Fix concurrency/file-presence findings from Stage 1 fact-finding

[mkarlesky]

Summary

Implements fixes for the four concurrency/file-presence findings identified in a prior fact-finding investigation of Ceedling's test (and, more lightly, release) build pipelines, plus the release-pipeline mutex issue found alongside them. Findings and experimental evidence are recorded in this session's memory (concurrency-file-presence-factfinding); this PR is the follow-on fix-and-validate stage.

• Finding 1 — release pipeline counter race. ReleaseBuildExecutor#compile_objects's skipped += 1 was an unguarded compound read-modify-write across worker threads. ReleaseState gains a :lock field, defaulted automatically; the increment is now synchronized — exactly matching the test pipeline's own established pattern. No validation experiment needed (well-understood pattern, direct fix).

• Finding 3 — no-guard preprocessing read (prime suspect for sporadic Windows file-miss reports). A file a shell-out (gcc) just wrote wasn't guaranteed visible to the very next read — surfacing as an uncaught, unretried crash. New FileWrapper#open_with_retry/#exist_with_retry? (four retries, ~0.37s worst case, scoped to SystemCallError only so a real programming bug is never masked) ride out a transient failure instead of crashing on the first attempt. Applied at all ten preprocessing call sites that read gcc output. A permanently-failing open now raises a filepath-naming CeedlingException; a permanently-missing file logs a NOTICE before falling back to prior behavior.

• Finding 2 — silent dependency-loss on a missed .d file. Same retry mechanism (exist_with_retry?) applied at all eight .d-dependency-file presence checks in the test and release pipelines, closing most of the "self-corrects into one extra recompile" gap directly.

• Finding 4 — two-pass compile barrier invariant. Confirmed safe today and load-bearing, not incidental. No code change — documented the invariant with explicit comments at stage_build_objects/apply_sibling_isolation, plus a committed integration-tier canary spec that will fail if the barrier is ever merged away or reordered.

Tests

• New spec/integration/preprocessing_transient_io_retry_spec.rb and spec/integration/test_build_parallel_barrier_spec.rb commit the fact-finding phase's deterministic fault-injection experiments as permanent regression coverage.
• New unit cases in file_wrapper_spec.rb (stub-based, no real filesystem/timing) prove the retry mechanism itself; call-site specs updated to track the renamed collaborator methods and stay green.

Verification

• Full unit suite (3260 examples) and integration suite (37 examples): green on host (macOS) and the madsciencelab-plugins Docker image (Linux).
• Targeted preprocessing/release-build system specs: green on both platforms (one pre-existing, unrelated host-locale failure in delta_builds_spec.rb confirmed present on unmodified next_version too, and passes cleanly in Docker).

🤖 Generated with Claude Code