[mkarlesky]

Third PR in the lint-and-coverage series, stacked on #22. It covers two parts of the triage report:

• the performance findings (async file access, concurrent discovery), which also fix two concurrency bugs found by stress testing;
• unit tests for the uncovered behavior in the pure modules.

Adapter extraction (PR D) and fake-Ceedling integration tests (PR E) follow separately.

Async file access

• Discovery, configuration, watcher, and report paths no longer block the extension host thread.
• EngineFileSystem is fully async.
• projectsFromSettings() and configurationFiles() return promises.
• The no-restricted-properties warnings on sync fs in src/ drop to zero, apart from the documented /proc reads in ceedlingProcess.ts.

Concurrent discovery

• Projects load concurrently across directories and one at a time within a directory (mapGrouped in the new src/concurrency.ts).
• Load outcomes apply in settings order, so error messages don't depend on timing.
• Listings run concurrently across directories, and test sources are read concurrently.
• Tree order is unchanged.

A lock per project directory (fixes a real race)

The race: Every Ceedling command that loads a project rewrites the vendored Unity and CMock files under build/vendor/ in place, dumpconfig included. Discovery's dumpconfig ran outside the old engine-wide mutex. So a rediscovery during a build in the same directory could leave gcc reading a half-written unity.h. The S2 stress run below reproduced this against real Ceedling as spurious TEST_CASE compile errors.

The fix:

• The engine-wide mutex is replaced by KeyedLock, one lock per project directory.
• dumpconfig, listings, runs, debug builds, and clean/clobber all hold their directory's lock.
• Different directories proceed independently, so builds in separate projects may now overlap.
• The async-mutex dependency is removed.

One shared read limit (fixes EMFILE)

All source reads share one Limiter capped at 16. Per-call bounds weren't enough: 600 simultaneous reparses plus two rediscoveries exhausted a 64-descriptor limit (S3).

Stress gate (one-off scripts, not committed)

All three ran against the final code with 0 violations and 0 unhandled rejections:

Run	Setup	Checks
S1	Compiled engine against a simulated Ceedling with random latency, failures, and cancellation. 60 seeds × 300 overlapping operations, up to 20 in flight.	No two project-loading commands in one directory at once; project args and cwd on every command; exactly one result per leaf and none after a run ends; no commands after dispose; no live processes or temp files at the end.
S2	Real Ceedling 1.0.0, 1.1.10, and 1.2.0 in the madsciencelab images. 150 s of overlapping discovery, runs, debug builds, clean, and reparses across four projects (two sharing a directory), with random cancellation.	Every completed run's results match the fixture's expected outcomes; no errored results; no leftover processes (via /proc); no leftover temp files.
S3	A 3,000-file × 20-case project on real disk under ulimit -n 64. Two concurrent discoveries plus 600 reparses.	Full file count, listed order, case count; no logged errors.

Tests

New unit tests:

• concurrency.test.ts covers mapGrouped, KeyedLock, and Limiter.
• A "concurrent discovery" suite in ceedlingEngine.test.ts covers directory grouping, error ordering, lock interplay between rediscovery and builds, the shared read limit, and read order.

Results: 311 unit tests and 4 integration tests pass, and lint reports 0 findings.

Coverage of the pure modules

File	Statements before → after	Branches before → after
ceedlingEngine.ts	94.7% → 100%	90.0% → 98.1%
ceedlingProcess.ts	80.3% → 99.0%	97.4% → 96.2%¹
problemMatching.ts	100%	80.6% → 100%
runPlan.ts	100%	98.2% → 100%
unityTestDiscovery.ts	96.6% → 99.0%	89.3% → 93.2%
All files (adapters included, still at 0%)	69.9% → 72.9%	91.4% → 96.9%

¹ The new injection point adds branches.

What the new tests reach:

• Engine: unparseable and missing dumpconfig output, a failed test listing, discovery warnings, an invalid file prefix for pretty labels, ignored and unmapped results, development-mode logging, a cancelled debug build, undeletable reports and temp files, project lookups, cancel and dispose, the shell and command settings, and the real filesystem and runner defaults.
• Process: killProcessTree against a real process tree on every OS, and its fallback when tree-kill fails. The fallback is reached through an injected platform and treeKill.

One behavior change: a TEST_RANGE with no bracketed range now gives a discovery warning, as a TEST_MATRIX with no lists already did. Unity's generator fails on both (undefined method 'map' for nil). This was checked in the Ceedling 1.2.0 image.

What stays uncovered:

• ?? fallbacks on captures that always participate;
• a race guard for a process exiting while /proc is being read;
• the engine's path.win32 selection. Unit tests always inject platform: 'linux', so this only runs on a real Windows host. The fake-Ceedling integration tests (PR E) run on the Windows CI leg and will cover it.

Docs

• docs/Development.md: the module table gains concurrency.ts. Discovery describes what overlaps, the shared read limit, and dumpconfig's two forms. A new paragraph explains why commands in one project directory run one at a time.
• CHANGELOG.md:
  • Changed: faster discovery, and commands in separate project directories may now run at the same time.
  • Added: the warning for a TEST_RANGE() or TEST_MATRIX() that Ceedling can't build.
  • The vendored-header race gets no entry: 2.0.1 held its mutex around files:test, so it only ever existed on next_version.

Upstream

Ceedling's non-atomic, in-place re-vendoring on every project-loading command is reported as ThrowTheSwitch/Ceedling#1361. This PR works around it.

🤖 Generated with Claude Code