fix(summarizer): raise on stream failure instead of swallowing it as an empty summary

[Optic00]

Problem

When a summary stream fails (Ollama returning 404 "model not found", a cloud API error, or an adapter error record), the failure was swallowed and reported as success: the user got an empty note ("No summary available for this meeting") with no error surfaced anywhere. Reproduced end-to-end against a real meeting with a broken Ollama: the backend logged ERROR - Ollama streaming failed: ... 404 but still printed STREAM_COMPLETE + Summary reprocessed successfully and exited 0.

Root cause: the streaming generators in src/summarizer.py logged provider errors and then ended the generator silently (logger.error(...); return). Everything downstream was already correct - all five consumer sites in simple_recorder.py catch generator exceptions and emit STREAM_ERROR + exit 1, and app/main.js maps that to failure events. The generators just never raised into that path. This is also the root cause of #301 (cloud streaming failures swallowed).

Fix

src/summarizer.py only - no consumer changes needed:

• _stream_completion: all four provider paths (Anthropic, Bedrock, OpenAI-compatible, Ollama) now log and re-raise the original exception, so the real message (e.g. model 'x' not found (404)) reaches the UI via STREAM_ERROR.
• _adapter_stream: an NDJSON error record, an HTTPError, and transport errors now raise after logging. Its second consumer (query_transcript_streaming, the chat path) already catches provider errors and yields a visible [Error: ...] chunk, so chat via the adapter now shows the error instead of an empty answer.
• summarize_transcript_streaming: defense-in-depth empty-stream guard - a stream that completes without raising but yields no non-whitespace content raises ValueError instead of saving an empty summary. Mirrors the existing empty-reduce guard in _map_reduce_streaming.

A failed regenerate can no longer clobber an existing good summary: the consumers exit before the write.

Tests

• Unit (tests/test_summarizer_stream_errors.py, 11 tests, written first and failing before the fix): error propagation for the Ollama, Anthropic, Bedrock, OpenAI-compatible and adapter paths (error record / HTTPError / transport), the empty-stream guard (plain + template path), and a success-passthrough regression guard.
• e2e (model-free T2, e2e/specs/summarize-failure.t2.spec.ts): drives a real reprocess against a mock Ollama answering /api/chat with the real 404 shape (new opt-in chatError option in mock-ollama.js, default off - existing specs unaffected) and asserts the operation reports success:false and the pre-existing summary on disk stays byte-for-byte unchanged.
• Full backend suite green, summarize-contract e2e re-run green (mock change is non-breaking), backend bundle rebuilt before the e2e run so the spec proves the fix end-to-end.

Fixes #301

---

Summary by cubic

Surface real errors from summary streaming and stop saving empty notes. Failures (e.g., Ollama 404, cloud API, adapter errors) now trigger STREAM_ERROR and never overwrite a good summary.

• Bug Fixes
  • All provider streaming paths (Anthropic, Bedrock, OpenAI-compatible, Ollama) now log and re-raise exceptions so the UI gets STREAM_ERROR.
  • Adapter stream raises on {"type":"error"}, HTTPError, and transport errors; chat path shows an inline [Error: ...] chunk.
  • Added empty-stream guard: if no non-whitespace is produced, raise ValueError instead of saving an empty summary.
  • Prevents failed regenerates from clobbering existing summaries. Fixes Cloud (Bedrock/OpenAI/Anthropic) streaming failures are swallowed, producing an empty "successful" summary #301.

^{Written for commit 49e73b4. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 4 files

_{Re-trigger cubic}