fix(db): fail closed when WAL valve cannot checkpoint past caps (#1539)

[colbymchenry]

Summary

• Root cause (code path): sync() armed WalCheckpointValve but never called backpressure(), so the hard/file caps were only enforced on indexAll. Daemon catch-up (with the query pool holding read marks) could therefore grow codegraph.db-wal without a bound — matching the 64 GiB incident on a Linux-kernel-sized repo.
• Fail-closed: after parked PASSIVE backfills cannot progress while still past the documented hard/file caps, throw WalValveAbortError instead of the old futility latch that disabled parking for 60s and allowed unbounded growth.
• Wire backpressure into orchestrator.sync (per changed file) and both resolveReferencesBatched call sites in CodeGraph.sync.

Relation to #1583

#1583 also wires sync backpressure (among many other memory/traversal bounds) but is CONFLICTING against main and mixes unrelated changes. This PR is the focused #1539 containment + fail-closed behavior the issue asked for. #1583's memory work remains separately useful and is not superseded.

Test plan

• npx vitest run __tests__/wal-deferral.test.ts — 24/24 under Node 22 (node:sqlite)
• New unit tests: abort when checkpoints stay busy over file cap; abort when machinery returns null over cap; soft foldNow under caps does not abort; sync invokes backpressure on changed-file store and orphan resolution
• tsc --noEmit clean

Refs #1539