Testing the parts that mattered: the AI pipeline, the default parser, and the isolation gates
16 source files had tests; 103 didn't. Rather than chase a coverage number, I tested the four things whose failure would hurt most: the async analysis pipeline, the default provider's parser, the SRS streak logic, and the per-user isolation boundaries.
The audit's bluntest finding: 16 test files for 103 source files, and the app's signature feature — the async AI analysis pipeline — had zero tests. The wrong response is to chase a coverage percentage. The right one is to ask which untested code, if it broke silently, would do the most damage, and test that.
Four answers:
1. The async analysis pipeline. ClipAnalysisService.runAnalysisPipeline orchestrates the whole "clip saved → analyze → persist" flow with transactional bookends around a network call and a self-injected proxy for the @Async hop. I test it synchronously: stub the self-proxy to return the instance, mock the repos and the provider, and assert the three branches — provider succeeds → analysis marked READY; provider not configured → FAILED and the provider is never called; clip already deleted → short-circuits quietly. The branching is the logic; the test pins it.
2. The default provider's parser. TubeShadow defaults to Gemini, and GeminiClient.parseResponse had no test even though it does the fiddly candidates → content → parts → text unwrap and then parses the inner JSON for the sentence-mining fields. I added the round-trip (including primary_translation / chunked_translation), the markdown-code-fence strip, and — importantly — the failure contract: a response truncated mid-string (the real gemini-2.5 thinking-token bug from an earlier log) must surface as a clean BusinessException, not a raw Jackson error.
3. The streak logic. computeStreak is pure and full of off-by-one traps: consecutive days ending today, a gap that breaks the run, the "today not done yet but yesterday was" case that should still show a streak, and the empty/stale → 0 case. It's package-private, so the test lives in the same package and feeds it LocalDates directly — no DB, no mocks, just the logic.
4. The isolation gates. The whole product promise is "your library is yours." Two boundaries now have explicit tests: ReviewService.respond rejects another user's review item (the findByIdAndUserId gate → NotFoundException), and DeckService.moveClip rejects both a clip you don't own and a target deck you don't own. These are the tests I'd want a reviewer to see, because they encode the security model as an assertion, not a hope.
14 tests, all green. Not a coverage trophy — coverage of the things that would actually hurt.