refactor(responses): make openai_stream_events IRR-only

[leseb]

Summary

openai_stream_events had a logical_stream flag that turned an architectural invariant into an operator choice. This removes the flag and makes the filter always compose the current IRR execution into one logical Responses stream — normalizing identity/sequence/output indices, suppressing intermediate lifecycle events and [DONE], emitting exactly one terminal event and sentinel, accumulating output and usage across rounds, and surfacing loop-terminal failures as post-commit SSE errors — while failing closed (500 SSE error) when placed outside an iterative_request_router. Every standalone example and test is migrated into an IRR step, including full-flow.yaml, which becomes a B-shaped gateway that keeps openai_response_store pre-IRR so streaming (stream: true) responses are now persisted and retrievable via GET /v1/responses/{id}. openai_agentic_loop keeps its fail-closed guard against unsafe terminal streaming by consuming the now-unconditional logical-stream marker, and unit tests, functional integration tests, inference fixtures, and generated docs are updated to the always-on/fail-closed contract.

Related issue

Closes #993

Validation

• Unit tests — cargo test -p praxis-ai-apis (2926 passed + 4 doctests, incl. the rewritten stream_events suite)
• Integration or functional tests — cargo test -p praxis-tests-integration --test suite examples (276), session_replay (9), codex_websocket (2), make test-inference-fixtures (all replay)
• make lint (also make build and make doc with rustdoc -D warnings)

Checklist

• I reviewed every changed line and can explain the change.
• New capabilities include an example config and functional example test.
• User-facing behavior and generated documentation are updated.
• Performance-sensitive changes include appropriate benchmark or load-test evidence. — N/A; this refactor removes a flag and adds no perf-sensitive path.
• Commits are signed and include a Signed-off-by trailer.

Breaking changes

Any config still setting openai_stream_events.logical_stream now fails to parse (deny_unknown_fields) — remove the flag, as the filter always composes. openai_stream_events must run inside an iterative_request_router step: a stream: true Responses request with the filter placed outside an IRR is rejected with a 500 SSE error. Migrate standalone placements into an IRR step (see stream-events.yaml and agentic-loop.yaml).

[praxis-bot]

praxis-bot review: refactor(responses): make openai_stream_events IRR-only

Clean refactor that hardens the contract: openai_stream_events now always composes logical streams and must live inside an iterative_request_router step. The logical_stream config flag is removed, the fail-closed guard rejects misconfigured placements, and all examples/tests/docs are updated accordingly. The architectural invariant is well-enforced and the test coverage (both unit and functional integration) is thorough.

One medium finding below.

---

Reviewed files: 23/23
Classification: Medium (1)

[franciscojavierarceo]

the current OpenResponses conformance job is still failing six cases with HTTP 504 upstream timeouts (one local schema case passes), so the required check is not green on this head. please rerun and diagnose the conformance job before merging; if the failures are environmental, make that evidence clear in the check or test setup.