fix(cycle): grade_takes and calibration_profile record the model they actually run — follow the gateway chat model (#3726)

Two cycle phases still carried the label-vs-actual defect that
propose_takes shed in v0.42.62 (and that the #2805 review noted was worth
tracking so it isn't lost):

- grade_takes hardcoded 'claude-sonnet-4-6' into judge_model_id, the
  evidence signature, and budget metering, while the default judge call
  passed NO model hint and rode the gateway's chat_model. On any brain
  with a non-default chat_model, the verdict cache and telemetry recorded
  a model that never ran.
- calibration_profile persisted TIER_DEFAULTS.reasoning to model_id but
  never passed a model to the patterns generator at all, so the recorded
  model and the executed one were unrelated.

Fix, matching the propose_takes convention: each phase resolves ONE
string via explicit override > getChatModel(), and that string drives the
chat call, the cache key, and the stored id.

- grade_takes: the judge hint gets the FULL provider-prefixed string; the
  stored judge_model_id and evidence signature keep the historical bare
  tail. Stock installs are unchanged: getChatModel() defaults to
  'anthropic:claude-sonnet-4-6', whose tail equals the old hardcoded
  value, so no verdict cache invalidates. A genuinely different
  chat_model invalidates, which is correct — the judge really changed.
- calibration_profile: getChatModel() is provider-prefixed, preserving
  the #2451 contract, and its default IS the old TIER_DEFAULTS.reasoning
  value — stock behavior unchanged. The generator now receives the same
  resolved string that is persisted to model_id.

Tests: regression per phase pinning configured-chat-model routing (full
string to the call, bare tail to the cache key for grade_takes, full
string persisted for calibration); all existing suites pass unchanged,
pinning the no-change-on-stock property. 112 tests across the four
affected suites; typecheck clean.

Co-authored-by: Paolo Belcastro <p3ob7o@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

src/core/cycle/calibration-profile.ts

@@ -27,8 +27,7 @@
 
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
 import { resolveOwnerHolder } from '../owner-holder.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 { 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
@@ -91,6 +90,8 @@ export type PatternStatementsGenerator = (input: {
   holder: string;
   attempt: number;
   feedback?: string;
+  /** Provider-prefixed model the phase resolved; drives the generator's chat call. */
+  modelHint?: string;
 }) => Promise<string[]>;
 
 /** Generator function for bias tags (test seam). */
@@ -233,7 +234,14 @@ class CalibrationProfilePhase extends BaseCyclePhase {
       configValue: await engine.getConfig('emotional_weight.user_holder'),
     });
     const promptVersion = opts.promptVersion ?? CALIBRATION_PROFILE_PROMPT_VERSION;
-    const modelId = opts.model ?? TIER_DEFAULTS.reasoning;
+    // Follow the gateway's configured chat model, matching propose_takes
+    // (v0.42.62) and grade_takes: previously the generator stayed pinned to
+    // the TIER_DEFAULTS.reasoning constant, ignoring a configured
+    // chat_model. getChatModel() is provider-prefixed, preserving the #2451
+    // contract (a bare id fed back into gateway.chat() throws), and its
+    // default IS 'anthropic:claude-sonnet-4-6' — identical to the old
+    // constant — so stock installs are unchanged.
+    const modelId = opts.model ?? getChatModel();
     const gradeCompletion = opts.gradeCompletion ?? 1.0;
     const patternsGenerator = opts.patternsGenerator ?? defaultPatternsGenerator;
     const biasTagsGenerator = opts.biasTagsGenerator ?? defaultBiasTagsGenerator;
@@ -269,6 +277,9 @@ class CalibrationProfilePhase extends BaseCyclePhase {
         scorecard,
         holder,
         attempt,
+        // The same resolved string that is persisted to model_id drives the
+        // generator's chat call — the phase can't record a model it didn't run.
+        modelHint: modelId,
         ...(feedback !== undefined ? { feedback } : {}),
       });
       return lines.join('\n');

src/core/cycle/grade-takes.ts

@@ -36,7 +36,8 @@
 
 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 { 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 +396,21 @@ 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';
+    // One resolved string drives the judge call, the verdict-cache key, and
+    // the stored judge_model_id — the convention propose_takes adopted in
+    // v0.42.62. Previously the default judge call passed NO model hint (it
+    // rode the gateway's chat_model) while 'claude-sonnet-4-6' was hardcoded
+    // into judge_model_id, the evidence signature, and budget metering — on
+    // brains with a different chat_model, telemetry priced and recorded a
+    // model that never ran.
+    const judgeModelFull = opts.model ?? getChatModel();
+    // Bare tail for the stored judge_model_id + evidence signature
+    // (historical convention). Stock installs are unchanged: getChatModel()
+    // defaults to 'anthropic:claude-sonnet-4-6', whose tail equals the old
+    // hardcoded value — zero verdict-cache invalidation. A genuinely
+    // different chat_model invalidates, which is correct: the judge really
+    // changed.
+    const judgeModelId = splitProviderModelId(judgeModelFull).model || judgeModelFull;
 
     const useEnsemble = opts.useEnsemble ?? false;
     const ensembleThreshold = opts.ensembleThreshold ?? 0.85;
@@ -468,7 +483,7 @@ 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}`);

test/calibration-profile.test.ts

@@ -355,3 +355,33 @@ describe('runPhaseCalibrationProfile — phase integration', () => {
     expect(result.summary).toContain('holder=people/charlie-example');
   });
 });
+
+describe('generator model follows the gateway chat model', () => {
+  test('configured chat_model drives the generator hint and the persisted model_id', async () => {
+    // Regression: the generator previously stayed pinned to the
+    // TIER_DEFAULTS.reasoning constant, ignoring a configured chat_model —
+    // unlike propose_takes (v0.42.62 convention). Stock behavior is
+    // unchanged (the gateway default equals the old constant).
+    const { configureGateway, resetGateway } = await import('../src/core/ai/gateway.ts');
+    configureGateway({ chat_model: 'openai:gpt-5', env: { OPENAI_API_KEY: 'test-key' } });
+    try {
+      const { engine, captured } = buildMockEngine({ scorecard: ENOUGH_RESOLVED_SCORECARD });
+      const hints: Array<string | undefined> = [];
+      const patternsGenerator: PatternStatementsGenerator = async ({ modelHint }) => {
+        hints.push(modelHint);
+        return ['You call early-stage tactics well — 8 of 10 held up.'];
+      };
+      await runPhaseCalibrationProfile(buildCtx(engine), {
+        patternsGenerator,
+        biasTagsGenerator: async () => [],
+        voiceGateJudge: passJudge,
+      });
+      expect(hints).toEqual(['openai:gpt-5']);
+      const insert = captured.find(c => c.sql.includes('INSERT INTO calibration_profiles'));
+      expect(insert).toBeDefined();
+      expect(insert!.params).toContain('openai:gpt-5'); // persisted model_id = full configured string
+    } finally {
+      resetGateway();
+    }
+  });
+});

test/grade-takes.test.ts

@@ -328,3 +328,39 @@ describe('runPhaseGradeTakes — phase integration', () => {
     expect((details.warnings as string[])[0]).toContain('judge timeout');
   });
 });
+
+// ─── judge model follows the gateway chat model ─────────────────────
+
+describe('judge model follows the gateway chat model (label = actual)', () => {
+  test('configured chat_model drives the judge hint (full string) and the stored judge_model_id (bare tail)', async () => {
+    // Regression: the default judge call previously passed NO model hint
+    // (riding the gateway's chat_model) while 'claude-sonnet-4-6' was
+    // hardcoded into judge_model_id + the evidence signature — on brains
+    // with a different chat_model, the cache and telemetry recorded a model
+    // that never ran.
+    const { configureGateway, resetGateway } = await import('../src/core/ai/gateway.ts');
+    configureGateway({ chat_model: 'openai:gpt-5', env: { OPENAI_API_KEY: 'test-key' } });
+    try {
+      const takes = [buildTake({ id: 1, sinceDate: '2023-01-01' })];
+      const { engine, captured } = buildMockEngine({ takes });
+      const hints: Array<string | undefined> = [];
+      const judge: JudgeFn = async ({ modelHint }) => {
+        hints.push(modelHint);
+        return { verdict: 'correct', confidence: 0.9, reasoning: 'held' };
+      };
+      const evidenceRetriever: EvidenceRetrieverFn = async () => 'evidence body';
+      const result = await runPhaseGradeTakes(buildCtx(engine), { judge, evidenceRetriever });
+      expect(result.status).toBe('ok');
+      expect(hints).toEqual(['openai:gpt-5']); // the chat call gets the FULL string
+      const inserts = captured.filter(c => c.sql.includes('INSERT INTO take_grade_cache'));
+      expect(inserts).toHaveLength(1);
+      expect(inserts[0]!.params[2]).toBe('gpt-5'); // stored judge_model_id is the bare tail
+      // The evidence signature is keyed on the same bare tail, so a stock
+      // install (default chat model tail == the old hardcoded value) sees
+      // zero cache invalidation from this change.
+      expect(inserts[0]!.params[3]).toBe(evidenceSignature('evidence body', 'gpt-5'));
+    } finally {
+      resetGateway();
+    }
+  });
+});