fix(backend): replace sunset Assistants file-chat with Chat Completions

[Git-on-my-level]

Summary

OpenAI retired the Assistants API on 2026-08-26. Production Cloud Run backend /v2/messages file-chat has been 100% down on the non-vision path since 20:02Z: every POST /v1/threads is 404, wrapped as RuntimeError('failed to create OpenAI thread').

This PR replaces that retired surface with the already-healthy direct Chat Completions lane (SCA-361 RCA):

• Images stay on gpt-5.6-luna as base64 image_url (live-verified).
• PDFs use Chat Completions file parts {"type":"file","file":{"file_id":…}} on gpt-4.1 (documented file-input contract; luna is not pinned for an untested file-part capability).
• Upload: images purpose=vision; PDFs purpose=user_data; other docs rejected at attach as UnsupportedChatFileError.
• Assistants thread/assistant create/persist/delete is gone. Legacy openai_thread_id / openai_assistant_id remain readable on old session docs.
• File-chat 4xx is classified at the graph boundary (unsupported_attachment / provider_rejected vs stream_failure) with status_code + sanitized param only.
• search_files_tool rides the same replacement and still returns a soft error string.

Closes #3361
Closes SCA-362

RCA (SCA-361)

Last-24h chat_response failures were the Assistants file-chat lane. Historical ~2% were Assistants 400s (message-create / run-create). The hard break is the sunset: 8/8 thread creates 404 after 2026-08-26 20:02Z. Files API and chat/completions 200 on the same key. Agentic search_files_tool 404s were swallowed as journey success.

Test plan

• Sunset: PDF chat via fake completions client; zero Assistants calls; no thread/assistant ids persisted
• Stale/deleted file_id → typed unsupported_attachment error frame + journey failure
• Completion stream error after first token → error frame + journey failure
• Upload: non-PDF doc → 400 UnsupportedChatFileError; PDF → purpose=user_data
• search_files_tool provider failure returns soft error string without raising
• Do not merge until review; merge later is merge-commit only (never squash)

Verification

Focused backend unit tests for the file-chat / graph / upload / tool / inventory surfaces passed locally (25 passed on the new Completions suite plus the updated neighbors).

Failure-Class: FC-pinned-vendor-request-contract-retired

[Git-on-my-level]

Backend unit suite unblocked — the failure was not in this PR's diff. Upstream main has been failing its own Backend Unit Tests since the Soniox module refactor (7c900a9383) left SafeSonioxSocket / process_audio_soniox imported in backend/utils/stt/streaming.py without a re-export marker; strict pyright flags both as unused even though receiver.py and test_soniox_streaming.py import them through utils.stt.streaming. Since this lane evaluates the merge ref, every open PR inherited the break.

• Rebased onto current origin/main (932e2f5f06) and marked the Soniox imports as intentional re-exports (# pyright: ignore[reportUnusedImport], matching the existing SafeDeepgramSocket compat line). Net-zero line count (1707 → 1707) so the line-count ratchet stays clean.
• Verification: enforced backend/scripts/typecheck.sh 0 errors (pinned pyright 1.1.403; was 2 errors); repo-runner unit suites for test_soniox_streaming.py (11 passed), the five file-chat suites (all passed) post-rebase; full pre-push gate passed.
• Note for maintainers: main itself is still red on this lane (latest main run at 2026-08-27T16:49Z fails with the same 2 errors) — the marker fix or equivalent should land on main separately to unblock other PRs.