tests: raise cosift coverage from 50% to 56%

[TeoSlayer]

Summary

• Adds 8 new *_test.go files (2,296 lines) targeting the lowest-coverage packages.
• No production code modified; only test files added.
• Module-average coverage: 50.4% → 56.0%.

Per-package coverage delta

Package	Before	After	Delta
internal/config	35.3%	94.1%	+58.8
internal/eval	34.7%	97.2%	+62.5
internal/store	55.0%	81.5%	+26.5
internal/embed	58.3%	86.8%	+28.5
internal/crawler	55.9%	57.9%	+2.0
cmd/cosift	33.5%	34.6%	+1.1
internal/index	75.9%	75.9%	—
internal/rerank	81.1%	81.1%	—
internal/server	83.7%	84.0%	+0.3
Module total	50.4%	56.0%	+5.6

What's covered

• config: Default, Load (missing file / malformed JSON / file present / empty data-dir fallback), applyEnvOverrides (PORT / COSIFT_LISTEN / COSIFT_DATA_DIR / whitespace ignored), LoadDotEnv (missing file / quoting / comments / no-override-existing).
• eval: LoadQuerySet, LoadCorpus, LoadSummary, SaveSummary (incl. write-error), Run (with and without paraphrases, retriever errors, paraphrase tolerance), rrfFuse, score, PrintTable, Diff.
• store (SQLite): Stats, GetDocTexts (incl. maxLen), ListDocSitemapEntries, CountByDomain, CountPassagesAllModels, CountParaphrases, CountHyDE, CountDocsWithPublishedAt, ListOutcomes, Vacuum, CountQueuedPerHost, RecrawlURL (both branches).
• store (Pebble): Metrics, Checkpoint, PutVectorMeta / GetVectorMeta, PutVectorNode / IterateVectorNodes, PutVectorNodesBatch, ClearVectorFamily, PutPQCodebook / GetPQCodebook / PutPQCodesBatch / IteratePQCodes, ClearPQFamily, UpsertDocumentBatch (incl. empty-URL rejection + re-upsert ID stability), IterDocsLite, PurgeFrontierByHost, CorpusStats, SumDocLengths, ListDomains (filter + pagination + defaults), TopDomains, TopQueuedHosts, GetDocMeta.
• embed: OpenAIClient.Model/Dim, NewOpenAIChat URL normalization, RoundRobinEmbedder (empty/single/distributing), ThrottledEmbedder (max<=0 unwraps, concurrency-cap actually enforced, ctx-cancel during queue), BatchingEmbedder (defaults, single-call, empty, coalescing, inner-error propagation), CachedEmbedder hit/miss accounting on disk and in-memory.
• crawler: parseProxies (empty, malformed, no-host skipping), Robots.Allowed against an httptest.NewServer (disallow / allow-overrides-disallow / crawl-delay / bad URL), Robots.Sitemaps, robots cache (one fetch across many calls), 404-treated-as-empty, parseRobots corner cases.
• cmd/cosift (pure helpers): authStatus, resolveAPIKey (slot precedence + OPENAI / OPENAI_API_KEY fallback), resolveEmbedAPIKey, firstEnv, contains, chunkerWith, sqrt, randUnit (magnitude + determinism), percentiles, sumDur, newSeededRand, benchResult.formatHuman (all modes + unknown fallback), sortedModes, union, neutralVocabForDistractors, generateDistractorText.

Notes / observations

• The 70% target was not reachable from test-only changes: cmd/cosift is 12.6k lines (38% of the codebase) and most uncovered code paths are top-level command handlers that wire a live crawler.Crawler, store.Store, and HTTP server together and depend on an OpenAI-shaped backend. Reaching them needs either OpenAI access or extensive HTTP-fake scaffolding beyond the scope of this PR.
• One small observation while reading: store.OpenMemory records path = ":memory:" but the migration path checks columnExists via raw SQL — works fine, just notable that the in-memory store inherits the full migration code path.
• No tests touch the network; all HTTP-using tests use httptest.NewServer.
• The repo's .env file (gitignored) was not read, modified, or referenced.

Test plan

• go test -race -count=1 -timeout 120s ./... passes locally (verified)
• CI codecov shows the coverage delta on each touched package

[codecov]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

ℹ️ You can also turn on project coverage checks and project coverage reporting on Pull Request comment

Thanks for integrating Codecov - We've got you covered ☂️