PTY snapshot suite: rare non-repeating case failures under parallel load (pass in isolation)

just snapshot-test migration と PTY snapshot runner から始め、rust-toolchain.toml と記載されている migration fixture を再現環境として使用します。コールドキャッシュでの並列実行と、分離した実行およびウォームキャッシュでの実行を比較し、提案されているタイミング、詳細度、またはリトライの計測を追加します。負荷に敏感な失敗が説明され、migration suite が説明のつかない間欠的な失敗を生成しなくなれば完了です。

Summary

Running the full migration* PTY snapshot suite in parallel occasionally fails a small, non-repeating set of cases (~0.4% of case-executions). Every affected case passes deterministically when run in isolation, and the failing names differ between runs — so this looks like load-sensitivity in the PTY runner environment rather than any individual fixture or product bug. Filing per m0g3r's suggestion in #2483 after it showed up there first.

Observations (macOS arm64, M-series, just snapshot-test migration)

Five full-suite runs across two commits, all with a fresh packages/cli/dist (freshness guard green):

• No case name repeats across runs. Union of failures over five runs: 10+ distinct fixtures, each failing exactly once.
• Every one of them passes in isolation (just snapshot-test <name>), immediately after the failing run, unchanged tree.
• The failure-heavy runs were the first suite run after a fresh build / cold caches; warm re-runs (C, D) were fully green. Cold-state work on first touch (managed runtime provisioning, registry-bridge first hits) overlapping with ~14 parallel PTY cases is my best guess at the mechanism, but I have not isolated it.
• m0g3r additionally verified in #2483 that the flaked fixtures are structurally unremarkable (step counts mid-pack; one of 21 double-vp migrate fixtures).

Why it may matter

Locally it's a shrug; in CI a 0.4% per-case flake across ~650 cases makes a meaningful fraction of runs red for reasons unrelated to the diff under test, and the changing names make it hard for contributors to tell signal from noise (it cost a round of triage on #2483).

Environment

• macOS arm64 (Darwin 25.2), 14 threads reported by the runner
• just snapshot-test migration (both flavors), toolchain per rust-toolchain.toml

Happy to re-run with any added instrumentation (timings, per-case retries, runner verbosity) if that helps narrow it — I can reproduce roughly one flaky run per two or three cold full-suite runs.