[beinan]

Summary

A fresh or cached WAL client could fail a payload fetch with Connection refused when its selected voter was unavailable, even though another configured replica held the committed record. Allow read-only Fetch requests to use the existing transport-failure voter fallback, within the original request deadline and byte budget. Mutation replay and unknown-result handling are unchanged; a missing record remains a local read, not a quorum absence proof.

Refs #378.

Testing

• Service-feature release Clippy passed.
• All 44 client tests passed, including exact payload retrieval past unavailable initial/cached voters, the original deadline when all voters are unavailable, and existing mutation no-replay checks.
• Workspace Rust tests: 1,341 passed. Python tests: 228 passed, 11 skipped, 1 expected failure. Rust formatting and workspace Clippy passed.
• Python formatting, lint and Pyright passed with the versions used by GitHub CI. The original local script stopped because locked Ruff 0.14.13 rejected an unchanged base-file import; completing these checks with CI's Ruff 0.17.5 and Pyright 1.1.414 required no source change or repeated tests.
• GitHub CI passed on commit 7092b01cd6bfffad239f6f3546f7d89ec491cda2, including WAL, HA, Lance, coverage, Python and wheel jobs.
• Release builds of the original fault baseline plus this exact fix passed, with all three original service configurations validated. The earlier cloud fault remains failed. Its separate audit verified all 20,480 full values for the observed committed prefixes; negative unknown outcomes remain unresolved without producer fencing. No throughput improvement is claimed.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!