test: production-harden the suite (error contract, tasks, async, real URLs, coverage gate)

[Abhinavexist]

Turns the happy-path suite (24 tests, ~81% line coverage) into a production-grade one: 119 tests, 100% line+branch coverage, gate floor at 90%.

What was missing (and is now covered)

• Error contract (test_errors.py) — status→exception mapping 400/401/403/404/429/500/503 → the right openai type (asserting status_code and the error body reaches the caller); retries fire on 429/5xx (respx side_effect + monkeypatched sleep, assert call_count); timeouts → APITimeoutError. Sync + async. (Previously zero error tests.)
• Task helpers (test_tasks.py) — all 8 (ocr/object_detection/gui_detection/transcribe/web_search/scrape/translate/forecast), sync + async: assert the <task> tag on the wire and result extraction; forecast precontext / fallback / skip-unrelated-entries paths; _extract edge cases. (Was ~48% covered.)
• Async parity (test_async.py) — real async coverage, not one smoke test.
• Streaming failure modes — empty stream, streamed tool-call accumulation, double-consume, malformed SSE, mid-iteration get_final_completion.
• Inputs — from_path (real temp file), blacklist via from_path, unknown-extension, auto_part routing.
• Guard / schema (test_guard.py, test_schema.py) edge cases.

Real URLs

Added tests/assets.py with the canonical Interfaze/JigsawStack doc asset URLs (OCR receipt, STT wav, forecast CSV, YOLO bus.jpg, arXiv PDF, HN, …) and replaced the fake https://x.com/... placeholders.

Coverage gate (self-contained)

pytest-cov in the dev extra + [tool.coverage.run] branch=true + addopts=--cov --cov-fail-under=90 in pyproject.toml. Travels with the tests, so any CI running pytest enforces it. Tests only — no src/ changes.

Findings the hardened tests surfaced (documented in the tests; flagging for follow-up)

1. gui_detection silently mis-routes extensionless image URLs. A CDN URL with no .ext (e.g. Unsplash) makes auto_part fall through to a generic file part instead of image_url (mime detection is extension-only), and tasks.gui_detection() has no filename/format hint to fix it. Real silent-failure gap for a common URL shape — worth an SDK follow-up.
2. Malformed SSE data: lines raise a raw json.JSONDecodeError from openai's SSE decoder (before interfaze code runs), rather than a typed error — a stream robustness gap.
3. All 5xx map to InternalServerError (a subclass of APIStatusError), not a distinct type per code — tests assert actual behavior.
4. Pre-existing ruff format drift in 6 src/ files (untouched here; deferred format-check from the CI PR).

[Khurdhula-Harshavardhan]

please help resolve the following.

[Abhinavexist]

⚠️ Semantic overlap with #20 — needs a design call, not just a rebase. #20 changes .stream() to yield OpenAI events instead of raw chunks (a deliberate breaking change). This PR hardens the chunk-based contract (e.g. list(stream) == chunk count, tool-call accumulation by iterating, role-from-first-delta). Each PR is green on its own, but merging both breaks this PR's stream tests against #20's event contract. Decide first whether .stream() is chunk- or event-based going forward; #20 obsoletes part of this PR's stream hardening, so those tests must be rewritten to the event API (or #20 deferred).

[Khurdhula-Harshavardhan]

please help resolve the following.