tests: cmd/cosift round-4 — fake fetcher + fake cluster

[TeoSlayer]

Summary

Round-4 coverage push for cmd/cosift. Lifts coverage from 50.5% (post-round-3) to 65.3% by extending the round-3 scaffolds (OpenAI httptest mock + populated Pebble fixture) with a fake peer cluster and whole-runner sweeps.

What's new

• fakeCluster — array of httptest.Server peer shards. Drives scatterSearch + the four gateway handlers (handleSearchGateway, handleFindSimilarGateway, handleAnswerGateway, handleResearchGateway) end-to-end without spinning up real cosift instances.
• Whole-runner sweeps for previously 0%-covered entries:
  • runServe — full lifecycle (free-port pick, /healthz round-trip, graceful shutdown) + the admin-token / trusted-proxies / chunker / defaults wiring branches + bad-CIDR error path.
  • runCrawl — sqlite + pebble backends, -seeds-file (happy + missing + comments/blanks), -duration, pre-cancelled ctx, unknown backend.
  • runEval — -api short-circuit against a fake httptest endpoint (with -save), local-bm25 path against a synthesized corpus, unknown-retriever guard, dense-without-key guard.
  • runBench — vector / bm25 / storage / -json modes (tiny N so the suite stays fast).
  • runAnswerEval -dry-run + bad-key/bad-corpus/bad-queries guards.
  • runAnswerEvalCompare — multi-strategy diff with per-query moves, no-moves, wrong-arg, missing-baseline, missing-new paths.
  • runCrawlErrors (empty + custom-limit) and runVerifyPebble (local + via -server happy + drift + -json).
• Direct unit tests for the iter-391 hnswPassageWriter bridge (UpsertPassage, UpsertPassageBatch, MarkURLInvalid + ctx-cancelled).
• Eval-side retriever adapters via stub Retriever + stub ChatClient: hybridAdapter, rerankAdapter (happy / inner-err / single-URL short-circuit / rerank-err fallback / candidateK default), paraphraseRetriever (chat-err / main-weight / inner-err / cache hit), plannerRetriever (happy / chat-err / unparseable / cap-at-5), httpAPIRetriever (happy / non-200 / dial-err / bad JSON).
• SSE coverage for streamAnswer + streamResearch via real httptest.NewServer (httptest.NewRecorder doesn't implement http.Flusher).
• doRerank counter wrapper (happy + err) via a realFakeReranker that satisfies the real rerank.Reranker interface.
• applyBM25EnvOverrides — valid / malformed / negative / unset.
• Admin endpoint happy-path + non-auth branches not already covered by round 2 (handleCheckpoint, handleCrawlEnqueue happy/hook-err/bad-body, handleSitemapImport happy/bad-body/hook-err, handleFrontierPurgeHost happy/bad-body, handleSitePack bad-host/missing-host, handlePQTrain no-hnsw, handlePQEncode no-codebook, handleWETImportBulk bad-body/manifest-unreachable, handleCrawlNow empty-urls).

Test plan

• go test -race -count=1 -timeout 180s ./cmd/cosift/... passes
• Coverage 65.3% (baseline 50.5%) — verified via go tool cover -func=…
• No production code modified — only cmd/cosift/zz_round4_test.go added (2061 LOC)
• Builds on existing round-3 scaffolds (openaiTestServer, populatedPebbleStore, seedTinyStore, captureStdoutCosift); no duplicate test names with rounds 2 or 3

What's still gated

• runReembed (16.2%), expandQuery (8.7%), handleEvalQuick (10.4%), handleEmbedBackfill (13.8%), handlePQTrain (12.9%), and the streaming-detail tails inside streamAnswer / streamResearch still have uncovered branches — most need either a working OpenAI key, a deep multi-step setup with passages persisted to disk, or the iter-402 in-serve crawler (startInProcessCrawl) which is currently entirely 0% and would require a substantial new fake-fetcher harness.
• main() is unreachable from unit tests.
• The four eval-internal adapter Search methods at main.go:2547 (denseAdapter), 2592 (rerankAdapter — partly covered), etc. — denseAdapter is still 0% because it requires a working Embedder; the mock embedder could be wired here in a follow-up round.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!