fix: serialize cold eval/task cache init across processes

[Arshgill01]

Fixes #3133.

Concurrent devenv shell entries into a project that does not yet have .devenv/ raced on eval-cache (and task-cache) SQLite initialization. One process created the database; another failed with Failed to initialize eval cache database: Database error: (code: 5) database is locked. Under more contention the same window also produced task-cache init failures, SQLITE_IOERR_DELETE_NOENT (code 5898) after a loser deleted the DB, and Failed to remove existing GC root: No such file or directory. Warm .devenv/ was already fine.

busy_timeout(10s) and WAL with DELETE fallback on SQLITE_IOERR_SHMMAP were not enough: sqlite3_open(CREATE) / PRAGMA journal_mode=WAL can return SQLITE_BUSY before the timeout is installed, and a migration error used to delete the database out from under the process that held it.

Changes

• Take an exclusive flock on a sibling of the SQLite file (*.db.init.lock, not the .db itself) around create + migrate in Database::new, so eval cache and task cache share the same cold-init serialization.
• Hold that lock on a dedicated OS thread so waiters do not stall the tokio runtime.
• Never delete the database on a busy/locked migration error; retry instead.
• Treat a missing GC-root symlink as success when replacing it, so two shells creating the same root no longer lose the TOCTOU race.
• Regression tests: concurrent cold/warm Database::new, concurrent cold TaskCache::new, and tests/cold-concurrent-shell (two concurrent devenv shell -q -- true against a missing .devenv/, then the same against a warm one).

Test plan

• cargo test -p devenv-cache-core (includes concurrent_cold_open_succeeds / concurrent_warm_open_succeeds)
• cargo test -p devenv-tasks --lib -- test_cold_cache_allows_concurrent_open test_warm_cache_allows_concurrent_open
• cargo clippy -p devenv-cache-core --all-targets -- -D warnings
• devenv-run-tests run tests --only cold-concurrent-shell (needs a full devenv/Nix environment; not run in this agent VM)

[github-actions]

🔍 Suggested Reviewers

Based on git blame analysis of the changed lines, the following contributors have significant experience with the modified code:

• @sandydoo - 100.0% of changed lines (9 lines)

Please consider reviewing this PR as you have authored significant portions of the code being modified. Your expertise would be valuable! 🙏

This comment was automatically generated by git-blame-auto-reviewer

Last updated: 2026-09-22T03:49:29.191Z

[Arshgill01]

Rebased onto latest main (behind_by: 0, mergeable). Kept the #3133 changelog entry with main’s newer 2.3.0/2.3.1 notes. Preserved flock around Database::new and GC-root NotFound-as-success on ensure_gc_root.

[Arshgill01]

Rebased onto latest main (behind_by: 0, mergeable). Kept the #3133 changelog entry with main’s newer 2.3.1/2.3.2 notes. Preserved flock around Database::new and GC-root NotFound-as-success on ensure_gc_root.