Markdown Viewer Agent Skills Files

A write that arrives while a brain folder with the Git durability hook is committing a backlog of Git effects now publishes after the group in flight, inside its 5 s wait. Before, it waited out the whole backlog and came back pending. On Postgres, a first sync of an already-imported source now issues 40 statements per unchanged file instead of 82.

Efficiency wave 9 (GBRA-75). Base is master at wave 8 (310371559). Measured on 4-vCPU AMD EPYC / 16 GiB machines with Bun 1.4.2 and synthetic brains, base and branch interleaved, p50 / p95.

Itemized changes

• Full Git groups yield to a queued write (persistence/effects.ts). A Git group in a folder with the durability hook already stood aside, for at most 20 claims, when a publication was queued or running on its worktree. Groups of 100 never did, though, so under a backlog every group was full. A write's publication found the worktree lock held, was released as writer_busy, and missed every gap until the backlog ended. Every group now yields, in a single update with publication_pending (250 ms). The attempt cap still bounds the delay under a steady stream of writes. Effects stay in claim order per worktree, and no group spans a publication of its own worktree.
• A released claim no longer looks like a stalled preparation (persistence/journal.ts). An uncharged release (writer_busy, owner_unavailable, recovery_required, capacity) now clears the claim phase. A row that waits on the worktree lock used to keep reading preparing / waiting_on=unknown. A charged preparation_deadline release keeps its phase, which kill accounting reads.
• One bounded-read session per waiver run (persistence/bounded-reads.ts withBoundedReadSession, sync-run.ts, sync-waivers.ts, postgres-engine.ts). On Postgres, a waiver run's screens now share one reserved connection and one transaction. SET LOCAL statement_timeout is set again before any read whose bound differs by more than 25 ms, so every read keeps its server-side budget. Before, each bounded read had its own BEGIN; SET LOCAL; read; COMMIT. A failed read rolls back the transaction, and reads it aborted run again in a new transaction. A lost connection, a pool with no long-hold capacity, or PGLite all fall back to the per-read path. Session reads are prepared statements unless the pool is a transaction pooler.
• Tests:
• persistence-git-coalescing-5530.slow.test.ts and its Postgres twin add two cases. A write behind six full groups, sent while one is committing, must publish within the wait with at most the in-flight group ahead of it; this fails on base on both engines. A write that finds the worktree locked is requeued as writer_busy with no claim phase.
• persistence-bounded-read-session.test.ts has 7 cases: sequencing, bound refresh, rollback and retry, joining a run's session, prepared reads, the PGLite and no-permit paths, a real lock wait that ends at its bound with nothing left running, and a re-sync. On the re-sync, base runs 101 bounded transactions for 8 files and the branch runs 16.
• managed-sync-foreground-priority.test.ts (Postgres arm) flake fixed. "a stream of writes from another process" now boots and connects its writer process before the drain starts. Before, the drain's lead depended on how fast a cold process started, so a slow runner could commit most groups before the first write. Forcing a 600 ms boot fails the old test 4 of 4 times and passes the new one 4 of 4.
• Crash robot: 600 s on each engine, every seam, with PgBouncer and pooler_disconnect on Postgres.