fix(openai): preserve Responses message phase across stateless replay (#2269)

[gold-silver-copper]

Summary

Second cassette-driven live bug hunt against the OpenAI, Anthropic, and Gemini providers (same doctrine as #2294): record new tests against the real APIs until a genuine defect surfaces, fix it, re-record what the fix invalidates.

Bug found and fixed

OpenAI Responses: the documented phase field was silently dropped from message output items and never replayed — the phase half of issue #2269.

• Symptom, proven by traffic already in the repo: every committed multi-turn gpt-5.6 cassette shows "phase":"final_answer" on the message items in each response and no phase in any follow-up request. OpenAI documents that follow-up requests should replay it and that dropping it can degrade performance.
• Root cause: OutputMessage had no phase field, so deserialization dropped it; nothing in the rig-message roundtrip could carry message-level wire data back to request assembly.
• Fix (responses_api/mod.rs):
  • OutputMessage gains phase: Option<String> — the raw path now round-trips it value-equal;
  • ingest parks it in the first text block's captured wire extras under openai_responses (the same vehicle annotations already use);
  • request assembly (assistant_text_replay_message) hoists it back to the message level on Message::Assistant, filtered out of the content-level extras so it serializes exactly where the wire wants it.
• Proof: tests/providers/openai/cassette/phase_roundtrip.rs + its cassette — a fresh two-turn recording in which turn 1's response carries phase and turn 2's recorded request body replays it on the assistant message ("id":"msg_…","phase":"final_answer","role":"assistant"), accepted by the live API. The harness matches outbound bodies, so a regression fails replay as a mock miss.
• Scope note: streaming ingestion does not yet route item extras into params (pre-existing, documented on OPENAI_RESPONSES_EXTRAS_KEY); this fix covers the blocking path, consistent with annotations. feat(openai): preserve Responses items for stateless replay #2269's other asks (compaction input items, caller-built raw requests) overlap open PR feat(openai): support server-side Responses compaction #2234 and are deliberately not touched.

Re-recorded

gpt_5_6_reasoning/five_turn_metadata_roundtrip.yaml — the one cassette the fix invalidates. The re-recording is a 4-line diff: the four follow-up request bodies now carry phase; responses unchanged.

New coverage recorded (independent of the fix)

• anthropic/stop_sequences — a stop_sequences hit (via additional_params) surfaces FinishReason::Stop, the raw response preserves stop_reason: "stop_sequence" + the matched stop_sequence: "five", and the emitted text stops before the sequence (claude-haiku-4-5).
• gemini/image_input — inline base64 image ingestion (inline_data request part → grounded description); the suite previously covered only the image generation direction.

Tracker sweep

Full sweep of all open issues/PRs was done earlier today and reported in #2294; the delta since is one new PR (#2293, Mistral — out of scope). This PR addresses the in-scope follow-up ranked highest there:

• feat(openai): preserve Responses items for stateless replay #2269 — phase half fixed here (Fixes intentionally not claimed: the issue also asks for compaction-item and raw-request support, which overlap open PR feat(openai): support server-side Responses compaction #2234).
• Remaining follow-ups unchanged from fix(openai): merge additional_params tools into chat completions tool list; record error-envelope coverage #2294's list: Response headers (incl. Retry-After) are discarded when non-success HTTP errors are constructed #2210/feat(core)!: preserve response headers on non-success HTTP errors #2211 (error response headers), feat: support OpenAI prompt cache breakpoints (Responses API) #2170/feat(openai): support manually placed prompt cache breakpoints in the Responses API #2171, feat: manually placed prompt-cache breakpoints for Anthropic #2168/feat(anthropic): support manually placed prompt-cache breakpoints #2169, fix(openai): accept Azure chat completion tool-call payloads #2255, fix(openai): forward additional_params in image generation request #2264, feat(anthropic): add Claude 4.6 adaptive thinking support #1480, fix(embeddings): respect token limits in EmbeddingsBuilder batching #1221, Standardize Rig GenAI telemetry on OpenTelemetry semantic conventions #1593 — each has an open fix PR or needs surface outside this run.

Recording matrix

Scenario	Provider	Kind	Outcome
phase_roundtrip (2-turn)	openai	tracker #2269	bug confirmed → fixed → recorded post-fix, pass
five_turn_metadata_roundtrip	openai	re-record (invalidated)	4-line diff, pass
stop_sequences	anthropic	coverage gap	recorded, pass
image_input	gemini	coverage gap	recorded, pass

Hygiene review done file-by-file: API keys/request_ids redacted, no secrets (two AIza-ish substrings in the gemini cassette are case-mismatched base64 image bytes, not key-shaped; the recorder's key-shape scan also passed), no unrelated captures, placeholders redact rather than fabricate. No hand-edited fixtures.

Verification

• Baselines on this branch before any change: openai 108 / anthropic 79 / gemini 131, all green (no pre-existing failures).
• After: cargo fmt clean; cargo clippy --all-targets --all-features clean; cargo test -p rig-core --all-features all pass; replay suites openai 109 / anthropic 80 / gemini 132 passed; wire-sharing suites also green: chatgpt 24, doubleword 21, bedrock 19.

Estimated API spend

Under $0.20: one two-turn and one five-turn gpt-5.6 recording at low reasoning effort (~2k tokens total), one claude-haiku-4-5 call (~100 tokens), one gemini-2.5-flash vision call (~1.3k input tokens).

[gold-silver-copper]

CI fix pushed: the phase capture originally sat in the shared From<Output> conversion, which also changed Copilot's replayed request bytes (its recorded gpt-5.3-codex responses carry phase), failing copilot::reasoning_roundtrip::nonstreaming as a cassette body mismatch. Capture is now in the blocking normalize path gated to provider == "openai" — the one backend whose acceptance the new recording proves; Copilot/ChatGPT/xAI request bytes stay identical to their recordings (no wire change without a recording to back it). Full local sweep after the fix: openai 109, copilot 69, chatgpt 65, xai 75, doubleword 62, anthropic 121, gemini 173, bedrock 73, rig-core 1325 — all green; fmt/clippy clean.