fix(search): skip unreadable blobs with a warning instead of aborting

[hammadmajid]

Description

Since 6393c33 a single blob that cannot be read (zero-byte or corrupt loose object, bad zlib stream, bad pack entry) aborted the entire search: workers pushed the read error onto errCh, Pipeline.Execute returned it, and Aggregator.AggregateChannel bailed on the first res.Error. One missing blob in git fsck meant exit 2 and no output at all. v0.1.0 tolerated this by accident; git grep and ripgrep both warn on unreadable inputs and keep going.

A per-blob read failure is now a soft error: the blob is skipped with a warning on stderr and the search continues, emitting results for every other blob. Context cancellation, matcher failures, history-walk and repository-discovery errors are unchanged and remain fatal.

Fixes #10

Type of Change

• Bug fix (non-breaking change fixing an issue)
• New feature (non-breaking change adding functionality)
• Performance improvement
• Refactoring or code cleanup
• Documentation update
• CI/CD or build workflow change

Key Changes

• internal/search: new BlobReadError{OID, Path, Err} (Path = first occurrence path). processTask produces it for reader.ReadObject failures; workers still deliver that result on resultsCh (so callers can warn) but no longer push it onto errCh. All other errors keep the existing fatal path.
• internal/aggregator: AggregateChannel skips BlobReadError results instead of returning; any other res.Error still aborts.
• cmd/grg: prints grg: warning: skipping blob <oid> (<path>): <err> per skipped blob to stderr before any match output; a new skippedBlobsError (exit 2, no extra message since each blob was already reported) folds the count into the exit code. The pre-existing exit-code error types were moved into cmd/grg/errors.go in a separate pure-move commit (main.go was over the 350-line target).
• README.md: exit-code section updated to the ripgrep semantics below. It also previously claimed 128 for repository errors, which the code never returned.

Exit-code semantics

Outcome	Exit
Match found, no blob skipped	0
No match, no blob skipped	1
Any blob skipped (matches may still have been printed)	2
-q and a match found (even with skipped blobs)	0
-q, no match, blob skipped	2
Fatal error (CLI, repo discovery, cancellation, ...)	2, unchanged

Warnings are printed to stderr even under -q, matching ripgrep (which has a separate --no-messages for suppressing them).

Verification & Testing

• Ran go test -v -count=1 ./...
• Ran go test -race -count=1 ./...
• Ran go vet ./...
• Added or updated unit/integration tests
• Tested manually against sample Git repository histories (covered by the integration tests below, which drive the built binary against a real repo)

Also ran golangci-lint run ./... (0 issues), go mod verify, and govulncheck ./... (0 vulnerabilities in reachable code).

Tests:

• internal/search: TestPipelineExecuteContext_ReaderError previously asserted that a read failure reached errCh, i.e. the exact fatal behaviour this fixes; it now asserts the soft BlobReadError path. New TestPipelineSkipsUnreadableBlob uses a fake ObjectReader where one OID returns ErrObjectNotFound / ErrCorruptObject and asserts the other blobs' matches are returned, the failing OID and first path are reported, and no fatal error is returned.
• internal/aggregator: TestAggregator_AggregateChannel_SkipsBlobReadError; the existing test asserting a generic res.Error is fatal is kept.
• test/integration/corrupt_object_test.go: creates a real repo, commits and later deletes a file (so the blob is only reachable through history, as in the issue), truncates its loose object to zero bytes, and asserts: matches from the healthy blob are printed, stderr has exactly one warning naming the OID and path, exit 2; exit 2 with no stdout when only the corrupt blob could match; -q semantics above; and a healthy repo still exits 0/1 with empty stderr.

Checklist

• My code adheres to the project's coding style and architecture guidelines (modular, files under 350 lines).
• No Cgo or external runtime dependencies have been introduced.
• Existing tests pass and new tests cover added functionality.
• Documentation / README has been updated where appropriate.

🤖 Generated with Claude Code

https://claude.ai/code/session_016ut4mEjEnEXdJZejxFm8gT