fix: address OpenAI compliance audit findings (streamed usage, request-id, docs)

[Abhinavexist]

Addresses the findings from the OpenAI compliance audit (run against main v0.1.0). These are orthogonal to the other open PRs (#7–#18) — none of them touched these paths.

Code fixes

• inputs.audio() mishandles data: URIs (mislabels/garbage format) #2 — streamed usage dropped (the real bug). _State.accumulate read id/model/created/delta/finish/tool_calls but never chunk.usage; worse, the usage chunk arrives with empty choices, so accumulate returned before reaching any usage read. Now it captures chunk.usage (and system_fingerprint) before the choices check and surfaces them on get_final_completion(). Verified: stream_options={"include_usage": True} → final.usage.total_tokens == 226.
• file()/video() drop computed MIME → server sees application/octet-stream #3 — _request_id lost. to_interfaze round-trips through model_dump()/model_validate(), which drops the openai-attached _request_id. Now carried onto the result. (Latent today — the API doesn't send x-request-id yet — but no longer silently discarded.)
• Appendix — pyproject Repository pointed at the non-existent interfaze/interfaze-py; fixed to InterfazeAI/interfaze-python.

Docs (README "Compatibility notes")

• Python create() raises ValidationError on tool-call responses (strict precontext) #1 — .stream() yields chunks, not events (unlike OpenAI's event-based .stream()); iterate chunk.choices[0].delta.
• Streamed <precontext>/<think> tags leak to live consumers #4 — returned text is post-processed (json_object fence unwrap, <think>/<precontext> extraction), so message.content may not be byte-identical to the wire.
• No LangChain integration; ChatOpenAI fallback drops custom response fields & rejects video #5 — inputs.* accept https URLs (Interfaze-only); not portable to OpenAI/Azure.
• Housekeeping: forecast task, precontext fixtures, system-message injection #6 — escape hatch: .parse()/.with_raw_response/.with_streaming_response live on interfaze.openai.
• Task-return note: tasks.* return the extracted result (dict/list/str), not a ChatCompletion.

Notes

• The .stream() rename the audit floats (Python create() raises ValidationError on tool-call responses (strict precontext) #1) is a breaking API decision — documented rather than renamed; happy to rename if you'd prefer.
• build()'s id/finish_reason defaults are kept (reasonable for a completed stream); logprobs accumulation is moot since Interfaze returns null.
• Tests: tests/test_compliance.py (streamed usage sync+async, request-id). Suite: 27 passing; ruff + mypy --strict clean. (CI runs once ci: GitHub Actions CI/CD (lint, types, tests, multimodal + token/usage, secret scan) #17 lands ci.yml on main; verified locally.)

[Abhinavexist]

Closing — the Issue #1 (.stream()) handling here was documentation-only, which doesn't resolve the OpenAI event-contract collision (a migrator's for event in stream code still breaks). Superseded by a new PR that properly aligns .stream() to OpenAI's event API, and re-carries the same #2 (streamed usage), #3 (request-id), pyproject-URL, and doc fixes.

[Khurdhula-Harshavardhan]

please help resolve the following.

[Khurdhula-Harshavardhan]

Rebase on main; this reverts #7 and re-breaks precontext.