@bongos/core 1.19.1036 → 1.19.1038

package/modules/grading/grader-score.js

@@ -12,7 +12,7 @@
 const path = require('node:path');
 const fs = require('node:fs');
 const { DEFAULT_GRADER_MODEL, FALLTHROUGH_CONFIDENCE, LOW_CONFIDENCE, THRESHOLD, VERSION, permissionPathCheck, routeRankCheck } = require('./grader-config.js');
-const { TRIVIAL_MAX_FILES, TRIVIAL_MAX_LINES, clamp, median, r1 } = require('./grader-model.js');
+const { TRIVIAL_MAX_FILES, TRIVIAL_MAX_LINES, clamp, r1 } = require('./grader-model.js');
 const { extractJson } = require('./grader-prompt.js');
 const { issueClaimsRuntimeDead } = require('./dead-feature-claim.js');
 
@@ -612,173 +612,6 @@ function serverSideGuards(gradeResult, { committedFiles, builderRank, diffStats,
   };
 }
 
-// ---------- v3: multi-shot aggregation ----------
-
-// Collapses N per-shot grade-results into the final v3 grade-result.
-// Pass requires unanimous PASS on all shots; anything else (any FAIL, any
-// error, any disagreement) → not passed. Issues are deduped across shots
-// by (file, summary). Displayed rubric_scores use the median across shots
-// so the trend dashboard stays one-number-per-task.
-function aggregateShots(shotResults, meta = {}) {
-  const N = shotResults.length;
-  const verdicts = shotResults.map((s) => ({
-    passed: s.passed,
-    ff: s.rubric_scores?.functional_fidelity ?? null,
-    cq: s.rubric_scores?.code_quality ?? null,
-    economy: s.rubric_scores?.economy ?? null,
-    cq_effective: s.rubric_scores?.code_quality_effective ?? null,
-    confidence: s.signals?.confidence ?? null,
-    model: s.signals?.model ?? null,
-    run_ms: s.signals?.run_ms ?? null,
-    shot_index: s.signals?.shot_index ?? null,
-    notes_md: s.notes_md,
-    issues: s.issues || [],
-    error:
-      s.signals?.parse_error
-        ? 'parse_error'
-        : s.signals?.schema_error
-          ? 'schema_error'
-          : s.signals?.subprocess_error || null,
-  }));
-
-  const ff = median(verdicts.map((v) => v.ff));
-  const cq = median(verdicts.map((v) => v.cq));
-  const economy = median(verdicts.map((v) => v.economy));
-  const cqEffective = median(verdicts.map((v) => v.cq_effective));
-  // NB (task 1003318): this v3 multi-shot path still collapses a null axis to
-  // 0 — it serves the trivial-diff single shot, which cannot overflow a window;
-  // the panel path's lens-outage handling lives in validateAndScore.
-  const blended = ff != null && cq != null ? r1((ff + cq) / 2) : 0;
-  // Top-level aggregated confidence = median of per-shot confidences. Without
-  // this, ship.js forwards a null top-level confidence to /tasks/:id/grade,
-  // and the server's validateAndScore (re-running on the POST body) hits the
-  // Number(null)=0 trap → FAIL even though every shot independently passed.
-  // Caught on the MAS Phase 1 self-ship (#395).
-  //
-  // Computed UNROUNDED (no r1()) — confidence sits near a 0.4 gate cap so
-  // rounding to one decimal could nudge a borderline value across the cap
-  // in either direction. Per-shot confidences are already 0-1 floats.
-  const confVals = verdicts
-    .map((v) => v.confidence)
-    .filter((n) => Number.isFinite(n))
-    .sort((a, b) => a - b);
-  const confidenceMedian = confVals.length === 0
-    ? null
-    : confVals.length % 2
-      ? confVals[(confVals.length - 1) / 2]
-      : (confVals[confVals.length / 2 - 1] + confVals[confVals.length / 2]) / 2;
-
-  // Dedupe issues across shots by (file, summary). Keep first-seen severity.
-  const seen = new Set();
-  const issues = [];
-  for (const v of verdicts) {
-    for (const it of v.issues) {
-      if (!it || typeof it !== 'object') continue;
-      const key = `${it.file ?? ''}::${(it.summary ?? '').trim()}`;
-      if (seen.has(key)) continue;
-      seen.add(key);
-      issues.push(it);
-    }
-  }
-
-  const passCount = verdicts.filter((v) => v.passed === true).length;
-  const failCount = verdicts.filter((v) => v.passed === false && !v.error).length;
-  const errorCount = verdicts.filter((v) => v.error).length;
-  const allPassed = N > 0 && passCount === N;
-
-  let verdict;
-  let passed;
-  let aggregationKind;
-  if (errorCount > 0) {
-    passed = false;
-    verdict = `FAIL — ${errorCount}/${N} shot(s) errored`;
-    aggregationKind = 'errored';
-  } else if (allPassed) {
-    passed = true;
-    verdict = `PASS — ${N}/${N} shots unanimous`;
-    aggregationKind = 'unanimous_pass';
-  } else if (passCount === 0) {
-    passed = false;
-    verdict = `FAIL — ${failCount}/${N} shots`;
-    aggregationKind = 'unanimous_fail';
-  } else {
-    passed = false;
-    verdict = `FAIL — ${passCount} pass, ${failCount} fail (no unanimous pass)`;
-    aggregationKind = 'disagreement';
-  }
-
-  const lines = [];
-  lines.push(`Subagent grader ${VERSION} (multi-shot, ${N}× shots) — ${verdict}`);
-  lines.push(`  functional_fidelity (median): ${ff ?? '?'}/10`);
-  lines.push(`  code_quality (median):        ${cq ?? '?'}/10`);
-  if (economy != null) {
-    lines.push(`  economy (advisory, median):   ${economy}/10`);
-  }
-  lines.push(`  blended (median):             ${blended}/10  (threshold ${THRESHOLD}, both axes must clear on EVERY shot)`);
-  lines.push('');
-  lines.push('Per-shot:');
-  for (let i = 0; i < verdicts.length; i++) {
-    const v = verdicts[i];
-    const status = v.error ? 'ERROR' : v.passed === true ? 'PASS' : 'FAIL';
-    const ffStr = v.ff != null ? `${v.ff}` : '?';
-    const cqStr = v.cq != null ? `${v.cq}` : '?';
-    const confStr = v.confidence != null ? `  conf=${v.confidence}` : '';
-    const errStr = v.error ? `  (${v.error})` : '';
-    lines.push(`  shot[${i}] ${status}  ff=${ffStr}  cq=${cqStr}${confStr}  ${v.model ?? '?'}${errStr}`);
-  }
-  if (issues.length > 0) {
-    lines.push('');
-    lines.push(`Issues (${issues.length} unique across shots):`);
-    for (const it of issues.slice(0, 20)) {
-      const sev = String(it.severity || '?').padEnd(7);
-      const file = String(it.file || '?');
-      const summary = String(it.summary || '').replace(/\s+/g, ' ').trim();
-      lines.push(`  [${sev}] ${file} — ${summary}`);
-    }
-    if (issues.length > 20) {
-      lines.push(`  ... ${issues.length - 20} more (full set in signals.shots[].issues)`);
-    }
-  }
-
-  // #486 (ADR 0028 path 1): sum each shot's own subprocess cost (envelope
-  // total_cost_usd) so the trivial-diff path reports panel_cost_usd the same way
-  // the full panel does. The single-shot trivial path is the common case here.
-  const panelCostUsd = shotResults.reduce((s, r) => s + (Number(r.cost_usd) || 0), 0);
-
-  return {
-    score: blended,
-    passed,
-    threshold: THRESHOLD,
-    grader_kind: 'subagent-multishot',
-    grader_version: VERSION,
-    panel_cost_usd: panelCostUsd, // #486: panel's own API spend (USD)
-    rubric_scores: {
-      functional_fidelity: ff,
-      code_quality: cq,
-      economy,
-      code_quality_effective: cqEffective,
-    },
-    signals: {
-      grader_kind: 'subagent-multishot',
-      grader_version: VERSION,
-      threshold: THRESHOLD,
-      // Top-level confidence = median across shots. ship.js forwards this to
-      // POST /tasks/:id/grade so the server-side re-validation has a real
-      // number rather than null (see Number(null)=0 fix in validateAndScore).
-      confidence: confidenceMedian,
-      shots: verdicts,
-      shot_count: N,
-      pass_count: passCount,
-      fail_count: failCount,
-      error_count: errorCount,
-      shot_aggregation: aggregationKind,
-      ...meta,
-    },
-    notes_md: lines.join('\n'),
-    issues,
-  };
-}
-
 // ---------- v3: subagent subprocess invoker ----------
 
 // Unchanged from v2 — spawns `claude -p --print --model <model> --output-
@@ -786,7 +619,6 @@ function aggregateShots(shotResults, meta = {}) {
 // via opts.model (set by gradeSubagent below).
 
 module.exports = {
-  aggregateShots,
   changeMagnitude,
   hasPanelEvidence,
   isSubstantiveIssue,

package/modules/grading/grader-subagent.js

@@ -11,10 +11,14 @@
 const path = require('node:path');
 const fs = require('node:fs');
 const { spawn } = require('node:child_process');
-const { DEFAULT_GRADER_MODEL, THRESHOLD, VERSION, api } = require('./grader-config.js');
-const { validateAndScore } = require('./grader-score.js');
+const { DEFAULT_GRADER_MODEL, api } = require('./grader-config.js');
+// task 1002840: the ModelProvider port (ADR 0089 §3) — runSubagent below routes
+// through it; the Claude spawn is its default provider.
+const modelProvider = require('./model-provider.js');
 
-function runSubagent({ prompt, opts = {} }) {
+// The `claude -p` spawn — the anthropic-cli ModelProvider. Body unchanged by
+// task 1002840; only its name moved, so runSubagent could become the port.
+function runClaudeCli({ prompt, opts = {} }) {
   return new Promise((resolve, reject) => {
     const timeoutMs = opts.timeoutMs ?? 240_000;
     const cwd = opts.cwd ?? process.cwd();
@@ -120,6 +124,17 @@ function runSubagent({ prompt, opts = {} }) {
     child.stdin.end();
   });
 }
+modelProvider.registerModelProvider({ name: modelProvider.DEFAULT_PROVIDER, kind: 'anthropic-cli', run: runClaudeCli });
+
+// runSubagent — the ModelProvider port every platform LLM call goes through
+// (the panel workers, the no-artifact shots, architect-audit, the agents module
+// via the grade port). A model with no registered provider prefix — every id in
+// use today — runs on the Claude spawn exactly as before; `<provider>:<model>`
+// runs on that provider, which answers in the same envelope + error codes.
+function runSubagent({ prompt, opts = {} }) {
+  const ref = modelProvider.parseModelRef(opts.model ?? DEFAULT_GRADER_MODEL);
+  return modelProvider.resolveModelProvider(ref.provider).run({ prompt, opts: { ...opts, model: ref.model } });
+}
 
 // ---------- task 1002660: shared context bundle ----------
 //
@@ -261,54 +276,6 @@ function extractAssistantResult(rawStdout) {
   }
 }
 
-// Run one shot end-to-end: build the prompt (caller did this), spawn the
-// subagent at the requested model, parse + validate. Returns a per-shot
-// grade-result; subprocess failures become passed=false per-shot results so
-// aggregateShots can still produce a final verdict.
-async function runOneShot({ prompt, model, shotIndex, runOpts }) {
-  try {
-    const { stdout, run_ms, exit_code, cost_usd } = await runSubagent({
-      prompt,
-      opts: { ...runOpts, model },
-    });
-    const assistantResult = extractAssistantResult(stdout);
-    const scored = validateAndScore(assistantResult, {
-      run_ms,
-      exit_code,
-      raw_chars: stdout.length,
-      model,
-      shot_index: shotIndex,
-    });
-    // #486 (ADR 0028 path 1): carry this shot's own subprocess cost so the
-    // trivial-diff path (gradeSubagentV3MultiShot → aggregateShots) can sum it
-    // into panel_cost_usd just like the full panel does.
-    scored.cost_usd = cost_usd ?? null;
-    return scored;
-  } catch (err) {
-    return {
-      score: 0,
-      passed: false,
-      threshold: THRESHOLD,
-      grader_kind: 'subagent',
-      grader_version: VERSION,
-      rubric_scores: {},
-      signals: {
-        grader_kind: 'subagent',
-        grader_version: VERSION,
-        model,
-        threshold: THRESHOLD,
-        shot_index: shotIndex,
-        run_ms: err.run_ms ?? null,
-        exit_code: err.exit_code ?? null,
-        subprocess_error: err.code || 'UNKNOWN',
-        stderr_tail: (err.stderr || '').slice(-500),
-      },
-      notes_md: `Shot ${shotIndex}: subprocess failed (${err.code || 'UNKNOWN'}: ${err.message}).`,
-      issues: [],
-    };
-  }
-}
-
 // ---------- v4: panel aggregation (MAS Phase 2.0) ----------
 
 // Collapses N per-worker WorkerResults into the final v4 grade-result.
@@ -342,7 +309,6 @@ module.exports = {
   diffAddedLinesByFile,
   extractAssistantResult,
   extractCostUsd,
-  runOneShot,
   runSubagent,
   runSubagentCached,
 };

package/modules/grading/grader.js

@@ -75,23 +75,28 @@
 //   module (which receives it via dependency injection to avoid a circular
 //   require).
 //
-// ## Legacy v1 (still exported)
+// ## Retired (task 1002842)
 //
-//   gradeDeterministic(...) — v1 deterministic path, kept for
-//   scripts/gds/audit-ships.js (retroactive audit). NOT invoked at ship time.
+//   The v1 deterministic process-audit grader (gradeDeterministic and its nine
+//   per-criterion scorers) and its one caller, scripts/gds/audit-ships.js, are
+//   deleted: the live panel grades every ship at ship time, so a retroactive
+//   heuristic audit has nothing left to cover. aggregateShots / runOneShot moved
+//   to grader-no-artifact.js, the species that depends on them; they are still
+//   re-exported here, and the trivial-diff fast path below still runs one shot.
 
 const path = require('node:path');
 const fs = require('node:fs');
 const { spawn } = require('node:child_process');
 
-const { ASSUMED_BUILDER_MODEL, DEFAULT_GRADER_MODEL, DEFAULT_SHOTS, FALLTHROUGH_CONFIDENCE, RUBRIC, THRESHOLD, V1, V3, VERSION, generatedProvenance, permissionPathPrePass, routeRankPrePass, workers } = require('./grader-config.js');
+const { ASSUMED_BUILDER_MODEL, DEFAULT_GRADER_MODEL, FALLTHROUGH_CONFIDENCE, RUBRIC, THRESHOLD, VERSION, generatedProvenance, permissionPathPrePass, routeRankPrePass, workers } = require('./grader-config.js');
 const { clamp, isTrivialDiff, median, r1, selectGraderModel, workerModelFor } = require('./grader-model.js');
 const { buildPrompt, extractJson } = require('./grader-prompt.js');
 // task 1004011: the second grade species — a grade over the CLAIM on an
 // --allow-empty ship, where there is no diff for the panel above to read.
-const { GRADER_KIND: NO_ARTIFACT_GRADER_KIND, buildNoArtifactPrompt, gradeNoArtifact } = require('./grader-no-artifact.js');
-const { aggregateShots, changeMagnitude, hasPanelEvidence, isSubstantiveIssue, serverSideGuards, validateAndScore } = require('./grader-score.js');
-const { buildContextBundle, diffAddedLinesByFile, extractAssistantResult, extractCostUsd, runOneShot, runSubagent, runSubagentCached } = require('./grader-subagent.js');
+// task 1002842: the generalist-shot plumbing now lives there too.
+const { GRADER_KIND: NO_ARTIFACT_GRADER_KIND, aggregateShots, buildNoArtifactPrompt, gradeNoArtifact, runOneShot } = require('./grader-no-artifact.js');
+const { changeMagnitude, hasPanelEvidence, isSubstantiveIssue, serverSideGuards, validateAndScore } = require('./grader-score.js');
+const { buildContextBundle, diffAddedLinesByFile, extractAssistantResult, extractCostUsd, runSubagent, runSubagentCached } = require('./grader-subagent.js');
 // task 1002903: the severity gate's unbacked low scores, named once for the record.
 const { unbackedLowScores } = require('./grader-workers/worker-verdict.js');
 // task 1002839: the central engine the code panel below registers into (ADR 0089 §1).
@@ -478,8 +483,8 @@ function aggregatePanel(workerResults, meta = {}) {
     grader_version: VERSION,
     // #486: the panel's own API spend (USD), for cost_log (source=grader).
     panel_cost_usd: panelCostUsd,
-    // rubric_scores keep the v3 keys so the trend dashboard + audit-ships
-    // legacy v1 entry-points continue to read them. Phase 2.1 added
+    // rubric_scores keep the v3 keys so the trend dashboard and the server's
+    // validateAndScore re-validation continue to read them. Phase 2.1 added
     // security_score; Phase 2.2 adds efficiency_score. Both additive —
     // consumers that don't know about them just ignore the fields.
     rubric_scores: {
@@ -651,12 +656,17 @@ async function gradeSubagent({ task, valueSummary, diff, changedFiles, diffStats
   });
 }
 
-// ---------- v3 multi-shot path (kept exported for explicit diagnostics) ----------
+// ---------- the trivial-diff fast path (#513) ----------
 //
-// Same shape as the original v3 gradeSubagent; renamed so the public name
-// (gradeSubagent) points at the v4 panel. Useful for re-running a single
-// generalist prompt N times for diagnostic comparison against the panel.
-async function gradeSubagentV3MultiShot({ task, valueSummary, diff, changedFiles, diffStats, shots = DEFAULT_SHOTS, builderModel, runOpts = {} }) {
+// A small copy/docs-only diff gets ONE generalist shot instead of the four-worker
+// panel — "lighter, not skipped" (ADR 0089 §7 keeps it). The name is historical:
+// the v3 multi-shot GENERATION is retired (task 1002842 removed its 3-shot default
+// and rubric block), and this is what is left of it — the working fast path.
+// Moving it onto a v4 panel needs a reviewer type that reports both
+// functional_fidelity and code_quality, which no v4 worker does; that is a cost
+// and behaviour decision, not a cleanup. `runShot` is a test seam defaulted to
+// the real spawner (the same pattern as gradeNoArtifact); production passes none.
+async function gradeSubagentV3MultiShot({ task, valueSummary, diff, changedFiles, diffStats, shots = 1, builderModel, runOpts = {}, runShot = runOneShot }) {
   const prompt = buildPrompt({ task, valueSummary, diff, changedFiles, diffStats, generatedProvenance: await generatedProvenance.verifyGeneratedProvenanceAsync({ changedFiles, cwd: runOpts.cwd }) }); // task 1002795
   const graderModel = selectGraderModel({ builderModel });
   const N = Math.max(1, Math.trunc(shots));
@@ -664,7 +674,7 @@ async function gradeSubagentV3MultiShot({ task, valueSummary, diff, changedFiles
   const startedAt = Date.now();
   const shotPromises = [];
   for (let i = 0; i < N; i++) {
-    shotPromises.push(runOneShot({ prompt, model: graderModel, shotIndex: i, runOpts }));
+    shotPromises.push(runShot({ prompt, model: graderModel, shotIndex: i, runOpts }));
   }
   const shotResults = await Promise.all(shotPromises);
   const totalMs = Date.now() - startedAt;
@@ -676,189 +686,6 @@ async function gradeSubagentV3MultiShot({ task, valueSummary, diff, changedFiles
   });
 }
 
-// =========================================================================
-// Legacy v1 (deterministic) — kept for scripts/gds/audit-ships.js only.
-// =========================================================================
-
-function scoreEstActualBand(actualMinutes, estMinutes) {
-  if (!Number.isFinite(actualMinutes) || actualMinutes <= 0) return 5;
-  if (!Number.isFinite(estMinutes) || estMinutes <= 0) return 6;
-  const ratio = actualMinutes / estMinutes;
-  if (ratio >= 0.5 && ratio <= 2.0) return 10;
-  if (ratio < 0.5) {
-    if (ratio <= 0.1) return 0;
-    return r1(10 * ((ratio - 0.1) / 0.4));
-  }
-  if (ratio > 2.0) {
-    if (ratio >= 10) return 0;
-    return r1(10 * (1 - (ratio - 2.0) / 8.0));
-  }
-  return 5;
-}
-
-function scoreValueSummary(valueSummary) {
-  const s = (valueSummary || '').trim();
-  if (s.length === 0) return 0;
-  if (s.length < 40) return 5;
-  if (s.length < 120) return 8;
-  return /\b(add|fix|update|remove|refactor|wire|land|migrate|enable|disable|introduce|deprecate|ship|deploy|prevent|capture|gate|allocate|carry|expose|surface)\b/i.test(s)
-    ? 10
-    : 8;
-}
-
-function scoreDiffHygiene(diffStats) {
-  if (!diffStats) return 5;
-  const files = Number(diffStats.files_changed) || 0;
-  const added = Number(diffStats.lines_added) || 0;
-  const removed = Number(diffStats.lines_removed) || 0;
-  if (files === 0 || added + removed === 0) return 0;
-  if (files > 50) return 5;
-  return 10;
-}
-
-function scoreScannerSmoke(verification) {
-  if (!verification) return 5;
-  return verification.scanner_clean && verification.smoke_clean ? 10 : 5;
-}
-
-function scoreScopingSignal(scopingConfidence) {
-  if (scopingConfidence == null) return 6;
-  const c = Number(scopingConfidence);
-  if (c <= 0) return 2;
-  if (c === 1) return 4;
-  if (c === 2) return 6;
-  if (c === 3) return 7;
-  if (c === 4) return 8;
-  return 10;
-}
-
-function scoreLocNetNegative(diffStats) {
-  if (!diffStats) return 5;
-  const added = Number(diffStats.lines_added) || 0;
-  const removed = Number(diffStats.lines_removed) || 0;
-  if (removed > added) return 10;
-  if (removed === added) return 5;
-  return 0;
-}
-
-function scoreLearningRow(learningRow) {
-  if (!learningRow) return 0;
-  const body = (learningRow.body_md || '').trim();
-  if (body.length > 100) return 10;
-  if (body.length > 0) return 5;
-  return 0;
-}
-
-function scoreArtRubricAvg(value) {
-  if (!Number.isFinite(value)) return null;
-  return clamp(r1(value), 0, 10);
-}
-
-function blendWeightsForKind(kind) {
-  const blended = {};
-  for (const [name, def] of Object.entries(V1.weights.common)) {
-    blended[name] = def.weight;
-  }
-  const kindBonus = V1.weights.kind_specific[kind];
-  if (kindBonus) {
-    for (const [name, def] of Object.entries(kindBonus)) {
-      if (name === 'loc_net_negative' && 'diff_hygiene' in blended) delete blended.diff_hygiene;
-      if (name === 'learning_row' && 'diff_hygiene' in blended) delete blended.diff_hygiene;
-      blended[name] = def.weight;
-    }
-  }
-  const total = Object.values(blended).reduce((a, b) => a + b, 0);
-  for (const k of Object.keys(blended)) blended[k] = blended[k] / total;
-  return blended;
-}
-
-// Legacy v1 entry. Renamed from `grade()` in V3.R18 (#209). Now ONLY
-// called from scripts/gds/audit-ships.js (retroactive #204 audit). The
-// runtime ship path uses gradeSubagent (v3 multi-shot).
-function gradeDeterministic({ task, verification, actualMinutes, valueSummary, learningRow }) {
-  const kind = task?.kind || 'unclassified';
-  const weights = blendWeightsForKind(kind);
-  const rubricScores = {};
-  const signals = {
-    kind,
-    est_minutes: task?.est_minutes ?? null,
-    actual_minutes: actualMinutes ?? null,
-    scoping_confidence: task?.scoping_confidence ?? null,
-    scanner_clean: verification?.scanner_clean ?? null,
-    smoke_clean: verification?.smoke_clean ?? null,
-    diff_stats: verification?.diff_stats ?? null,
-    art_rubric_avg: verification?.art_rubric_avg ?? null,
-    value_summary_chars: (valueSummary || '').trim().length,
-  };
-
-  if ('est_actual_band' in weights) {
-    rubricScores.est_actual_band = scoreEstActualBand(actualMinutes, task?.est_minutes);
-  }
-  if ('value_summary' in weights) {
-    rubricScores.value_summary = scoreValueSummary(valueSummary);
-  }
-  if ('diff_hygiene' in weights) {
-    rubricScores.diff_hygiene = scoreDiffHygiene(verification?.diff_stats);
-  }
-  if ('scanner_smoke' in weights) {
-    rubricScores.scanner_smoke = scoreScannerSmoke(verification);
-  }
-  if ('scoping_signal' in weights) {
-    rubricScores.scoping_signal = scoreScopingSignal(task?.scoping_confidence);
-  }
-  if ('loc_net_negative' in weights) {
-    rubricScores.loc_net_negative = scoreLocNetNegative(verification?.diff_stats);
-  }
-  if ('learning_row' in weights) {
-    rubricScores.learning_row = scoreLearningRow(learningRow);
-  }
-  if ('art_rubric_avg' in weights) {
-    const s = scoreArtRubricAvg(verification?.art_rubric_avg);
-    if (s == null) {
-      const w = weights.art_rubric_avg;
-      delete weights.art_rubric_avg;
-      const remaining = Object.values(weights).reduce((a, b) => a + b, 0);
-      if (remaining > 0) {
-        for (const k of Object.keys(weights)) weights[k] = weights[k] / remaining;
-      }
-    } else {
-      rubricScores.art_rubric_avg = s;
-    }
-  }
-
-  let score = 0;
-  for (const [name, w] of Object.entries(weights)) {
-    if (rubricScores[name] != null) score += rubricScores[name] * w;
-  }
-  score = r1(clamp(score, 0, 10));
-
-  const threshold = V1.threshold;
-  const passed = score >= threshold;
-
-  const lines = [];
-  lines.push(`Auto-grader ${V1.version} (deprecated) — task_kind=${kind}.`);
-  lines.push(`Score ${score}/10 vs threshold ${threshold} → ${passed ? 'PASS' : 'FAIL'}.`);
-  lines.push('');
-  lines.push('Per-criterion:');
-  for (const [name, w] of Object.entries(weights)) {
-    const s = rubricScores[name];
-    if (s == null) continue;
-    const contrib = r1(s * w);
-    lines.push(`  - ${name.padEnd(20)} ${String(s).padStart(4)}/10  × ${w.toFixed(2)} = ${contrib}`);
-  }
-
-  return {
-    score,
-    passed,
-    threshold,
-    grader_kind: 'auto',
-    grader_version: V1.version,
-    rubric_scores: rubricScores,
-    signals,
-    notes_md: lines.join('\n'),
-  };
-}
-
 module.exports = {
   // v4 surface (active panel-based gate)
   aggregatePanel,
@@ -888,25 +715,13 @@ module.exports = {
   runSubagent,
   runSubagentCached,              // System 5 (task 1359): opt-in content-addressed cached path
   runOneShot,
-  gradeSubagentV3MultiShot,       // explicit v3 multi-shot diagnostic path
+  gradeSubagentV3MultiShot,       // the trivial-diff fast path (#513) — one generalist shot
   extractJson,
   extractAssistantResult,
   extractCostUsd,                 // #486: parse total_cost_usd from the envelope
   THRESHOLD,
   VERSION,
   DEFAULT_GRADER_MODEL,
-  DEFAULT_SHOTS,
   ASSUMED_BUILDER_MODEL,
-  // legacy v1 (audit-ships.js only)
-  gradeDeterministic,
-  scoreEstActualBand,
-  scoreValueSummary,
-  scoreDiffHygiene,
-  scoreScannerSmoke,
-  scoreScopingSignal,
-  scoreLocNetNegative,
-  scoreLearningRow,
-  scoreArtRubricAvg,
-  blendWeightsForKind,
   RUBRIC,
 };