fix+docs: D33 — round-5 cleanup batch (F1/F3/F5/F8/F9/F10/F11/F12)

cold-audit catch from 2026-05-24 (round 5)

Round-5 cold-audit cleanup batch. 8 items + 1 release-discipline
reconciliation. Largest batch by line count (582+/39-) but every item
is small-and-focused. 3 P2 items (F1/F3/F5 of which F3 + F1 are real
correctness/observability fixes; F5 backfills /health to spec).

Changes (10 files, +583/-39):

**P2 fixes**

1. **F1 — ALIGNMENT.md mistral authority pin self-contradicted plugin**
   (ALIGNMENT.md): row cited `vibe --prompt --output json` but mistral.mjs
   uses `--output streaming` (the plugin header at lines 360-369 even
   justifies WHY: `--output json` emits single blob, breaks NDJSON
   line-buffered parser). Constitution self-contradicting itself —
   missed across 4 prior rounds. Pin updated to `--output streaming`
   with DOCS-1 reference.

2. **F3 — Deterministic function_call synth ID**
   (lib/ir/openai-to-ir.mjs): deprecated `function_call` translation
   produced `id: \`fc-${Date.now()}\`` → ID flows into normalized
   tool_calls → cache key SHA-256. Two identical requests separated
   by ≥1ms → different cache keys → cache always misses for
   `function_call` request shape. Violates ADR 0005 invariant
   "same inputs → same key, no random, no timestamp."

   Fixed: id is now `fc-<16-hex>` from SHA-256 of `${name}\0${arguments}`.
   NUL separator prevents the (name='ab',args='c') vs (name='a',args='bc')
   collision. 2^64 collision resistance is more than sufficient for
   tool_call ID disambiguation (per-request semantic key, not crypto
   primitive).

**P3 fixes**

3. **F5 — /health invokes per-plugin healthCheck()** (server.mjs +
   docs/openai-spec-pin.md): ADR 0002 says "healthCheck — startup AND
   /health endpoint use this." Pre-D33 /health returned only
   {enabled, available} counts. Now async, iterates loadedProviders,
   awaits each plugin's healthCheck() in try/catch. Returns
   `providers: {enabled, available, status: {<name>: {ok, latencyMs?, error?}}}`.

4. **F8 — X-OLP-Cache reports fallback-hop cache hits** (server.mjs):
   pre-D33 cacheStatus computed from `preCheckHit && fallbackHops === 0`
   — only counted primary-hop cache hits. When fallback fires + the
   fallback hop's getOrCompute returns from cache, header reported
   `miss` despite no spawn happening.

   Fixed: peek BEFORE getOrCompute inside executeHopFn, set
   `lastHopWasCached` closure variable on every hop (last-write-wins
   = serving hop's state). cacheStatus combines
   `lastHopWasCached || (preCheckHit && fallbackHops === 0)`.

   F8 chose option (b) peek-then-getOrCompute over option (a)
   getOrCompute API change because option (a) would break ~15 test
   callsites for marginal benefit. Accepted race window same as
   existing preCheckHit pattern.

5. **F9 — validateProvider hints error message updated** (lib/providers/
   base.mjs): pre-D33 message listed cacheable as missing and
   maxSpawnTimeMs as required. Now: `'hints must be an object with
   { requiresTTY, concurrentSpawnSafe, maxConcurrent } + optional
   { maxSpawnTimeMs, cacheable }'`.

6. **F10 — Dead cache-write branch removed** (server.mjs): the
   `if (hasStopChunk)` check in the streaming stop-less exhaustion
   branch was unreachable (the stop-chunk completion path returns
   earlier inside the for-await loop). Removed the dead code + added
   a comment documenting the invariant.

**Governance/policy**

7. **F11 — Phase rolling mode policy formalized** (CLAUDE.md +
   CHANGELOG.md): 22+ D-day commits accumulated under "Unreleased"
   without per-D version bumps — Iron Rule 5 (release-kit bump-before-
   push) appeared to be silently violated. Reality: per-D bumps would
   produce 30+ noise tags during Phase 1. F11 formalizes the policy:
   intra-Phase D-day commits accumulate under Unreleased; bump+tag
   fires explicitly at Phase close (maintainer-triggered, not
   automated). CLAUDE.md release_kit overlay gains `phase_rolling_mode`
   block documenting the exception with self-pointer ("if Rule 5
   appears silently violated, check this section first"). CHANGELOG
   "Unreleased" gets a notice at top.

   **No version bump, no git tag in D33** — policy formalization only.

8. **F12 — /v1/models created is stable per-model timestamp**
   (models-registry.json + lib/providers/index.mjs + server.mjs +
   docs/openai-spec-pin.md): pre-D33 used Math.floor(Date.now()/1000)
   per request — violates OpenAI spec which treats `created` as
   per-model attribute. Clients caching models by created would see
   spurious updates on every poll.

   Fixed: models-registry.json gains `bootstrapCreated: 1778630400`
   top-level constant + per-model `created` fields where known
   (anthropic claude-{opus,sonnet,haiku} with estimated release dates;
   devstral models from "25-12" suffix; codex models pinned to
   bootstrap pending verified release dates). handleModels uses
   `getModelCreated(modelId)` helper from lib/providers/index.mjs.
   Aliases share canonical's timestamp.

**Tests** (test-features.mjs): 401 → 414 (+13):
- F3 ×3 (same input → same id → same cache key; different name → different)
- F5 ×4 (empty/single/multi/throwing-plugin /health shapes)
- F8 ×1 (2-hop primary-fail + secondary-cache-hit → X-OLP-Cache: hit)
- F12 ×5 (stability/fallback/alias-equals-canonical)

Pre-commit fold-in (per evidence-first checkpoint #4):
- **D33 reviewer flagged F3 empty-args asymmetry** (Concern #1): hash
  input used `?? ''` (empty stays) but emitted IR field used
  `|| '{}'` (empty becomes '{}'). Consequence: `arguments: ''` and
  `arguments: '{}'` emit identical IR but compute different ids →
  different cache keys for semantically-identical requests. The exact
  cache-stability bug F3 was supposed to fix.

  Folded in: canonicalize empty-args to '{}' BEFORE hashing. Hash
  input now matches IR emission exactly. Same line change resolves
  the asymmetry.

Authority:
- ALIGNMENT.md self-amendment (F1 pin correction)
- ADR 0005 invariant "same inputs → same key, no random, no timestamp"
  (F3 restoration)
- ADR 0002 § Provider contract "/health uses healthCheck" (F5)
- ADR 0004 § Observability headers (F8 X-OLP-Cache correctness)
- ADR 0005 § Cache write conditions item 1 (F10 truncation-not-cached
  invariant explicit)
- Iron Rule 5 (F11 release-kit reconciliation)
- OpenAI /v1/models spec — `created` per-model stable (F12)
- CC 开发铁律 v1.6 § 10.x — Round-5 Cold Audit caught all 8

Reviewer (Iron Rule v1.6 § 10.x Mode A, fresh-context opus, independent
of drafter): APPROVE_WITH_MINOR. Verified:
- F1 plugin cross-reference (mistral.mjs:360-369) accurately documents
  the rationale
- F3 collision resistance + NUL separator + restored cache invariant
- F5 all 4 cases (empty/single/multi/throwing) work
- F8 closure semantics across multi-hop chains (verified hop-fail +
  fallback-hit case)
- F10 dead code removal preserves the stop-chunk completion path
- F11 phase_rolling_mode policy honest about what happened and what
  the going-forward rule is
- F12 stability across consecutive /v1/models calls; alias-canonical
  parity
- 414/414 tests pass

3 remaining non-blocking suggestions (F3-vs-modern-tool_calls path
canonicalization symmetry; F12 codex models explicit-vs-fallback
writeup mismatch; F8 servingHopWasCached naming) tracked as future
polish; not folded.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

lib/ir/openai-to-ir.mjs

@@ -10,6 +10,7 @@
  * they receive only the IR (ADR 0003 § "IR is not exposed externally").
  */
 
+import { createHash } from 'node:crypto';
 import { IR_VERSION, validateIRRequest } from './types.mjs';
 
 // ── Custom error ──────────────────────────────────────────────────────────
@@ -74,13 +75,27 @@ function translateMessage(msg) {
       },
     }));
   } else if (msg.function_call && typeof msg.function_call === 'object') {
-    // Deprecated OpenAI function_call field — map to single tool_calls entry
+    // Deprecated OpenAI function_call field — map to single tool_calls entry.
+    // ID is deterministic: sha256(name + NUL + arguments).slice(0,16) so that
+    // two identical function_call requests produce the same IR → same cache key.
+    // Using Date.now() here would violate ADR 0005 invariant "same inputs → same key,
+    // no random, no timestamp" (F3 cold-audit finding, round-5).
+    const fcName = msg.function_call.name ?? '';
+    // Canonicalize empty-args to '{}' BEFORE hashing so the hash input matches the
+    // emitted IR field exactly. Otherwise `arguments: ''` and `arguments: '{}'` would
+    // emit identical IR but compute different ids → different cache keys for
+    // semantically-identical requests. Folded in per D33 F3 reviewer non-blocking #1.
+    const fcArgs = msg.function_call.arguments || '{}';
+    const fcKey = createHash('sha256')
+      .update(`${fcName}\0${fcArgs}`)
+      .digest('hex')
+      .slice(0, 16);
     irMsg.tool_calls = [{
-      id: `fc-${Date.now()}`,
+      id: `fc-${fcKey}`,   // deterministic — same input → same id
       type: 'function',
       function: {
-        name: msg.function_call.name ?? '',
-        arguments: msg.function_call.arguments ?? '',
+        name: fcName,
+        arguments: fcArgs,
       },
     }];
   }

lib/providers/base.mjs

@@ -114,7 +114,7 @@ export function validateProvider(p) {
   }
 
   if (!p.hints || typeof p.hints !== 'object') {
-    errors.push('hints must be an object with { requiresTTY, concurrentSpawnSafe, maxConcurrent, maxSpawnTimeMs }');
+    errors.push('hints must be an object with { requiresTTY, concurrentSpawnSafe, maxConcurrent } + optional { maxSpawnTimeMs, cacheable }');
   } else {
     if (typeof p.hints.requiresTTY !== 'boolean') errors.push('hints.requiresTTY must be a boolean');
     if (typeof p.hints.concurrentSpawnSafe !== 'boolean') errors.push('hints.concurrentSpawnSafe must be a boolean');

lib/providers/index.mjs

@@ -80,6 +80,44 @@ export function getAliasMap() {
   return new Map(_aliasMap);
 }
 
+// ── Per-model stable created timestamps ───────────────────────────────────
+// F12 (round-5 cold-audit): OpenAI spec treats `created` as a stable per-model
+// attribute. Synthesizing Date.now() on each /v1/models request causes spurious
+// updates for clients caching models by `created`. This map provides the stable
+// per-model timestamp sourced from models-registry.json entries.
+//
+// Fallback: models-registry.json top-level `bootstrapCreated` is used when a
+// model entry has no `created` field.
+
+/** @type {number} Stable fallback timestamp for models with no per-entry `created` field. */
+export const REGISTRY_BOOTSTRAP_CREATED = modelsRegistryRaw?.bootstrapCreated ?? 1778630400;
+
+/** Map<modelId, createdUnixSeconds> — built at module load, never mutated. */
+const _modelCreatedMap = new Map();
+for (const providerEntry of Object.values(modelsRegistryRaw?.providers ?? {})) {
+  for (const modelEntry of providerEntry?.models ?? []) {
+    if (modelEntry?.id && typeof modelEntry.created === 'number') {
+      _modelCreatedMap.set(modelEntry.id, modelEntry.created);
+    }
+  }
+}
+
+/**
+ * Returns the stable Unix-epoch `created` timestamp for a given model ID.
+ * Prefers the per-entry value from models-registry.json; falls back to
+ * REGISTRY_BOOTSTRAP_CREATED when the entry has no `created` field.
+ *
+ * Per F12 (round-5 cold-audit): DO NOT use Date.now() for the `created` field
+ * in /v1/models responses — OpenAI clients may cache models by `created` and
+ * would see spurious updates on every poll if the timestamp varies per request.
+ *
+ * @param {string} modelId
+ * @returns {number} Unix timestamp in seconds
+ */
+export function getModelCreated(modelId) {
+  return _modelCreatedMap.get(modelId) ?? REGISTRY_BOOTSTRAP_CREATED;
+}
+
 // ── Registry functions ────────────────────────────────────────────────────
 
 /**