mirror of
https://github.com/Stirling-Tools/Stirling-PDF.git
synced 2026-09-03 05:10:16 +03:00
Connor Yoh
cabba5a94b
Review caught a real fail-open regression. The multiple-endpoint query keyed on a constant, which asserts the value is caller-independent, but on two paths it is not: the self-hosted-offline check and the legacy ?endpoints= fallback both resolve only the endpoints they were handed. Consumers pass disjoint sets, so whichever mounted first froze a partial map into the shared entry, and a later consumer projected endpoints nobody had checked. Those read as available on exactly the paths where desktop means to fail closed, and being mount-order dependent it would not have reproduced reliably. Keying by the endpoint set restores the old per-consumer semantics and removes the mount-order dependence. Prefix invalidation still covers every set, so the reconnect swap is unaffected. This costs the cross-consumer dedup the PR claimed: that saving only held while the whole-map response made the shared entry complete, which those two paths break. Also stops the readiness effect invalidating on mount, which was forcing a redundant second fetch behind the first.