[Tyrie Vella (tyrielv)]

Problem and Context

Mount runs credential retrieval and local Git config writes in parallel. On Windows, a config replacement can make a concurrent reader report Permission denied.

git credential fill reads Git config before it returns credentials. VFS for Git treats a transient config-read failure as a credential failure. A mount with a cache server can then continue without the server config.

A timed process wait can also return before redirected output callbacks finish. The retry classifier can then miss the config error, or a successful fill can lose its credential output.

PR #2130 retries explicit Git config reads. This change covers credential fill, which also reads config. It does not depend on #2130.

Changes

• Retry credential fill only for a failed, untruncated result with empty stdout. Stderr must contain only the supported English Git config-read diagnostics. The access warning must report Permission denied or Device or resource busy.
• Share the fill retry path between repository credentials and certificate passwords. Make at most five attempts, with delays of 20, 40, 80, and 160 milliseconds.
• Deduct elapsed time from the process timeout before each retry. Stop retries when the remaining budget cannot cover the delay.
• Preserve successful fills, missing-password results, other credential failures, and the final error message. Do not retry credential approval or rejection.
• Drain redirected stdout and stderr after Git exits, before the shared process runner returns the result. Leave the timeout-failure path unchanged.
• Emit CredentialConfigReadFailure for each matching failed attempt, including exhaustion. WillRetry distinguishes a retry from the final failure. Use RetryDelayMilliseconds for the delay.
• Keep config paths, URLs, credential values, and stderr out of retry metadata. Keep stderr in the local log.
• Add unit tests for recovery, exhaustion, non-retryable failures, truncation, timeout budgets, and both fill callers.
• Add unit tests with real Git and an exclusive config-file lock, for both finite and infinite timeouts. The tests release the lock after the first failure or retain it through exhaustion. They use a local helper with test credentials.
• Add unit tests for output completion after process exit. The tests pause a real Git output callback and verify that timed and infinite waits await its completion.

[Keith Klein (KeithIsSleeping)]

[KeithBot]: ⚠️ 1 issue · iteration 1 (d9b1b9a)

1. 🟡 Medium Config-read retry skipped when timed WaitForExit returns before stderr

Validation

• ❌ PR credential fixtures (50 tests)
• ❌ Locked-fill probe, 30 s timeout
• ✅ Build: unit tests, natives, Mount
• ✅ Full unit suite at source + base
• ✅ Real-git race + loopback HTTP auth
• ✅ Mutants and base-code red control
• ✅ CI checks on head commit
• ⏭️ Functional tests and installer — not run

_{🤖 KeithBot · claude-opus-5.5, gpt-6.1-sol, gemini-4-argon · reply on a thread to dispute}

[Keith Klein (KeithIsSleeping)]

[KeithBot]: ✅ Resolved · iteration 2 (baabb71)

• Config-read retry skipped when timed WaitForExit returns before stderr: fixed
• New changes since iteration 1: no new issues

Validation

• ⚠️ Stderr-holding helper, base vs source — partial
• ✅ PR fixtures and auth tests at source
• ✅ Locked first fill, 16-way, 30 s and -1
• ✅ Full unit suite (r1, r3 runs)
• ✅ CI checks on head commit
• ⏭️ Functional tests and installer — not run

Approving.

_{🤖 KeithBot · claude-opus-5.5, gpt-6.1-sol, gemini-4-argon · reply on a thread to dispute}

[Nathan Baird (ShiningMassXAcc)]

Reviewed the current head: the previous concerns are addressed and current-head validation passed.