fix(deepseek): transport-neutral bounded-JSON policy with terminal SSE synthesis (#875)

[lidge-jun]

Summary

Stack 04 of the bug-stack campaign (devlog/_plan/260805_bug_stack_campaign/050), stacked on #1025. Fixes #875.

The registry's upstream-streaming hint was WS-scoped: ordinary HTTP turns kept stream:true into DeepSeek's terminal-less SSE and stalled after tool calls (reporter reproduced at e44d234 with websockets:false, bodyKind:sse).

• modelResponsesUpstreamStreaming (renamed, transport-neutral) applies after final wire resolution for openai-responses adapters — DeepSeek Flash now uses bounded JSON upstream on every transport.
• Shared bounded-JSON event sequence (src/server/responses-json-events.ts): the WS bridge and the new HTTP synthesis emit the same canonical frames — response.created → one response.output_item.done per item → status-preserving terminal; HTTP streaming clients get SSE with exactly one [DONE].Non-streaming clients and WS turns keep plain JSON (gate excludesinboundTransport: "websocket"`).
• Supersedes PR fix(deepseek): apply bounded JSON policy on HTTP SSE for Flash #1006's mechanism with the contributor's core policy retained (attribution) — its duplicated event algorithm, missing [DONE] trailer, and bypass of the SSE item-ID repair branch are avoided.

Tests

• Stall activation: fake upstream fails loudly if stream:true is ever sent (the old hang), and the client receives the full synthesized sequence with [DONE]; function_call id/call_id byte-identical.
• WS/HTTP serializer parity + status preservation (failed/incomplete) in tests/responses-json-events.test.ts and tests/ws-endpoint.test.ts.
• Independent review: GO-WITH-FIXES(1) folded (WS transport exclusion).
• bun run typecheck 0 errors; bun run privacy:scan pass.
• Full bun run test on Linux (ssh lidge): 8288 pass / 0 fail (baseline 8222).

Limitations

DeepSeek Flash loses progressive token delivery — the intended provider-specific reliability tradeoff. structure/04 documents the transport-neutral policy.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (2)

• ^dev$
• ^preview$

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: e7938d5e-fd76-431d-a35d-beda3aed50db

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[lidge-jun]

Rebased onto the current stack head, with a review blocker folded in.

Blocker: this layer was not independently safe with item-id repair.

The bounded-JSON policy answers a streaming client by synthesizing SSE from a completed JSON body, and reframes the same body into events for WS turns. Neither path goes through the SSE relay, so neither picked up the relay's item-id rewrite. The consequence is worse than a missing feature: for a provider that has id repair configured, enabling this reliability policy would silently disable it — canonical ids while streaming, raw upstream ids the moment the policy engaged. The later #938 PR happened to fix this, which meant this PR was only safe if its child landed with it.

710492cb2 moves that normalization into this layer: repairResponsesJsonItemIds for whole-object rewriting, applied to the synthesized-SSE and WS-reframe paths after the raw recording. New tests prove repaired ids on both bounded-JSON paths and that a provider without repair keeps the body byte-identical. #1027 was rebased on top and now carries only its own repairInvalidIds work.