code: bridge local Responses requests to core conversations

[erikleblansch]

Change

Add an explicitly started loopback Responses subset backed by the real core private-conversation IPC. Preserve exact text, namespaced function/custom calls and call/result correlation; return SSE only after terminal cleanup-confirmed core results. Legacy private Q&A source stays byte-identical. VOLPAROSSA remains the model/cooperation owner; this adapter adds no model, peer scheduler or tool executor.

Boundaries

Same-owner0600socket, random in-memory bearer, fixed loopback endpoint, single active request, bounded frames. Private source/results are never published, trained on, persisted or sent to OpenAI; no automatic download/start. Incomplete/invalid generations produce response.incomplete, never executable output. Unsupported grammar, strict schemas, hosted/stateful tools and reasoning reject explicitly. Exact native transport hints are accepted as documented no-ops, not model data or cache proof.

Verification

41 focused checks pass across real HTTP/Unix framing with a synthetic protocol peer and unchanged legacy Q&A tests; syntax/diff checks pass. Smol normal-Codex prompts remain too large and reject. The explicit Qwen conversation profile targets the core follow-up, not already measured model capability. No actual model/native Codex editing loop/editor test claim. Existing native lifecycle and separate core private-question proofs remain separately scoped. See docs/RESPONSES_PROVIDER.md.

Native coding integration candidate — 9126580

The explicit native harness now connects the pinned source-built Codex app-server, unchanged full upstream prompt, a VOLPAROSSA Qwen model catalog and the existing core Responses provider. Exact per-command approvals permit actual reading, model-selected arithmetic editing and real tests in a disposable synthetic project; normal editor threads remain read-only with declined approvals. Runtime source/patch/notices are verified. Closed counters require at least four real cleanup-confirmed responses, actual native tool completions, an independent post-edit test and clean shutdown.

28 targeted protocol, provider, namespace and helper tests passed on this exact candidate. These are offline contracts, not a passed native model-driven coding trial. A separate core Debian KVM scenario is being wired for that actual trial; private code is not dispatched to peers and no OpenAI inference fallback exists. See docs/NATIVE_CODING_TRIAL.md.

Integration boundary verified 2026-10-02

This PR is integrated as the local Responses/conversation bridge and explicit native-runtime harness, not a completed coding agent. All five checks on exact head 2f7014b0014b90e488b15364d6e596d7d1a30782 pass. Actual source-built Codex plus the core Qwen service have exchanged real cleanup-confirmed responses and executed a read command. The latest original core trial 36932657647 still fails the complete task: the continuation proposes a command refused by the narrow fixture, with no edit/test success. Runtime/private cleanup and unchanged host state pass. No rejected command text was retained, so its precise semantics are not claimed.

Native interactive editor operation is separate Draft #4. General coding quality, full real-editor model-driven read/edit/test, and private peer delegation remain unfinished. Normal existing editor behavior and its explicit local-only scope remain unchanged by this bridge.