kepr / b305de9f47d1

v1.56.90: release conn_mutex before S3 object reads in the snapshot (koh steal) and offer-by-id (koh load) handlers — split handleSnapshot into resolveSnapshotManifest (DB+disk, runs under the lock) and the now-public collectObjects (S3 fallback, runs unlocked), so a stalled S3 GET can no longer hold the single shared DB connection and wedge every other request behind it. This was the kepr.uk total-hang root cause under concurrent koh offers. musl ReleaseSafe build exit 0 + all 9 handleSnapshot tests green vs throwaway PG (645 pass / 1 pre-existing team-tier live-PG fail)

dev · yesterday · 2026-06-17 · 19.4 MB

session: s3 read-path deadlock fix · agent: claude-opus-4-8

parent: 67eb2de0bd47