fix: OpenAI compliance audit — .stream() events, streamed usage, request-id, parse/raw-response

[Abhinavexist]

Supersedes #19 (closed — its Issue-1 handling was documentation-only, which didn't resolve the contract collision). This does the proper fix and re-carries #19's other fixes.

Issue #1 (the proper fix): .stream() now yields OpenAI events

Our .stream() shadowed OpenAI's method name but yielded raw ChatCompletionChunks instead of event objects — so a migrator's for event in stream: if event.type == "content.delta" ran, then crashed. Now .stream() yields the real OpenAI ChatCompletionStreamEvents (content.delta, content.done, chunk, tool_calls.function.arguments.delta/done, …) via openai's ChatCompletionStreamState, with Interfaze's inline <think>/<precontext> side-channels filtered out of the content events (buffered across chunk boundaries, flushed before content.done). get_final_completion() still returns the extended InterfazeChatCompletion (reasoning/precontext/vcache/usage).

Live-verified against the API with reasoning_effort:"high":

event types: {'chunk': 9, 'content.delta': 8, 'content.done': 1}
tag leak in content events: False
content: "The sky is blue because of a phenomenon called **Rayleigh scattering**…"
final.reasoning present: True (314 chars) · final content has <think>: False

• Added text_deltas() for plain-token consumers; create(stream=True) unchanged (raw chunks, OpenAI Way A).
• BREAKING: iterating .stream() now yields events, not chunks (it was never OpenAI-compatible before). Deliberate — that's the whole point of the audit finding.

Also carried from #19 (unchanged, still needed)

• inputs.audio() mishandles data: URIs (mislabels/garbage format) #2 streamed usage — _State captures chunk.usage + system_fingerprint (the usage chunk has empty choices, so it's read before the choices check). get_final_completion().usage is now populated.
• file()/video() drop computed MIME → server sees application/octet-stream #3 _request_id carried through to_interfaze.
• pyproject Repository → InterfazeAI/interfaze-python.
• README "Compatibility notes" (updated: .stream() now documented as event-yielding), inputs portability, escape hatch, task-return.

Note

Depends on openai.lib.streaming.chat.ChatCompletionStreamState (openai's helper-lib surface, like our existing _thenUnwrap use). The event tests act as the guard — they'd fail if a future openai changed its accumulation. Worth a tighter openai pin if this contract must be durable.

Verify: ruff + mypy --strict clean; 36 tests pass (event-contract, no-leak-in-events, tool-call events, text_deltas, async parity, double-consume, streamed usage).

[Abhinavexist]

Rebased onto current main (post #7–#11). Also folded in audit issue #6: parse() / with_raw_response / with_streaming_response are now exposed on the wrapper (delegated to the OpenAI client) — no more AttributeError.

⚠️ Merge-order note: this PR and #16 both restructure src/interfaze/_stream.py (this one → event-based .stream(); #16 → fence-strip on the aggregated path). They're each clean against main independently, but whichever merges second will need a quick re-rebase to fold the other's _stream.py change in.

[Abhinavexist]

Also — beyond the #16 textual overlap: this PR's event-based .stream() semantically conflicts with #18, which hardens the chunk-based .stream() contract. Merging both breaks #18's stream tests. Live-verified: parse() (#6) returns clean unfenced json_schema output against the real API — no fence-strip needed (unlike json_object).

[Abhinavexist]

Confirming the breaking change is intentional: iterating .stream() now yields typed OpenAI ChatCompletionStreamEvents (content.delta/content.done, tool-call events) instead of raw chunks. Rationale: audit issue #1 — .stream() must be OpenAI-compatible so for event in stream is drop-in. create(stream=True) is unchanged and still yields raw ChatCompletionChunks for chunk consumers. Live-verified against the API.