows(path: string): BatchReadResult {
      question_id: typeof obj.question_id === 'string' ? obj.question_id : `line-${lineNo}`,
      question: obj.question,
      hypothesis: obj.hypothesis,
-      ...(typeof obj.answer === 'string' && obj.answer.length > 0 ? { answer: obj.answer } : {}),
    });
  }
  if (summarySkipped > 0) {
@@ -706,11 +697,7 @@ async function runBatchMode(parsed: ParsedArgs, opts: RunCrossModalOpts): Promis
      fn: async (row, idx) => {
        process.stderr.write(`[eval cross-modal batch] ${idx + 1}/${rows.length} ${row.question_id} starting...\n`);
        return await runEvalFn({
-          // With a gold answer the judges can actually verify correctness;
-          // without one they see only {question, hypothesis} and cannot.
-          task: row.answer
-            ? `${row.question}\n\nExpected answer (gold label from the benchmark dataset): ${row.answer}`
-            : row.question,
+          task: row.question,
          output: row.hypothesis,
          slug: row.question_id,
          dimensions,
@@ -33,7 +33,6 @@ import {
  type AliasMap,
 } from '../eval/longmemeval/extract.ts';
 import { extractCandidateEntities } from '../core/think/entity-extract.ts';
-import { splitProviderModelId } from '../core/model-id.ts';
 import { resolveEntitySlugWithSource, type ResolutionSource } from '../core/entities/resolve.ts';
 import { formatTrajectoryBlock } from '../core/trajectory-format.ts';

@@ -470,22 +469,14 @@ export async function runEvalLongMemEval(args: string[], runOpts: RunOpts = {}):
  });

  // Wrap Anthropic SDK so its `.messages.create` shape matches ThinkLLMClient.
-  // Same pattern as src/core/think/index.ts:247-249 — EXCEPT think's default
-  // client routes through the gateway, which parses `provider:model` recipe
-  // ids. This eval's client is a raw SDK by design (hermetic, no gateway
-  // dependency), and resolveModel returns RECIPE ids (`anthropic:claude-…`);
-  // passing one through unstripped 404s every answer/extractor call, which
-  // surfaces downstream as all-upstream_error batches in the nightly probe.
-  const toSdkModel = (m: string): string => splitProviderModelId(m).model || m;
+  // Same pattern as src/core/think/index.ts:247-249.
  const realClient = new Anthropic();
  const client: ThinkLLMClient = runOpts.client ?? {
-    create: (params, callOpts) =>
-      realClient.messages.create({ ...params, model: toSdkModel(params.model) }, callOpts),
+    create: (params, callOpts) => realClient.messages.create(params, callOpts),
  };
  // v0.40.2.0 — separate extractor client (defaults to same SDK).
  const extractorClient: ThinkLLMClient = runOpts.extractorClient ?? {
-    create: (params, callOpts) =>
-      realClient.messages.create({ ...params, model: toSdkModel(params.model) }, callOpts),
+    create: (params, callOpts) => realClient.messages.create(params, callOpts),
  };
  const trajectoryEnabled = !opts.noTrajectory;
  const extractorModel = trajectoryEnabled
@@ -760,11 +751,6 @@ async function runOneQuestion(
    // v0.40.1.0 (Track D / T2) — copy question_type into the row so the
    // by_type_summary can be rebuilt from the file on resume runs.
    question_type: q.question_type,
-    // Gold answer for downstream consumers that verify correctness (the
-    // cross-modal --batch judge folds it into the task; evaluate_qa.py
-    // ignores unknown fields). Without it a judge can't validate a terse
-    // factual hypothesis against a haystack it never saw.
-    ...(q.answer !== undefined ? { answer: q.answer } : {}),
    hypothesis,
    retrieved_session_ids: retrievedSessionIds,
    ...(recallHit !== undefined ? { recall_hit: recallHit } : {}),
@@ -50,6 +50,10 @@ const PER_TASK_KEYS: Array<{ key: string; tier: ModelTier; description: string }
  { key: 'models.eval.contradictions_judge', tier: 'utility',  description: 'Contradiction probe judge (v0.34 temporal-aware)' },
  { key: 'models.expansion',                tier: 'utility',   description: 'Query expansion for hybrid search' },
  { key: 'models.chat',                     tier: 'reasoning', description: 'Default `gateway.chat()` model' },
+  { key: 'models.propose_takes',            tier: 'reasoning', description: 'propose_takes claim extractor' },
+  { key: 'models.grade_takes',              tier: 'reasoning', description: 'grade_takes verdict judge' },
+  { key: 'models.calibration_profile',      tier: 'reasoning', description: 'Calibration profile generator' },
+  { key: 'models.brainstorm',               tier: 'reasoning', description: '`gbrain brainstorm` orchestrator' },
 ];

 interface ModelEntry {
@@ -1,63 +0,0 @@
-/**
- * Nightly conversation-parser probe audit trail.
- *
- * One event per REAL probe run lands in
- * `~/.gbrain/audit/parser-probe-YYYY-Www.jsonl` (ISO-week rotation via the
- * shared audit-writer primitive; honors `GBRAIN_AUDIT_DIR`).
- * Scheduler-cadence skips (`rate_limited`) are NOT logged — the autopilot
- * loop ticks every few minutes, so logging every skip would flood the
- * audit file with rows that carry no signal.
- *
- * Read by `gbrain doctor`'s `conversation_parser_probe_health` check and
- * by the autopilot wiring's 24h rate-limit gate (`parserProbeRanWithin`).
- */
-
-import { createAuditWriter } from './audit/audit-writer.ts';
-import type { NightlyProbeResult } from './conversation-parser/nightly-probe.ts';
-
-export type ParserProbeAuditEvent = NightlyProbeResult;
-
-const writer = createAuditWriter<ParserProbeAuditEvent>({
-  featureName: 'parser-probe',
-  errorLabel: 'gbrain',
-  errorMessagePrefix: 'parser-probe audit ',
-  errorTrailer: '; probe continues',
-});
-
-/** Append one parser-probe event. Best-effort; never throws. */
-export function logParserProbeEvent(event: ParserProbeAuditEvent): void {
-  writer.log(event);
-}
-
-/**
- * Read recent parser-probe events (current + previous ISO week, filtered
- * to the window). Missing files and corrupt rows are skipped silently.
- */
-export function readRecentParserProbeEvents(
-  days = 7,
-  now: Date = new Date(),
-): ParserProbeAuditEvent[] {
-  return writer.readRecent(days, now);
-}
-
-/** Exposed for tests pinning the rotation edge cases. */
-export function computeParserProbeAuditFilename(now: Date = new Date()): string {
-  return writer.computeFilename(now);
-}
-
-/**
- * 24h rate-limit gate for the autopilot wiring: true when any audited run
- * happened within `windowMs` of `now`. Only REAL outcomes are audited (see
- * module header), so a pass/fail today blocks re-runs until tomorrow while
- * scheduler-cadence skips never extend the window.
- */
-export function parserProbeRanWithin(
-  windowMs: number,
-  now: Date = new Date(),
-): boolean {
-  const cutoff = now.getTime() - windowMs;
-  return readRecentParserProbeEvents(2, now).some((ev) => {
-    const ts = Date.parse(ev.ts);
-    return Number.isFinite(ts) && ts >= cutoff;
-  });
-}
@@ -61,7 +61,10 @@ export const MAX_OUTPUT_TOKENS_CEIL = 32_000;
 * (with a readable error) instead of the provider's opaque HTTP 400.
 */
 export const ANTHROPIC_OUTPUT_CAPS: Record<string, number> = {
+  'claude-fable-5': 64_000,
+  'claude-opus-4-8': 32_000,
  'claude-opus-4-7': 32_000,
+  'claude-sonnet-5': 64_000,
  'claude-sonnet-4-6': 64_000,
  'claude-haiku-4-5': 64_000,
  'claude-haiku-4-5-20251001': 64_000,
@@ -32,6 +32,7 @@
 */

 import type { BrainEngine } from '../engine.ts';
+import { resolveModel } from '../model-config.ts';
 import { chat as defaultChat, embedQuery, type ChatResult, type ChatOpts } from '../ai/gateway.ts';
 import { hybridSearch, hybridSearchCached } from '../search/hybrid.ts';
 import { fetchFar, type CloseRef, type FarPage } from './domain-bank.ts';
@@ -538,7 +539,14 @@ async function _runBrainstormInner(
  const embedFn = opts.embedQueryFn ?? embedQuery;

  // ---- Phase 0: cost preview + TTY grace ----
-  const modelStr = opts.modelOverride ?? 'anthropic:claude-sonnet-4-6';
+  // Tier-resolved (mirrors the cycle phases): honors models.brainstorm >
+  // models.default > models.tier.reasoning; the fallback keeps stock
+  // behavior identical (reasoning tier default IS claude-sonnet-4-6).
+  const modelStr = opts.modelOverride ?? await resolveModel(engine, {
+    configKey: 'models.brainstorm',
+    tier: 'reasoning',
+    fallback: 'anthropic:claude-sonnet-4-6',
+  });
  const { aborted, estimate } = await previewCostAndWait({
    profile,
    model: modelStr,
@@ -105,16 +105,6 @@ export interface GBrainConfig {
       */
      max_usd?: number;
    };
-    /**
-     * v0.41.16.0 — nightly conversation-parser probe. Per D10: default ON
-     * for `search.mode=tokenmax` brains, opt-in for conservative/balanced.
-     * ~$0.05/night with the committed fixtures × Haiku polish. Gated
-     * INSIDE the autopilot tick body, like nightly_quality_probe.
-     */
-    conversation_parser_probe?: {
-      /** Enable for non-tokenmax modes. Defaults to false. */
-      enabled?: boolean;
-    };
    /**
     * v0.42.x (#1685 GAP D) — extract_atoms backlog auto-drain. Default ON so a
     * pack-gated silent backlog never piles up unseen; daily-spend-capped so the
@@ -17,11 +17,11 @@
 * Cost: ~$0.05/night with default fixtures × Haiku polish. Bounded
 * by the active BudgetTracker the autopilot loop creates per-tick.
 *
- * Wired into the autopilot loop (step 4.6 in autopilot.ts), following
- * the same shape as `src/core/cycle/nightly-quality-probe.ts`
- * (v0.40.1.0 Track D / T6): the wiring resolves fixtures from the
- * gbrain package root, writes real outcomes to the parser-probe audit
- * trail (`audit-parser-probe.ts`), and never crashes the loop.
+ * **Wiring into the autopilot loop is deferred to a follow-up**
+ * (filed in TODOS.md). v0.41.16.0 ships the phase as a callable
+ * module so doctor + future cron drivers can invoke it; the
+ * scheduler wire-up follows the same shape as
+ * `src/core/cycle/nightly-quality-probe.ts` (v0.40.1.0 Track D / T6).
 *
 * Test seam: all dependencies are injected via NightlyProbeDeps so
 * unit tests don't touch real LLMs or real fixtures.
@@ -44,12 +44,7 @@ export const DEFAULT_DIMENSIONS: string[] = [
 * `--slot-a-model`, `--slot-b-model`, `--slot-c-model` on the CLI.
 */
 export const DEFAULT_SLOTS: SlotConfig[] = [
-  // Every default MUST be listed in its recipe's chat touchpoint (pinned by
-  // test/cross-modal-default-slots.test.ts) — `openai:gpt-4o` sat here after
-  // the OpenAI recipe dropped it, so slot A errored "not listed for OpenAI
-  // chat" on every install and the 3-slot panel could never reach its
-  // 2-model quorum without a Google key (verdict: permanently inconclusive).
-  { id: 'A', model: 'openai:gpt-5.2' },
+  { id: 'A', model: 'openai:gpt-4o' },
  { id: 'B', model: 'anthropic:claude-opus-4-7' },
  { id: 'C', model: 'google:gemini-1.5-pro' },
 ];
@@ -26,8 +26,8 @@
 */

 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
-import { chat as gatewayChat } from '../ai/gateway.ts';
-import { TIER_DEFAULTS } from '../model-config.ts';
+import { chat as gatewayChat, getChatModel } from '../ai/gateway.ts';
+import { resolveModel } from '../model-config.ts';
 import { gateVoice, type VoiceGateGenerator, type VoiceGateJudge } from '../calibration/voice-gate.ts';
 import { patternStatementTemplate, type PatternStatementSlots } from '../calibration/templates.ts';
 // v0.41 T10 — domain widening. The aggregator module resolves the active
@@ -229,7 +229,16 @@ class CalibrationProfilePhase extends BaseCyclePhase {
  ): Promise<{ summary: string; details: Record<string, unknown>; status?: PhaseStatus }> {
    const holder = opts.holder ?? 'garry';
    const promptVersion = opts.promptVersion ?? CALIBRATION_PROFILE_PROMPT_VERSION;
-    const modelId = opts.model ?? TIER_DEFAULTS.reasoning;
+    // Resolved once (see propose-takes.ts for the chain): models.calibration_profile
+    // > models.default > env > the gateway's chat model (itself resolved
+    // through models.chat + the reasoning tier). Provider-prefixed per #2451
+    // — a bare id would make gateway.chat() throw "missing a provider
+    // prefix". Drives the generator's chat call, the budget label, and the
+    // persisted model_id, so the three can never disagree.
+    const modelId = opts.model ?? await resolveModel(engine, {
+      configKey: 'models.calibration_profile',
+      fallback: getChatModel(),
+    });
    const gradeCompletion = opts.gradeCompletion ?? 1.0;
    const patternsGenerator = opts.patternsGenerator ?? defaultPatternsGenerator;
    const biasTagsGenerator = opts.biasTagsGenerator ?? defaultBiasTagsGenerator;
@@ -265,6 +274,7 @@ class CalibrationProfilePhase extends BaseCyclePhase {
        scorecard,
        holder,
        attempt,
+        modelHint: modelId,
        ...(feedback !== undefined ? { feedback } : {}),
      });
      return lines.join('\n');
@@ -36,7 +36,9 @@

 import { createHash } from 'node:crypto';
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
-import { chat as gatewayChat } from '../ai/gateway.ts';
+import { chat as gatewayChat, getChatModel } from '../ai/gateway.ts';
+import { resolveModel } from '../model-config.ts';
+import { splitProviderModelId } from '../model-id.ts';
 import { GBrainError } from '../types.ts';
 import type { OperationContext } from '../operations.ts';
 import type { BrainEngine, Take, TakeResolution } from '../engine.ts';
@@ -395,7 +397,24 @@ class GradeTakesPhase extends BaseCyclePhase {
    const autoResolve = opts.autoResolve ?? false; // D17 default OFF
    const autoResolveThreshold = opts.autoResolveThreshold ?? 0.95; // D12 conservative
    const resolvedByLabel = opts.resolvedByLabel ?? 'gbrain:grade_takes';
-    const judgeModelId = opts.model ?? 'claude-sonnet-4-6';
+    // Resolve the judge model ONCE (see propose-takes.ts for the chain —
+    // same label-vs-actual split fixed here: the judge call rode the
+    // gateway's chat_model while the grade cache key, evidence signature,
+    // and budget label recorded a hardcoded 4.6).
+    // NOTE: changing the resolved judge model invalidates the grade cache
+    // (judge_model_id is part of its key) — a one-time, budget-capped
+    // re-grade wave that is CORRECT, since the actual judge did change.
+    const judgeModelFull = opts.model ?? await resolveModel(engine, {
+      configKey: 'models.grade_takes',
+      fallback: getChatModel(),
+    });
+    // Bare tail for cache keys / evidence signatures / stored ids — the
+    // grade cache has always been keyed on bare ids; the gateway default
+    // resolves provider-prefixed, and normalizing preserves cache continuity
+    // on stock installs (no spurious re-judge wave from a prefix change). A
+    // genuinely different configured judge still invalidates, which is
+    // correct. The FULL string drives the actual judge call.
+    const judgeModelId = splitProviderModelId(judgeModelFull).model || judgeModelFull;

    const useEnsemble = opts.useEnsemble ?? false;
    const ensembleThreshold = opts.ensembleThreshold ?? 0.85;
@@ -468,7 +487,7 @@ class GradeTakesPhase extends BaseCyclePhase {
      // Call the single-model judge. Errors on a single take log warning + continue.
      let verdict: JudgeVerdict;
      try {
-        verdict = await judge({ take, evidence, modelHint: opts.model });
+        verdict = await judge({ take, evidence, modelHint: judgeModelFull });
      } catch (err) {
        const msg = err instanceof Error ? err.message : String(err);
        result.warnings.push(`judge failed on take ${take.id}: ${msg}`);
@@ -70,28 +70,6 @@ export async function runLongMemEvalForProbe(args: LongMemEvalProbeArgs): Promis
 * the batch input) or unparseable (cross-modal wrote garbage). Both
 * cases are paste-ready in the error message.
 */
-/**
- * QA-shaped judge dimensions for the nightly probe. The batch judge's
- * DEFAULT_DIMENSIONS rubric (DEPTH / SOURCING / SPECIFICITY / …) is built
- * for rich agent responses; LongMemEval hypotheses are deliberately terse
- * factual answers ("in widget-co") that can never score ≥7 on DEPTH or
- * SOURCING — so with the default rubric the probe FAILs every night even
- * when retrieval + answering are perfectly healthy. The probe owns its
- * invocation of the eval tool and passes dimensions matching the
- * fixture's QA shape instead.
- *
- * NOTE: the `--dimensions` CLI flag splits on commas, so these dimension
- * descriptions must stay comma-free.
- */
-export const PROBE_QA_DIMENSIONS: string[] = [
-  // No faithfulness/grounding dimension on purpose: the judge never sees
-  // the haystack, so any accurate detail beyond the terse gold label reads
-  // as "invented" and correct answers fail (verified empirically — a
-  // correct "before + dates" answer scored 4/10 on such a dimension).
-  'CORRECTNESS — Does the hypothesis state the same fact as the expected answer? A terse direct answer is ideal.',
-  'DIRECTNESS — Does it answer THIS question without hedging or padding or answering something else?',
-];
-
 export async function runCrossModalBatchForProbe(
  args: CrossModalProbeArgs,
 ): Promise<{ exitCode: number; summary: CrossModalBatchSummary }> {
@@ -103,8 +81,6 @@ export async function runCrossModalBatchForProbe(
    args.summaryPath,
    '--max-usd',
    String(args.maxUsd),
-    '--dimensions',
-    PROBE_QA_DIMENSIONS.join(','),
    '--yes',
    '--json',
  ]);
@@ -62,42 +62,6 @@ export interface NightlyProbeDeps {
  now: () => Date;
 }

-/**
- * Dual-plane flag resolution (same precedent as `mcp.publish_skills` in
- * serve-http.ts): the DB config row — what `gbrain config set` writes —
- * wins when present; the file plane (~/.gbrain/config.json) is the
- * fallback. Doctor's paste-ready enable hint says `gbrain config set
- * autopilot.nightly_quality_probe.enabled true`, so the gate MUST read
- * the DB plane — a file-only read turns that hint into a silent no-op.
- */
-export function resolveProbeEnabled(
-  dbVal: string | null | undefined,
-  fileVal: unknown,
-): boolean {
-  if (dbVal != null) return dbVal === 'true';
-  return fileVal === true;
-}
-
-/**
- * Same dual-plane rule for the per-run cost cap. Malformed or negative
- * values on either plane fall through to the next plane / the default.
- */
-export function resolveProbeMaxUsd(
-  dbVal: string | null | undefined,
-  fileVal: unknown,
-  fallback: number = DEFAULT_MAX_USD,
-): number {
-  if (dbVal != null) {
-    const n = Number(dbVal);
-    if (Number.isFinite(n) && n >= 0) return n;
-  }
-  if (fileVal != null) {
-    const n = Number(fileVal);
-    if (Number.isFinite(n) && n >= 0) return n;
-  }
-  return fallback;
-}
-
 /**
 * Pure function: decide whether the probe should run given the audit
 * history. Returns reason when skipping.
@@ -137,17 +101,21 @@ export async function runNightlyQualityProbe(deps: NightlyProbeDeps): Promise<Ni
    return { outcome: 'disabled', exit_code: 0, detail: 'feature flag off' };
  }

-  // 24h rate limit — skip WITHOUT an audit row. The autopilot loop invokes
-  // the probe every cycle (~5-10 min), so all but one invocation per day
-  // lands here; logging each skip floods the audit file (~hundreds of
-  // rows/day) and — because doctor treats any non-pass outcome as bad
-  // signal — flips nightly_quality_probe_health to a permanent WARN the
-  // moment the probe is enabled. A skip is a non-event: the real runs are
-  // the signal, and their rows are what gates the next 24h window.
+  // 24h rate limit — skip + audit "rate_limited".
  const now = deps.now();
  const recent = readRecentQualityProbeEvents(2, now); // 2-day window is enough for 24h check
  const decision = shouldRunNightly(now, recent);
  if (!decision.run) {
+    logQualityProbeEvent({
+      outcome: 'rate_limited',
+      exit_code: 0,
+      pass_count: 0,
+      fail_count: 0,
+      inconclusive_count: 0,
+      error_count: 0,
+      est_cost_usd: 0,
+      detail: 'already ran within 24h window',
+    });
    return { outcome: 'rate_limited', exit_code: 0, detail: 'already ran within 24h' };
  }

@@ -5,11 +5,16 @@
 * a tuned LLM extractor, writes the extracted gradeable claims to the
 * `take_proposals` queue. User accepts/rejects via `gbrain takes propose`.
 *
- * Idempotency contract (D17 schema spec):
- *   The unique index on (source_id, page_slug, content_hash, prompt_version)
- *   means an unchanged page never re-spends LLM tokens. Bumping
- *   PROPOSE_TAKES_PROMPT_VERSION cleanly invalidates the cache so a tuned
- *   prompt re-runs proposals on every page.
+ * Idempotency contract (D17 schema spec; per-claim rows since migration v125):
+ *   Every scan of a (source_id, page_slug, content_hash, prompt_version)
+ *   tuple leaves at least one row — one per extracted claim, or a single
+ *   status='empty' sentinel when extraction yields nothing — so an unchanged
+ *   page never re-spends LLM tokens. Pre-v125 only proposal rows were
+ *   written: a zero-claim page never entered the cache and was re-extracted
+ *   on EVERY cycle (observed live: ~60 such pages × every cycle ≈ 1,400
+ *   wasted extractor calls / ~$15 per day — ~90% of total autopilot spend).
+ *   Bumping PROPOSE_TAKES_PROMPT_VERSION cleanly invalidates the cache so a
+ *   tuned prompt re-runs proposals on every page.
 *
 * F2 fence dedup:
 *   The phase reads the page's existing `<!-- gbrain:takes:begin -->` fence
@@ -40,6 +45,7 @@
 import { randomUUID, createHash } from 'node:crypto';
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
 import { chat as gatewayChat, getChatModel } from '../ai/gateway.ts';
+import { resolveModel } from '../model-config.ts';
 import { writeReceipt } from '../extract/receipt-writer.ts';
 import { upsertExtractRollup } from '../extract/rollup-writer.ts';
 import { GBrainError } from '../types.ts';
@@ -307,6 +313,18 @@ class ProposeTakesPhase extends BaseCyclePhase {
    const promptVersion = opts.promptVersion ?? PROPOSE_TAKES_PROMPT_VERSION;
    const pageLimit = opts.pageLimit ?? 100;
    const skipPagesWithFence = opts.skipPagesWithFence ?? false;
+    // Resolve the extractor model ONCE: models.propose_takes >
+    // models.default > GBRAIN_MODEL env > the gateway's chat model (which
+    // reconfigureGatewayWithEngine already resolved through models.chat +
+    // the reasoning tier). One resolved provider-prefixed string drives the
+    // actual chat call, the budget estimate, AND the stored model_id — so
+    // the recorded model can never disagree with the model that ran (#2451
+    // convention: stored ids are provider-prefixed, nested prefixes like
+    // openrouter:anthropic/... stay intact).
+    const extractorModelId = opts.model ?? await resolveModel(engine, {
+      configKey: 'models.propose_takes',
+      fallback: getChatModel(),
+    });
    const proposalRunId = `propose-${new Date().toISOString().slice(0, 19).replace(/[-:T]/g, '')}-${randomUUID().slice(0, 8)}`;

    const result: ProposeTakesResult = {
@@ -330,8 +348,6 @@ class ProposeTakesPhase extends BaseCyclePhase {
      opts.reporter.start('propose_takes.pages' as never, pages.length);
    }

-    const modelId = opts.model ?? getChatModel();
-
    for (const page of pages) {
      result.pages_scanned += 1;
      this.tick(opts);
@@ -361,7 +377,7 @@ class ProposeTakesPhase extends BaseCyclePhase {

      // Budget pre-check before the LLM call. Estimate: ~1500 input tokens + 500 output.
      const budget = this.checkBudget({
-        modelId,
+        modelId: extractorModelId,
        estimatedInputTokens: 1500,
        maxOutputTokens: 500,
      });
@@ -380,7 +396,7 @@ class ProposeTakesPhase extends BaseCyclePhase {
          pagePath: page.slug,
          pageBody: body,
          existingTakes,
-          modelHint: opts.model,
+          modelHint: extractorModelId,
        });
      } catch (err) {
        const msg = err instanceof Error ? err.message : String(err);
@@ -388,16 +404,42 @@ class ProposeTakesPhase extends BaseCyclePhase {
        continue;
      }

-      // Write proposals to take_proposals. Each row is a separate INSERT
-      // because the composite idempotency key is on the per-page tuple — a
-      // bulk UPSERT would collapse a same-page-multi-claim run into one row.
-      for (const p of proposals) {
+      // Zero-claim scans MUST still enter the idempotency cache. Pre-v125
+      // only proposal rows were written, so a page whose extraction yielded
+      // no gradeable claims never got a row for its (page, content_hash) —
+      // the cache check above missed on every subsequent cycle and the LLM
+      // call was re-spent on the same unchanged page, forever. The sentinel
+      // row (status='empty', empty claim_text) is invisible to the review
+      // queue (pending_idx is partial on status='pending'); it exists only
+      // so the cache check hits.
+      if (proposals.length === 0) {
        await engine.executeRaw(
+          `INSERT INTO take_proposals
+             (source_id, page_slug, content_hash, prompt_version, proposal_run_id,
+              status, claim_text, kind, holder, weight, domain, dedup_against_fence_rows, model_id)
+           VALUES ($1, $2, $3, $4, $5, 'empty', '', 'none', 'brain', 0, NULL, NULL, $6)
+           ON CONFLICT (source_id, page_slug, content_hash, prompt_version, md5(claim_text)) DO NOTHING`,
+          [sourceId, page.slug, ch, promptVersion, proposalRunId, extractorModelId],
+        );
+        continue;
+      }
+
+      // Write proposals to take_proposals, one row per claim. The v125
+      // idempotency index includes md5(claim_text), so a same-page
+      // multi-claim run keeps EVERY claim — the pre-v125 four-column unique
+      // index made claims 2..N conflict with claim 1 and ON CONFLICT DO
+      // NOTHING silently dropped them (the review queue only ever saw the
+      // first claim of each page version). RETURNING id keeps
+      // proposals_inserted honest: it counts rows that actually landed,
+      // not insert attempts.
+      for (const p of proposals) {
+        const landed = await engine.executeRaw<{ id: number }>(
          `INSERT INTO take_proposals
             (source_id, page_slug, content_hash, prompt_version, proposal_run_id,
              claim_text, kind, holder, weight, domain, dedup_against_fence_rows, model_id)
           VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12)
-           ON CONFLICT (source_id, page_slug, content_hash, prompt_version) DO NOTHING`,
+           ON CONFLICT (source_id, page_slug, content_hash, prompt_version, md5(claim_text)) DO NOTHING
+           RETURNING id`,
          [
            sourceId,
            page.slug,
@@ -410,10 +452,10 @@ class ProposeTakesPhase extends BaseCyclePhase {
            p.weight,
            p.domain ?? null,
            JSON.stringify(existingTakes),
-            modelId,
+            extractorModelId,
          ],
        );
-        result.proposals_inserted += 1;
+        if (landed.length > 0) result.proposals_inserted += 1;
      }
    }

@@ -58,8 +58,11 @@ const SUMMARY_SLUG_RE = /^[a-z0-9][a-z0-9\-]*(\/[a-z0-9][a-z0-9\-]*)*$/;
 * resolver returns for known Anthropic aliases.
 */
 const MODEL_CONTEXT_TOKENS: Record<string, number> = {
+  'claude-fable-5': 1_000_000,
+  'claude-opus-4-8': 1_000_000,
  'claude-opus-4-7': 1_000_000,
  'claude-opus-4-6': 1_000_000,
+  'claude-sonnet-5': 1_000_000,
  'claude-sonnet-4-6': 200_000,
  'claude-sonnet-4-5': 200_000,
  'claude-haiku-4-5-20251001': 200_000,
@@ -5671,6 +5671,41 @@ export const MIGRATIONS: Migration[] = [
 `);
    },
  },
+  {
+    version: 125,
+    name: 'take_proposals_empty_scan_sentinels_and_per_claim_rows',
+    // v0.42.x — kill the propose_takes rescan loop + stop dropping claims.
+    //
+    // Two defects, one schema touch:
+    //   1. Zero-claim scans never entered the idempotency cache (only
+    //      proposal rows were written), so pages whose extraction yielded
+    //      nothing were re-extracted on EVERY cycle. Observed live: ~60
+    //      such pages per run ≈ 1,400 wasted extractor calls / ~$15 per
+    //      day — ~90% of total autopilot LLM spend. Fix: the phase now
+    //      writes a status='empty' sentinel row per zero-claim scan; the
+    //      status CHECK gains the 'empty' value. Sentinels are excluded
+    //      from the partial pending index, so the review queue never sees
+    //      them.
+    //   2. The 4-column unique index collapsed a same-page multi-claim run
+    //      to its FIRST claim: rows 2..N conflicted and ON CONFLICT DO
+    //      NOTHING silently dropped them (verified live: exactly one row
+    //      per (page, hash) across 3 days of runs). Fix: the idempotency
+    //      index gains md5(claim_text) — per-claim rows, while the
+    //      4-column prefix still serves the per-scan cache lookup.
+    //
+    // Existing data is index-safe by construction: the old index guaranteed
+    // at most one row per 4-tuple, so the widened index has no duplicates
+    // to trip on. The DROP+ADD CONSTRAINT pair is idempotent as a unit.
+    idempotent: true,
+    sql: `
+      ALTER TABLE take_proposals DROP CONSTRAINT IF EXISTS take_proposals_status_check;
+      ALTER TABLE take_proposals ADD CONSTRAINT take_proposals_status_check
+        CHECK (status IN ('pending','accepted','rejected','superseded','empty'));
+      DROP INDEX IF EXISTS take_proposals_idempotency_idx;
+      CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
+        ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
+    `,
+  },
 ];

 export const LATEST_VERSION = MIGRATIONS.length > 0
@@ -75,11 +75,6 @@ export const CANONICAL_PRICING: Record<string, ModelPricing> = {
  'openai:gpt-4o':                        { input:  2.50, output: 10.00 },
  'openai:gpt-4o-mini':                   { input:  0.15, output:  0.60 },
  'openai:gpt-5':                         { input:  5.00, output: 20.00 },
-  // gpt-5.2: rates from the OpenAI recipe chat touchpoint (verified
-  // 2026-04-20). Needed here because it's the cross-modal DEFAULT_SLOTS
-  // slot-A model — without a canonical entry estimateCost silently drops
-  // slot A from the --max-usd pre-flight and est_cost_usd audit rows.
-  'openai:gpt-5.2':                       { input:  1.25, output: 10.00 },
  'openai:gpt-5.5':                       { input:  4.00, output: 16.00 },

  // ── Google ─────────────────────────────────────────────────────────────
@@ -762,7 +762,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
  proposed_at                 TIMESTAMPTZ  NOT NULL DEFAULT now(),
  proposal_run_id             TEXT         NOT NULL,
  status                      TEXT         NOT NULL DEFAULT 'pending'
-                                           CHECK (status IN ('pending','accepted','rejected','superseded')),
+                                           CHECK (status IN ('pending','accepted','rejected','superseded','empty')),
  claim_text                  TEXT         NOT NULL,
  kind                        TEXT         NOT NULL,
  holder                      TEXT         NOT NULL,
@@ -777,7 +777,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
  predicted_brier_bucket_n    INTEGER
 );
 CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-  ON take_proposals (source_id, page_slug, content_hash, prompt_version);
+  ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
 CREATE INDEX IF NOT EXISTS take_proposals_pending_idx
  ON take_proposals (source_id, status, proposed_at DESC)
  WHERE status = 'pending';
@@ -1274,8 +1274,14 @@ CREATE INDEX IF NOT EXISTS calibration_profiles_published_idx
  WHERE published = true;

 -- take_proposals: propose_takes phase queue. Idempotency cache via the
-- composite unique index (source_id, page_slug, content_hash, prompt_version)
-- mirrors v0.23 dream_verdicts. proposal_run_id supports --rollback by run.
+-- composite unique index (source_id, page_slug, content_hash, prompt_version,
+-- md5(claim_text)) — the 4-column prefix is the per-scan cache key (mirrors
+-- v0.23 dream_verdicts); md5(claim_text) makes rows per-claim so multi-claim
+-- pages keep every claim (v125). status='empty' rows are zero-claim scan
+-- sentinels: they hold the cache slot for a page version whose extraction
+-- yielded nothing — the phase never re-spends the LLM call on that page
+-- version. Excluded from the partial pending index. proposal_run_id supports
+-- --rollback by run.
 CREATE TABLE IF NOT EXISTS take_proposals (
  id                          BIGSERIAL PRIMARY KEY,
  source_id                   TEXT         NOT NULL REFERENCES sources(id) ON DELETE CASCADE,
@@ -1286,7 +1292,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
  proposed_at                 TIMESTAMPTZ  NOT NULL DEFAULT now(),
  proposal_run_id             TEXT         NOT NULL,
  status                      TEXT         NOT NULL DEFAULT 'pending'
-                                           CHECK (status IN ('pending','accepted','rejected','superseded')),
+                                           CHECK (status IN ('pending','accepted','rejected','superseded','empty')),
  claim_text                  TEXT         NOT NULL,
  kind                        TEXT         NOT NULL,
  holder                      TEXT         NOT NULL,
@@ -1301,7 +1307,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
  predicted_brier_bucket_n    INTEGER
 );
 CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-  ON take_proposals (source_id, page_slug, content_hash, prompt_version);
+  ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
 CREATE INDEX IF NOT EXISTS take_proposals_pending_idx
  ON take_proposals (source_id, status, proposed_at DESC)
  WHERE status = 'pending';
@@ -1270,8 +1270,14 @@ CREATE INDEX IF NOT EXISTS calibration_profiles_published_idx
  WHERE published = true;

 -- take_proposals: propose_takes phase queue. Idempotency cache via the
-- composite unique index (source_id, page_slug, content_hash, prompt_version)
-- mirrors v0.23 dream_verdicts. proposal_run_id supports --rollback by run.
+-- composite unique index (source_id, page_slug, content_hash, prompt_version,
+-- md5(claim_text)) — the 4-column prefix is the per-scan cache key (mirrors
+-- v0.23 dream_verdicts); md5(claim_text) makes rows per-claim so multi-claim
+-- pages keep every claim (v125). status='empty' rows are zero-claim scan
+-- sentinels: they hold the cache slot for a page version whose extraction
+-- yielded nothing — the phase never re-spends the LLM call on that page
+-- version. Excluded from the partial pending index. proposal_run_id supports
+-- --rollback by run.
 CREATE TABLE IF NOT EXISTS take_proposals (
  id                          BIGSERIAL PRIMARY KEY,
  source_id                   TEXT         NOT NULL REFERENCES sources(id) ON DELETE CASCADE,
@@ -1282,7 +1288,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
  proposed_at                 TIMESTAMPTZ  NOT NULL DEFAULT now(),
  proposal_run_id             TEXT         NOT NULL,
  status                      TEXT         NOT NULL DEFAULT 'pending'
-                                           CHECK (status IN ('pending','accepted','rejected','superseded')),
+                                           CHECK (status IN ('pending','accepted','rejected','superseded','empty')),
  claim_text                  TEXT         NOT NULL,
  kind                        TEXT         NOT NULL,
  holder                      TEXT         NOT NULL,
@@ -1297,7 +1303,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
  predicted_brier_bucket_n    INTEGER
 );
 CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-  ON take_proposals (source_id, page_slug, content_hash, prompt_version);
+  ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
 CREATE INDEX IF NOT EXISTS take_proposals_pending_idx
  ON take_proposals (source_id, status, proposed_at DESC)
  WHERE status = 'pending';
@@ -1,102 +0,0 @@
-/**
- * Tests for the parser-probe audit trail + the 24h rate-limit gate.
- *
- * Uses GBRAIN_AUDIT_DIR override pointed at a tmpdir for hermeticity
- * (same pattern as audit-slug-fallback.serial.test.ts). Serial because
- * the env override is process-global.
- */
-import { afterEach, beforeEach, describe, expect, test } from 'bun:test';
-import { mkdtempSync, rmSync, readdirSync } from 'node:fs';
-import { tmpdir } from 'node:os';
-import { join } from 'node:path';
-
-import {
-  computeParserProbeAuditFilename,
-  logParserProbeEvent,
-  parserProbeRanWithin,
-  readRecentParserProbeEvents,
-  type ParserProbeAuditEvent,
-} from '../src/core/audit-parser-probe.ts';
-
-let auditDir: string;
-let savedEnv: string | undefined;
-
-beforeEach(() => {
-  auditDir = mkdtempSync(join(tmpdir(), 'parser-probe-audit-'));
-  savedEnv = process.env.GBRAIN_AUDIT_DIR;
-  process.env.GBRAIN_AUDIT_DIR = auditDir;
-});
-
-afterEach(() => {
-  if (savedEnv === undefined) delete process.env.GBRAIN_AUDIT_DIR;
-  else process.env.GBRAIN_AUDIT_DIR = savedEnv;
-  rmSync(auditDir, { recursive: true, force: true });
-});
-
-function makeEvent(overrides: Partial<ParserProbeAuditEvent> = {}): ParserProbeAuditEvent {
-  return {
-    schema_version: 1,
-    ts: new Date().toISOString(),
-    outcome: 'pass',
-    fixtures_total: 12,
-    fixtures_passed: 12,
-    recall_mean: 0.98,
-    participants_recall_mean: 0.97,
-    adversarial_false_positives: 0,
-    failed_fixture_ids: [],
-    ...overrides,
-  };
-}
-
-describe('parser-probe audit trail', () => {
-  test('log + readRecent round-trip', () => {
-    logParserProbeEvent(makeEvent({ outcome: 'fail', reason: '2 fixture(s) failed' }));
-    const events = readRecentParserProbeEvents(7);
-    expect(events.length).toBe(1);
-    expect(events[0]!.outcome).toBe('fail');
-    expect(events[0]!.reason).toBe('2 fixture(s) failed');
-    const files = readdirSync(auditDir);
-    expect(files.length).toBe(1);
-    expect(files[0]).toMatch(/^parser-probe-\d{4}-W\d{2}\.jsonl$/);
-  });
-
-  test('filename uses ISO-week rotation with the parser-probe prefix', () => {
-    // Year-boundary edge pinned by the shared writer's own tests; here we
-    // pin the prefix wiring.
-    expect(computeParserProbeAuditFilename(new Date('2026-07-06T12:00:00Z'))).toBe(
-      'parser-probe-2026-W28.jsonl',
-    );
-  });
-
-  test('readRecent filters by window', () => {
-    const old = new Date(Date.now() - 10 * 86400000).toISOString();
-    logParserProbeEvent(makeEvent({ ts: old }));
-    expect(readRecentParserProbeEvents(7).length).toBe(0);
-  });
-});
-
-describe('parserProbeRanWithin — 24h rate-limit gate', () => {
-  const DAY_MS = 24 * 60 * 60 * 1000;
-
-  test('false when no runs are audited', () => {
-    expect(parserProbeRanWithin(DAY_MS)).toBe(false);
-  });
-
-  test('true when a run landed within the window', () => {
-    logParserProbeEvent(makeEvent({ ts: new Date(Date.now() - 60_000).toISOString() }));
-    expect(parserProbeRanWithin(DAY_MS)).toBe(true);
-  });
-
-  test('false when the last run is older than the window', () => {
-    logParserProbeEvent(makeEvent({ ts: new Date(Date.now() - 25 * 3600_000).toISOString() }));
-    expect(parserProbeRanWithin(DAY_MS)).toBe(false);
-  });
-
-  test('non-pass outcomes also hold the window (mirrors quality-probe semantics)', () => {
-    logParserProbeEvent(makeEvent({
-      outcome: 'no_embedding_key',
-      ts: new Date(Date.now() - 3600_000).toISOString(),
-    }));
-    expect(parserProbeRanWithin(DAY_MS)).toBe(true);
-  });
-});
@@ -31,15 +31,10 @@ describe('autopilot wiring: nightly quality probe', () => {
    expect(SOURCE).toContain(`runCrossModalBatchForProbe`);
  });

-  test('feature flag gate present: dual-plane read (DB row wins, file plane fallback)', () => {
+  test('feature flag gate present: cfg.autopilot.nightly_quality_probe.enabled', () => {
    // Per D10: the scheduler ONLY checks the feature flag. The 24h rate-limit
    // lives inside runNightlyQualityProbe itself (no scheduler-side precheck).
-    // The flag resolves through resolveProbeEnabled so `gbrain config set
-    // autopilot.nightly_quality_probe.enabled true` (the doctor hint, DB
-    // plane) and ~/.gbrain/config.json (file plane) BOTH work — a file-only
-    // read made the printed hint a silent no-op.
-    expect(SOURCE).toContain(`getConfig('autopilot.nightly_quality_probe.enabled')`);
-    expect(SOURCE).toMatch(/resolveProbeEnabled\(dbEnabled,\s*cfg\?\.autopilot\?\.nightly_quality_probe\?\.enabled\)/);
+    expect(SOURCE).toContain(`nightly_quality_probe?.enabled === true`);
  });

  test('NO scheduler-side rate-limit check (D10 simplification)', () => {
@@ -69,23 +64,12 @@ describe('autopilot wiring: nightly quality probe', () => {
    expect(SOURCE).toContain(`now:`);
  });

-  test('resolveRepoRoot prefers the gbrain package root (committed fixture home), not the brain repoPath', () => {
-    // The DI harness in nightly-quality-probe.test.ts passes process.cwd()
-    // (= the gbrain repo in CI), which papered over the wiring passing
-    // repoPath (= sync.repo_path, the user's BRAIN repo, where the fixture
-    // never exists). Pin the package-root resolution + existence check.
-    expect(SOURCE).toMatch(/fileURLToPath\(new URL\('\.\.\/\.\.', import\.meta\.url\)\)/);
-    expect(SOURCE).toContain(`'longmemeval-nightly.jsonl'`);
-    expect(SOURCE).toMatch(/fixtureAtPkgRoot \? pkgRoot : repoPath/);
-  });
-
  test('hasEmbeddingProvider reads from gateway.isAvailable("embedding") (codex round-2 #12 — in-process, not subprocess)', () => {
    expect(SOURCE).toContain(`isAvailable('embedding')`);
    expect(SOURCE).toContain(`gateway`);
  });

-  test('max_usd resolves dual-plane (default = 5 pinned by resolveProbeMaxUsd unit tests)', () => {
-    expect(SOURCE).toContain(`getConfig('autopilot.nightly_quality_probe.max_usd')`);
-    expect(SOURCE).toMatch(/resolveProbeMaxUsd\(dbMaxUsd,\s*cfg\?\.autopilot\?\.nightly_quality_probe\?\.max_usd\)/);
+  test('max_usd default = 5 when config unset (matches plan default per D10)', () => {
+    expect(SOURCE).toMatch(/max_usd\s*\?\?\s*5/);
  });
 });
@@ -1,77 +0,0 @@
-/**
- * Source-shape regression tests for the autopilot wiring of
- * `runConversationParserNightlyProbe` (step 4.6).
- *
- * Same rationale as autopilot-nightly-probe-wiring.test.ts: the loop is
- * hard to drive end-to-end, so these pin the structural protections —
- * the dual-plane flag read, the D10 tokenmax mode-gate, the package-root
- * fixture resolution, the audit-flood guard, and the try/catch posture.
- *
- * The probe's own gate/scoring logic is pinned by the module's unit
- * tests; the audit trail by audit-parser-probe.serial.test.ts.
- */
-
-import { describe, test, expect } from 'bun:test';
-import { readFileSync } from 'node:fs';
-import { resolve } from 'node:path';
-
-const AUTOPILOT_SRC = resolve('src/commands/autopilot.ts');
-const SOURCE = readFileSync(AUTOPILOT_SRC, 'utf-8');
-
-describe('autopilot wiring: conversation-parser probe', () => {
-  test('invokes the phase module and the audit trail', () => {
-    expect(SOURCE).toContain(`runConversationParserNightlyProbe`);
-    expect(SOURCE).toContain(`conversation-parser/nightly-probe`);
-    expect(SOURCE).toContain(`logParserProbeEvent`);
-    expect(SOURCE).toContain(`audit-parser-probe`);
-  });
-
-  test('flag reads dual-plane: DB row (gbrain config set) wins, file plane fallback', () => {
-    expect(SOURCE).toContain(`getConfig('autopilot.conversation_parser_probe.enabled')`);
-    expect(SOURCE).toContain(`cfg?.autopilot?.conversation_parser_probe?.enabled === true`);
-  });
-
-  test('D10 mode-gate present: tokenmax brains run the probe by default', () => {
-    expect(SOURCE).toMatch(/parserEnabled \|\| searchMode === 'tokenmax'/);
-  });
-
-  test('fixtures resolve from the gbrain package root, NOT the brain repoPath', () => {
-    // The committed fixtures live in the gbrain source tree; resolving
-    // them against sync.repo_path would point into the user's brain repo.
-    expect(SOURCE).toMatch(/fileURLToPath\(new URL\('\.\.\/\.\.', import\.meta\.url\)\)/);
-    expect(SOURCE).toContain(`'conversation-formats', 'all.jsonl'`);
-    expect(SOURCE).toContain(`'conversation-formats', 'adversarial.jsonl'`);
-  });
-
-  test('missing fixtures skip quietly (no audit row, once-per-process stderr note)', () => {
-    // Compiled-binary installs carry no source tree; writing failure rows
-    // would flip doctor to WARN on every binary install.
-    expect(SOURCE).toContain(`parserProbeFixtureWarned`);
-  });
-
-  test('rate_limited outcomes are NOT audit-logged (flood guard)', () => {
-    expect(SOURCE).toMatch(/outcome !== 'rate_limited'\) logParserProbeEvent\(result\)/);
-  });
-
-  test('rate-limit gate delegates to the audit module, not inline event reads', () => {
-    expect(SOURCE).toContain(`parserProbeRanWithin(24 * 60 * 60 * 1000)`);
-  });
-
-  test('LLM-key gate reads gateway.isAvailable("chat") in-process', () => {
-    expect(SOURCE).toContain(`isAvailable('chat')`);
-  });
-
-  test('probe call wrapped in try/catch that does NOT bump consecutiveErrors', () => {
-    expect(SOURCE).toMatch(/catch[\s\S]*?autopilot\.parser_probe[\s\S]*?do NOT bump consecutiveErrors/);
-  });
-
-  test('DI shape: the exact 7 fields of the parser probe NightlyProbeDeps', () => {
-    expect(SOURCE).toContain(`isEnabled:`);
-    expect(SOURCE).toContain(`searchMode:`);
-    expect(SOURCE).toContain(`hasLlmKey:`);
-    expect(SOURCE).toContain(`resolveFixturePath:`);
-    expect(SOURCE).toContain(`resolveAdversarialPath:`);
-    expect(SOURCE).toContain(`shouldSkipForRateLimit:`);
-    expect(SOURCE).toContain(`now:`);
-  });
-});
@@ -38,6 +38,7 @@ function buildMockEngine(opts: { scorecard: TakesScorecard }): {
 } {
  const captured: CapturedSql[] = [];
  const engine = {
+    async getConfig() { return null; },
    kind: 'pglite',
    async getScorecard() {
      return opts.scorecard;
@@ -1,47 +0,0 @@
-/**
- * Consistency guard: every cross-modal DEFAULT_SLOTS model must be listed
- * in its recipe's chat touchpoint. `openai:gpt-4o` drifted out of the
- * OpenAI recipe while remaining the slot-A default — the gateway then
- * rejected slot A ("not listed for OpenAI chat") on every install, and the
- * 3-slot judge panel could never reach its 2-model quorum without a Google
- * key, pinning every batch verdict at inconclusive (which the nightly
- * quality probe surfaces as a doctor WARN).
- */
-import { describe, expect, test } from 'bun:test';
-
-import { DEFAULT_SLOTS } from '../src/core/cross-modal-eval/runner.ts';
-import { getRecipe } from '../src/core/ai/recipes/index.ts';
-import { splitProviderModelId } from '../src/core/model-id.ts';
-import { canonicalLookup } from '../src/core/model-pricing.ts';
-
-describe('cross-modal DEFAULT_SLOTS ↔ recipe consistency', () => {
-  test('every default slot model is listed in its recipe chat touchpoint', () => {
-    for (const slot of DEFAULT_SLOTS) {
-      const { provider, model } = splitProviderModelId(slot.model);
-      expect(provider).not.toBeNull();
-      const recipe = getRecipe(provider!);
-      expect(recipe, `slot ${slot.id}: unknown recipe "${provider}"`).toBeDefined();
-      const chatModels = recipe!.touchpoints.chat?.models ?? [];
-      expect(
-        chatModels,
-        `slot ${slot.id}: "${model}" not listed for ${provider} chat — the judge slot can never run`,
-      ).toContain(model);
-    }
-  });
-
-  test('every default slot model has a canonical pricing entry', () => {
-    // Without one, estimateCost silently drops the slot from the
-    // --max-usd pre-flight and est_cost_usd audit rows (~1/3 under-count).
-    for (const slot of DEFAULT_SLOTS) {
-      expect(
-        canonicalLookup(slot.model),
-        `slot ${slot.id}: "${slot.model}" missing from CANONICAL_PRICING`,
-      ).toBeDefined();
-    }
-  });
-
-  test('slots span three distinct providers (uncorrelated blind spots)', () => {
-    const providers = new Set(DEFAULT_SLOTS.map(s => splitProviderModelId(s.model).provider));
-    expect(providers.size).toBe(3);
-  });
-});
@@ -1,51 +0,0 @@
-/**
- * Tests for computeConversationParserProbeHealthCheck — the pure function
- * behind doctor's conversation_parser_probe_health check, which replaced
- * the v0.41.13.0 hardcoded "Skipped" stub when the autopilot wiring
- * landed. Mirrors the branch coverage style of the quality-probe check.
- */
-import { describe, expect, test } from 'bun:test';
-
-import { computeConversationParserProbeHealthCheck } from '../src/commands/doctor.ts';
-
-const ev = (outcome: string, reason?: string, ts = new Date().toISOString()) => ({
-  outcome,
-  ts,
-  ...(reason !== undefined ? { reason } : {}),
-});
-
-describe('computeConversationParserProbeHealthCheck', () => {
-  test('disabled + no events → ok with paste-ready enable hint', () => {
-    const check = computeConversationParserProbeHealthCheck(false, []);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('gbrain config set autopilot.conversation_parser_probe.enabled true');
-  });
-
-  test('enabled + no events yet → ok, next run by autopilot', () => {
-    const check = computeConversationParserProbeHealthCheck(true, []);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('no probe events');
-  });
-
-  test('disabled flag but events exist (tokenmax mode-gate ran it) → events win over the hint', () => {
-    const check = computeConversationParserProbeHealthCheck(false, [ev('pass')]);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('all pass');
-  });
-
-  test('any non-pass outcome in the window → warn, latest surfaced with reason', () => {
-    const check = computeConversationParserProbeHealthCheck(true, [
-      ev('pass'),
-      ev('adversarial_false_positive', '1 adversarial fixture(s) parsed to non-empty'),
-    ]);
-    expect(check.status).toBe('warn');
-    expect(check.message).toContain('adversarial_false_positive');
-    expect(check.message).toContain('parsed to non-empty');
-  });
-
-  test('all pass → ok with run count', () => {
-    const check = computeConversationParserProbeHealthCheck(true, [ev('pass'), ev('pass')]);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('2 probe run(s)');
-  });
-});
@@ -47,6 +47,7 @@ function buildMockEngine(opts: { takes: Take[] }): {
  const captured: CapturedSql[] = [];
  const resolves: CapturedResolve[] = [];
  const engine = {
+    async getConfig() { return null; },
    kind: 'pglite',
    async listTakes() {
      return opts.takes;
@@ -46,12 +46,14 @@ interface CapturedResolve {
 function buildMockEngine(opts: {
  takes: Take[];
  cachedGrades?: Set<string>; // composite-key strings already in take_grade_cache
+  config?: Record<string, string>; // engine.getConfig plane (models.grade_takes etc.)
 }): { engine: BrainEngine; captured: CapturedSql[]; resolves: CapturedResolve[] } {
  const captured: CapturedSql[] = [];
  const resolves: CapturedResolve[] = [];
  const cached = opts.cachedGrades ?? new Set<string>();

  const engine = {
+    async getConfig(key: string) { return opts.config?.[key] ?? null; },
    kind: 'pglite',
    async listTakes() {
      return opts.takes;
@@ -224,6 +226,28 @@ describe('runPhaseGradeTakes — phase integration', () => {
    expect(resolves).toHaveLength(0); // no canonical mutation
  });

+  test('models.grade_takes config drives the judge call; cache key stays bare-tailed', async () => {
+    // Pre-fix the judge call rode the gateway's chat_model while the cache
+    // key / budget label recorded a hardcoded 'claude-sonnet-4-6'. The phase
+    // now resolves models.grade_takes; the judge gets the FULL string and
+    // the cache row keys on the bare tail (continuity with historical rows).
+    const takes = [buildTake({ id: 1, sinceDate: '2023-01-01' })];
+    const { engine, captured } = buildMockEngine({
+      takes,
+      config: { 'models.grade_takes': 'anthropic:claude-sonnet-5' },
+    });
+    const hints: Array<string | undefined> = [];
+    const judge: JudgeFn = async ({ modelHint }) => {
+      hints.push(modelHint);
+      return { verdict: 'correct', confidence: 0.9, reasoning: 'held' };
+    };
+    const result = await runPhaseGradeTakes(buildCtx(engine), { judge });
+    expect(result.status).toBe('ok');
+    expect(hints).toEqual(['anthropic:claude-sonnet-5']); // actual call gets the FULL string
+    const inserts = captured.filter(c => c.sql.includes('INSERT INTO take_grade_cache'));
+    expect(inserts[0]!.params[2]).toBe('claude-sonnet-5'); // judge_model_id is the bare tail
+  });
+
  test('D17: auto-resolve OFF by default — even high-confidence verdict does NOT mutate takes', async () => {
    const takes = [buildTake({ id: 1, sinceDate: '2023-01-01' })];
    const { engine, resolves } = buildMockEngine({ takes });
@@ -1,74 +0,0 @@
-// Regression test for the nightly-quality-probe config-plane split-brain.
-//
-// The doctor check prints a paste-ready enable hint — `gbrain config set
-// autopilot.nightly_quality_probe.enabled true` — which writes the DB config
-// plane. But both the autopilot gate and the doctor check used to read ONLY
-// the file plane (~/.gbrain/config.json via loadConfig), so following the
-// hint was a silent no-op: the probe never ran and doctor kept reporting
-// "disabled (opt-in)".
-//
-// resolveProbeEnabled / resolveProbeMaxUsd pin the dual-plane rule (same
-// precedent as `mcp.publish_skills` in serve-http.ts): DB row wins when
-// present, file plane is the fallback.
-import { describe, expect, test } from 'bun:test';
-
-import {
-  resolveProbeEnabled,
-  resolveProbeMaxUsd,
-} from '../src/core/cycle/nightly-quality-probe.ts';
-
-describe('resolveProbeEnabled — dual-plane flag resolution', () => {
-  test('DB plane "true" enables regardless of file plane (the doctor hint path)', () => {
-    expect(resolveProbeEnabled('true', undefined)).toBe(true);
-    expect(resolveProbeEnabled('true', false)).toBe(true);
-  });
-
-  test('explicit DB "false" wins over file-plane true (config set off sticks)', () => {
-    expect(resolveProbeEnabled('false', true)).toBe(false);
-  });
-
-  test('file plane is the fallback when no DB row exists', () => {
-    expect(resolveProbeEnabled(null, true)).toBe(true);
-    expect(resolveProbeEnabled(undefined, true)).toBe(true);
-    expect(resolveProbeEnabled(null, undefined)).toBe(false);
-    expect(resolveProbeEnabled(null, false)).toBe(false);
-  });
-
-  test('file plane stays strict boolean — string "true" in config.json does not enable', () => {
-    // Matches the pre-fix autopilot gate (`=== true`); the doctor check used
-    // Boolean(...) and could disagree with autopilot on a string value.
-    // Both call sites now share this helper, so they can no longer diverge.
-    expect(resolveProbeEnabled(null, 'true')).toBe(false);
-    expect(resolveProbeEnabled(null, 1)).toBe(false);
-  });
-
-  test('non-"true" DB strings are off (mcp.publish_skills semantics)', () => {
-    expect(resolveProbeEnabled('1', true)).toBe(false);
-    expect(resolveProbeEnabled('yes', true)).toBe(false);
-    expect(resolveProbeEnabled('', true)).toBe(false);
-  });
-});
-
-describe('resolveProbeMaxUsd — dual-plane cost cap resolution', () => {
-  test('DB plane wins when parseable', () => {
-    expect(resolveProbeMaxUsd('2.5', 10)).toBe(2.5);
-    expect(resolveProbeMaxUsd('0', 10)).toBe(0);
-  });
-
-  test('malformed or negative DB value falls through to file plane', () => {
-    expect(resolveProbeMaxUsd('banana', 3)).toBe(3);
-    expect(resolveProbeMaxUsd('-1', 3)).toBe(3);
-  });
-
-  test('file plane used when no DB row; default when both absent/invalid', () => {
-    expect(resolveProbeMaxUsd(null, 7)).toBe(7);
-    expect(resolveProbeMaxUsd(null, '4')).toBe(4);
-    expect(resolveProbeMaxUsd(null, undefined)).toBe(5);
-    expect(resolveProbeMaxUsd(null, 'banana')).toBe(5);
-    expect(resolveProbeMaxUsd(undefined, -2)).toBe(5);
-  });
-
-  test('explicit fallback override is honored', () => {
-    expect(resolveProbeMaxUsd(null, undefined, 12)).toBe(12);
-  });
-});
@@ -132,20 +132,17 @@ describe('runNightlyQualityProbe (DI stub harness)', () => {
    });
  });

-  test('enabled + recent run within 24h → outcome: rate_limited, NO audit row', async () => {
+  test('enabled + recent run within 24h → outcome: rate_limited', async () => {
    // Pre-seed a recent audit event by running the probe once first.
    await withEnv({ GBRAIN_AUDIT_DIR: auditTmp }, async () => {
      // First run succeeds.
      await runNightlyQualityProbe(makeDeps());
-      // Second run, same hour → rate_limited. A skip is a non-event: the
-      // autopilot loop invokes the probe every cycle (~5-10 min), so
-      // logging each skip would flood the audit file and flip doctor's
-      // any-non-pass-is-bad filter to a permanent WARN.
+      // Second run, same hour → rate_limited.
      const r2 = await runNightlyQualityProbe(makeDeps());
      expect(r2.outcome).toBe('rate_limited');
      const events = await readEvents();
-      expect(events.length).toBe(1);
-      expect(events[0].outcome).toBe('pass');
+      expect(events.length).toBe(2);
+      expect(events[1].outcome).toBe('rate_limited');
    });
  });

@@ -0,0 +1,160 @@
+/**
+ * propose_takes rescan-loop + dropped-claims regression tests (migration v125).
+ *
+ * Two live-observed defects, both fixed by the v125 schema + phase change:
+ *
+ *   1. RESCAN LOOP — a page whose extraction yielded zero claims never
+ *      entered the idempotency cache (only proposal rows were written), so
+ *      every cycle re-spent the extractor call on the same unchanged page.
+ *      Live impact: ~60 such pages × every cycle ≈ 1,400 wasted LLM calls
+ *      (~$15) per day — ~90% of total autopilot spend. Fix: status='empty'
+ *      sentinel row per zero-claim scan.
+ *
+ *   2. DROPPED CLAIMS — the 4-column unique index collapsed a same-page
+ *      multi-claim run to its first claim (rows 2..N conflicted, ON CONFLICT
+ *      DO NOTHING dropped them silently; verified live: exactly 1 row per
+ *      (page, hash) over 3 days). Fix: idempotency index gains
+ *      md5(claim_text).
+ *
+ * Hermetic: PGLite engine + injected extractor; no gateway, no LLM.
+ */
+import { describe, test, expect, beforeAll, afterAll } from 'bun:test';
+import { PGLiteEngine } from '../src/core/pglite-engine.ts';
+import { runPhaseProposeTakes, type ProposeTakesExtractor, type ProposedTake } from '../src/core/cycle/propose-takes.ts';
+import type { OperationContext } from '../src/core/operations.ts';
+
+let engine: PGLiteEngine;
+
+function ctx(): OperationContext {
+  return {
+    engine,
+    remote: false,
+    config: {} as OperationContext['config'],
+    logger: { info() {}, warn() {}, error() {}, debug() {} } as unknown as OperationContext['logger'],
+  } as unknown as OperationContext;
+}
+
+/** Extractor stub that counts invocations per page slug. */
+function countingExtractor(
+  claimsBySlug: Record<string, ProposedTake[]>,
+): { extractor: ProposeTakesExtractor; calls: string[] } {
+  const calls: string[] = [];
+  const extractor: ProposeTakesExtractor = async ({ pagePath }) => {
+    calls.push(pagePath);
+    return claimsBySlug[pagePath] ?? [];
+  };
+  return { extractor, calls };
+}
+
+beforeAll(async () => {
+  engine = new PGLiteEngine();
+  await engine.connect({});
+  await engine.initSchema();
+  await engine.putPage('notes/zero-claims', {
+    type: 'note',
+    title: 'pure narrative',
+    compiled_truth: 'A quiet walk in the park. Nothing opinionated happened at all today.',
+  });
+  await engine.putPage('notes/three-claims', {
+    type: 'note',
+    title: 'opinionated',
+    compiled_truth: 'I bet acme-example wins the market. widget-co will struggle. fund-a is overexposed.',
+  });
+});
+
+afterAll(async () => {
+  await engine.disconnect();
+});
+
+describe('rescan loop — zero-claim scans enter the cache', () => {
+  test('second run cache-hits: extractor is NOT called again on unchanged pages', async () => {
+    const claims = {
+      'notes/three-claims': [
+        { claim_text: 'acme-example wins the market', kind: 'bet' as const, holder: 'brain', weight: 0.7 },
+        { claim_text: 'widget-co will struggle', kind: 'take' as const, holder: 'brain', weight: 0.6 },
+        { claim_text: 'fund-a is overexposed', kind: 'take' as const, holder: 'brain', weight: 0.55 },
+      ],
+    };
+
+    const first = countingExtractor(claims);
+    const r1 = await runPhaseProposeTakes(ctx(), { extractor: first.extractor });
+    expect(r1.status).toBe('ok');
+    // Both pages extracted on the first pass.
+    expect(first.calls).toContain('notes/zero-claims');
+    expect(first.calls).toContain('notes/three-claims');
+
+    const second = countingExtractor(claims);
+    const r2 = await runPhaseProposeTakes(ctx(), { extractor: second.extractor });
+    expect(r2.status).toBe('ok');
+    // THE regression: pre-fix the zero-claim page missed the cache every
+    // run and was re-extracted here. (Run 1's receipt page legitimately
+    // appears once — it's a new page — and its zero-claim scan now caches
+    // too; pre-fix, receipts re-scanned forever as well.)
+    expect(second.calls).not.toContain('notes/zero-claims');
+    expect(second.calls).not.toContain('notes/three-claims');
+
+    // Run 2 inserted nothing, so no new receipt page exists: run 3 must be
+    // fully quiescent — zero extractor calls, zero cache misses.
+    const third = countingExtractor(claims);
+    const r3 = await runPhaseProposeTakes(ctx(), { extractor: third.extractor });
+    expect(r3.status).toBe('ok');
+    expect(third.calls).toEqual([]);
+    expect((r3.details as Record<string, unknown>).cache_misses).toBe(0);
+  });
+
+  test('zero-claim scan wrote an "empty" sentinel invisible to the pending queue', async () => {
+    const sentinel = await engine.executeRaw<{ status: string; claim_text: string }>(
+      `SELECT status, claim_text FROM take_proposals WHERE page_slug = 'notes/zero-claims'`,
+      [],
+    );
+    expect(sentinel.length).toBe(1);
+    expect(sentinel[0].status).toBe('empty');
+    expect(sentinel[0].claim_text).toBe('');
+    const pending = await engine.executeRaw<{ n: number }>(
+      `SELECT COUNT(*)::int AS n FROM take_proposals WHERE page_slug = 'notes/zero-claims' AND status = 'pending'`,
+      [],
+    );
+    expect(Number(pending[0].n)).toBe(0);
+  });
+});
+
+describe('dropped claims — per-claim rows survive the idempotency index', () => {
+  test('a 3-claim page stores 3 rows and reports an honest inserted count', async () => {
+    const rows = await engine.executeRaw<{ claim_text: string }>(
+      `SELECT claim_text FROM take_proposals WHERE page_slug = 'notes/three-claims' AND status = 'pending' ORDER BY id`,
+      [],
+    );
+    // Pre-fix the 4-column unique index kept only the FIRST claim.
+    expect(rows.length).toBe(3);
+    expect(rows.map(r => r.claim_text)).toEqual([
+      'acme-example wins the market',
+      'widget-co will struggle',
+      'fund-a is overexposed',
+    ]);
+  });
+
+  test('content change re-extracts and stores the new version separately', async () => {
+    await engine.putPage('notes/zero-claims', {
+      type: 'note',
+      title: 'pure narrative',
+      compiled_truth: 'Updated: I now believe acme-example is undervalued and will re-rate within a year.',
+    });
+    const claims = {
+      'notes/zero-claims': [
+        { claim_text: 'acme-example is undervalued', kind: 'take' as const, holder: 'brain', weight: 0.6 },
+      ],
+    };
+    const run = countingExtractor(claims);
+    const r = await runPhaseProposeTakes(ctx(), { extractor: run.extractor });
+    expect(r.status).toBe('ok');
+    // Changed page re-extracts; the unchanged 3-claim page stays cached.
+    expect(run.calls).toEqual(['notes/zero-claims']);
+    expect((r.details as Record<string, unknown>).proposals_inserted).toBe(1);
+    const all = await engine.executeRaw<{ status: string }>(
+      `SELECT status FROM take_proposals WHERE page_slug = 'notes/zero-claims' ORDER BY id`,
+      [],
+    );
+    // Old hash's sentinel + new hash's pending claim coexist.
+    expect(all.map(r2 => r2.status).sort()).toEqual(['empty', 'pending']);
+  });
+});
@@ -41,12 +41,16 @@ interface CapturedSql {
 function buildMockEngine(opts: {
  pages: Page[];
  existingProposals?: Set<string>; // composite-key strings already in take_proposals
+  config?: Record<string, string>; // engine.getConfig plane (models.tier.* etc.)
 }): { engine: BrainEngine; captured: CapturedSql[] } {
  const captured: CapturedSql[] = [];
  const existing = opts.existingProposals ?? new Set<string>();

  const engine = {
    kind: 'pglite',
+    async getConfig(key: string) {
+      return opts.config?.[key] ?? null;
+    },
    async listPages() {
      return opts.pages;
    },
@@ -59,7 +63,12 @@ function buildMockEngine(opts: {
        if (existing.has(key)) return [{ id: 1 } as unknown as T];
        return [];
      }
-      // INSERT — return nothing
+      // INSERT ... RETURNING id — emulate a successful insert so the
+      // honest proposals_inserted counter (counts RETURNING rows, not
+      // attempts) sees the row land. Conflicted inserts would return [].
+      if (sql.includes('INSERT INTO take_proposals') && sql.includes('RETURNING id')) {
+        return [{ id: 1 } as unknown as T];
+      }
      return [];
    },
  } as unknown as BrainEngine;
@@ -267,6 +276,29 @@ describe('runPhaseProposeTakes — phase integration', () => {
    expect(inserts[0]!.params[9]).toBe('market'); // domain
  });

+  test('extractor model resolves through models.propose_takes config', async () => {
+    // Pre-fix the phase's only model knob was the gateway chat model — there
+    // was no per-phase config key. The phase now resolves once via
+    // resolveModel(models.propose_takes > models.default > env > gateway
+    // chat model); the extractor hint and the stored model_id must both
+    // reflect the configured override (full provider-prefixed string, #2451).
+    const pages = [buildPage({ slug: 'wiki/concepts/tier-routing', body: 'Tier-routed models will win.' })];
+    const { engine, captured } = buildMockEngine({
+      pages,
+      config: { 'models.propose_takes': 'anthropic:claude-sonnet-5' },
+    });
+    const seen: Array<string | undefined> = [];
+    const extractor: ProposeTakesExtractor = async ({ modelHint }) => {
+      seen.push(modelHint);
+      return [{ claim_text: 'tier-routed models win', kind: 'bet', holder: 'brain', weight: 0.7 }];
+    };
+    const result = await runPhaseProposeTakes(buildCtx(engine), { extractor });
+    expect(result.status).toBe('ok');
+    expect(seen).toEqual(['anthropic:claude-sonnet-5']); // chat call gets the FULL string
+    const inserts = captured.filter(c => c.sql.includes('INSERT INTO take_proposals'));
+    expect(inserts[0]!.params[11]).toBe('anthropic:claude-sonnet-5'); // stored model_id matches the call
+  });
+
  test('cache hit: page already in take_proposals is skipped', async () => {
    const body = 'A page that was already processed.';
    const pages = [buildPage({ slug: 'wiki/old-page', body })];