chore: D25 — round-2 P3 batch (F5/F6/F7/F9/F10/F11/F13 + D22 follow-up)

cold-audit catch from 2026-05-24

Final round-2 cold-audit cleanup batch. 8 small P3 items, batched per
Iron Rule 11 IDR cleanup-batch convention (precedent: D19 / D20). No
item exceeds ~35 lines; only F9 is a code change, the rest are docs +
registry updates.

Changes (7 files, +140 / -10):

1. ALIGNMENT.md (+8 / -2)
   - **F7**: Authority pin rows for anthropic / codex / mistral updated
     from "TBD on Phase-1 spawn" to the actual citations cited in each
     plugin's header:
     · anthropic — @anthropic-ai/claude-code v2.1.89 (OCP fork audit pin)
     · openai — https://developers.openai.com/codex/cli/reference + features URL
     · mistral — https://docs.mistral.ai/mistral-vibe/terminal/quickstart + config URL
     Plugins remain Candidate (per Provider Inventory) — D25 just removes the
     `TBD` marker; Enabled transition still requires Phase audit.
   - **F6**: New 3rd entry in § One-shot Triggered Audits — "OpenAI Codex ToS
     formal pin (trigger: Phase 2 E2E enable for Codex, OR 2026-12-31)".
     Closes the cross-reference from ADR 0006 § Decision table.

2. README.md (+4 / -2)
   - **F5**: Cache-key bullet (Architecture section) replaced the stale
     7-tuple with a link to ADR 0005 § Cache key composition + the
     post-D15 11-field tuple inlined.
   - **D22 follow-up**: "328-test suite" → "Comprehensive test suite
     covering IR, cache, fallback, and integration paths" (version-less
     to prevent re-drift on every D-day).

3. docs/adr/0002-plugin-architecture.md (+4 / -2) — **F11**: Amendment 1
   wording corrected. The original Authority line + maxSpawnTimeMs
   description said "fallback engine's spawn-timeout enforcement loop"
   — but the enforcement actually lives in each provider plugin's
   `_spawnAndStream` (setTimeout + proc.kill + reject pattern). The
   fallback engine merely treats SPAWN_TIMEOUT as a hard trigger per
   ADR 0004 § Trigger taxonomy bullet 4. Both wording sites updated.

4. docs/adr/0005-cache-cross-provider.md (+1) — **F13**: Amendment 2
   gains a "Note on null-coalescing collisions" paragraph documenting
   that the `?? null` serialization treats undefined / null / [] as
   equivalent cache keys for array-typed fields. Intentional — both
   `tools: []` and `tools` omitted semantically mean "no tools." If a
   future provider distinguishes empty-array vs absent, the serialization
   needs revision.

5. models-registry.json (+35) — **F10**: 5 candidate entries added for
   the providers ALIGNMENT.md names but registry omitted. All five with
   `candidate: true`, `models: []`, tier per ALIGNMENT.md inventory:
   · grok / kimi → Tier C
   · minimax / glm / qwen → Tier B
   Closes the release_kit overlay's "Supported Providers from
   models-registry.json" claim. alignment.yml KNOWN_PROVIDERS validation
   array already includes all 8 names, so registry → workflow validation
   continues to pass.

6. server.mjs (+20 / -6) — **F9**: Streaming success path now distinguishes
   stop-terminated vs exhausted-without-stop. Pre-D25 code unconditionally
   cached after the for-await loop ended, treating any exhaustion as
   "completed." Now: only cache if `lastChunk?.type === 'stop'`. If the
   generator exhausts without emitting stop (truncation), log
   `streaming_no_stop_chunk` warn event and do NOT persist. Mirrors
   D16's buffered-path semantics ("response completed successfully (no
   truncation, no error mid-stream)") with the simpler skip-write
   pattern (streaming path doesn't use getOrCompute → no singleflight
   eviction needed). D23's cacheableForFirstHop guard preserved as the
   outer condition.

7. test-features.mjs (+78) — F9 test 15e in Suite 15:
   - Mock provider whose spawn yields delta chunks then implicit-returns
     without stop (provider-injection pattern same as 15d — the real
     plugins synthesize a stop on clean proc exit so __setSpawnImpl can't
     simulate this case)
   - Asserts response succeeds AND second identical request triggers
     fresh spawn (proves no caching happened) AND X-OLP-Cache: miss on
     both responses

Tests: 348 → 349 (+1 from 15e). All pass on Node 20.

Authority:
- F5 → ADR 0005 § Cache key composition (post-D15 Amendment 2)
- F6 → ALIGNMENT.md self + ADR 0006 self-reference
- F7 → plugin headers (verified during D25 implementation)
- F9 → ADR 0005 § Cache write conditions item 1 + D16 truncation precedent
- F10 → ALIGNMENT.md § Provider Inventory tier classification
- F11 → ADR 0004 § Trigger taxonomy bullet 4 (the actual SPAWN_TIMEOUT
  hard-trigger documentation)
- F13 → ADR 0005 Amendment 2 (extends with the null-coalescing note)
- D22 fu → no spec authority; version-less framing prevents future drift
- CC 开发铁律 v1.6 § 10.x — Round-2 Cold Audit caught all 8 items

Reviewer (Iron Rule v1.6 § 10.x Mode A, fresh-context opus, independent
of drafter): APPROVE. Independently verified F7 citations against
plugin headers (anthropic.mjs:5-6, codex.mjs:23/31, mistral.mjs:26/32);
F10 tier classifications against ALIGNMENT.md § Provider Inventory;
F6 cross-reference now self-consistent (ADR 0006 → ALIGNMENT.md);
alignment.yml workflow validation passes; F9 truncation semantics
mirror D16 buffered-path; test 15e correctly uses provider-injection
since real plugin synthesizes stop on clean exit.

Three non-blocking suggestions noted (README cache-key bullet slightly
verbose with both link + inline list; F9 could optionally synthesize
a finish_reason: 'length' stop chunk for client-visible truncation
observability; test 15e single-delta variant could be extended to
multi-delta). None folded in — all genuine polish, not correctness
gaps.

---

**Round-2 cold-audit cleanup complete.** 5 D-days (D21-D25 minus the
already-completed D24) closed all 13 round-2 findings (P2: F1/F2/F3/F4
in D21/D22/D23/D24; P3: F5-F13 distributed across D25 + earlier issues
#2/#3). v0.1 is one cold-audit-round-3 away from being ready to tag.

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

docs/adr/0005-cache-cross-provider.md

@@ -23,6 +23,7 @@
 - **Change:** Expand cache key composition to include `max_tokens`, `top_p`, `stop`, and `tool_choice` in addition to the existing seven fields. The four new fields are appended after the existing set so that the ordering of existing fields is stable (pre-D15 cache entries are invalidated on first request — a forced miss — but the schema is forward-compatible). Field serialization follows the existing `?? null` null-coalescing pattern: a request with `max_tokens: undefined` serializes as `null` and is treated as equivalent to an absent field; `max_tokens: 100` serializes as `100`. This is consistent with how `temperature` and `response_format` are handled.
 - **Rationale:** ADR 0005's own stated invariant — "different output → different cache entry" — requires that any IR field affecting output be included in the cache key. The original v1.0 key composition was correct for the seven named fields but provided incomplete coverage of ADR 0003's IR field set. The four omitted fields are all defined in ADR 0003 § Optional fields and are all sourced directly from the OpenAI `/v1/chat/completions` spec parameters that affect generation.
 - **Forward-looking note:** Future IR field additions (e.g., `seed`, `frequency_penalty`, `presence_penalty`, `logit_bias`, `logprobs`, `n`) must be evaluated for cache-key inclusion at the time they are added to the IR. The default is to **include** the field in the cache key unless an explicit rationale documents why omitting it is safe (e.g., the field has no effect on the provider's output for any value, or it is a purely client-side metadata field). This evaluation must appear in the amending ADR per ALIGNMENT.md Rule 2(c) spirit.
+- **Note on null-coalescing collisions:** The `?? null` serialization pattern treats `field: undefined`, `field: null`, and `field: []` (for array fields such as `tools`) as equivalent cache keys. This is intentional — semantically these all mean "field absent or empty." In practice, `tools: []` (explicit empty array) and `tools` omitted (undefined) both mean "no tools available" and produce identical model output, so they correctly share a cache entry. If a future provider distinguishes empty-array from absent (e.g., `tools: []` explicitly meaning "tools disabled" vs `tools` omitted meaning "tools not specified"), the serialization for that field must be revisited and this amendment will need revision.
 - **Procedural mechanism:** CC 开发铁律 v1.6 § 10.x (Cold Audit caught it). This follows the calibration-loop precedent established by D11 / Amendment 1 of ADR 0002: the diff-review pass that approved the original ADR 0005 did not cross-reference all ADR 0003 optional fields; the cold-audit pass on 2026-05-23 caught the gap as Finding 7.
 
 ## Context