test: coverage audit — real fixtures, dead tests removed, streaming and MMR covered

[lucaosti]

Closes #154.

What this does

The coverage-audit layer of the quality gate (#146, layer 1/6): measure what's actually tested, find what isn't, and close the highest-value gaps rather than chase a percentage.

Skipped tests: 19 → 0

17 of the 19 skips had nothing to do with hardware or models — they were guarded on two fixture files that were never committed: docs/src/bitcoin_technical_document.pdf and bitcoin_creative_commons_en.pptx. *.pdf/*.pptx are globally gitignored (local upload cruft), and these two fixtures got caught by that same rule instead of being carved out as an exception.

Generated real, structured fixtures rather than placeholders — a 4-page PDF (headings, paragraphs, a formula, a table) and a 5-slide PPTX — using pymupdf and python-pptx, both already backend dependencies, so nothing new was added. All 17 previously-skipped tests pass against them unmodified.

The remaining 2 (test_chunker.py, test_ingester_parser.py) guarded on module_1_ingestor / module_2_parser / module_3_micro_chunker — never declared as a dependency anywhere, permanently skipped since introduction. They describe a three-level chunking hierarchy that predates the current two-level parent/child design in pipeline.py. Deleted, not fixed: a test for an architecture the code no longer has is misleading, not just inert.

Assertion-free scan

4 flagged by a static scan of every test_* function for a missing assert/pytest.raises. All four are legitimate does-not-raise checks — pytest fails a test on an uncaught exception, so their absence of a raise is the assertion. Strengthened one (test_module_aliases_are_registered_without_error) with an explicit check of the registered alias, since that was cheap and is the actual point of the test.

New coverage: streaming and MMR

The two real gaps the audit surfaced: the SSE streaming paths (chat_service.stream_answer, study_service.stream_dispatch) at ~53-54% despite being the primary interaction mode today, and reranker.mmr_select's actual diversity algorithm — previously only its no-model fallback was exercised.

• 11 streaming tests: cache short-circuit (no retrieval on a hit), QUIZ/ORAL/RETRIEVE always buffered never streamed, progressive token yield, empty-stream → buffered-generate fallback, a QVAC [ERROR] token mid-stream treated as failure rather than shown to the user, full degradation to a raw source snippet when everything is down.
• 4 MMR tests with a fake embedding model returning controlled vectors: a diverse lower-relevance chunk beats a redundant higher-relevance one for the second slot, lambda_=1.0 matches pure relevance order, inference failure degrades to a plain top-k slice.

One test documents a finding rather than an assumption I got right the first time: when both streaming and buffered generation fail, the fallback answer is raw_answer — QVAC's own bare string from the initial retrieval call — not pack.context_block(). So retrieved evidence sitting right there in the pack can be silently discarded if raw_answer happens to be empty. The identical branch exists in the non-streaming dispatch() path, so this isn't new. Filed as #172 rather than changed here — this issue is about testing and reporting, not product behaviour.

Numbers

	Before	After
Backend tests	915	946
Skipped	19	0
Coverage (branch)	78.02%	82.04%
chat_service.py	53%	80%
study_service.py	54%	73%
reranker.py	54%	77%
pipeline.py	61%	73%

mypy clean across 84 files.

Not done here (by design)

Per-endpoint success/validation/not-found matrix and per-service error-path mapping — #154 already flagged auth coverage as complete (test_authorization_matrix.py) and asked for a report, not a rewrite of every layer. The remaining line-level gaps (ChromaDB fallback path in chat_service, two-hop compare/derive merge logic, the low-level _stream_generate SSE parser) are listed in the closing comment on #154 as scoped candidates for whichever of layers 2-6 fits each.