verify chorale liveness: PASS on AWS Fargate after resync-on-lag fix

[fabracht]

Context

stitch 0.2.0 shipped via PR #1 (squash commit 5169cce) with the resync-on-broadcast-lag fix, typed mqdb conflict mapping, and event_channel_capacity 1024→4096.

The original bug surfaced as liveness: FAIL in the chorale AWS scale test (10×200 tasks) against the embedded mqdb backend. That end-to-end repro is the acceptance gate for this fix and was not run before merge — it requires the AWS Fargate harness, which isn't available locally.

What's verified

• Mechanism proven with in-process regression tests in crates/stitch/src/store.rs (store::tests): lag → resync, connection-status gate, sustained-overflow re-lag.
• Concurrent same-key convergence test in crates/stitch/tests/memory_store.rs.
• Full cargo test -p stitch (94 tests) + cargo clippy --workspace --all-targets green on 5169cce.

What's NOT verified

• The actual chorale scale repro (liveness: PASS) on AWS Fargate.

Acceptance

• Run the chorale 10×200 scale test on the AWS Fargate harness against main @ 5169cce (or later).
• Confirm liveness: PASS.
• If still failing, capture logs (look for mutation receiver lagged; resyncing from server and any residual Error::Conflict) and reopen investigation.

Note

There is no CI on this repo (no .github/workflows), so main has no automated regression gate for this path beyond the unit/regression tests above.

[fabracht]

Round-2 split:

• Fix A (stitch) — cache mirror dropped out-of-order Update-before-Insert deliveries (surfaced as Error::NotFound), diverging the local cache and stalling liveness at ~23/2000. Fixed in PR upsert out-of-order remote updates in cache mirror #3 (upsert on NotFound; bumped to 0.2.1). Covered by store::tests::mirror_*.
• Fix B (mqdb) — pending_sync index entries surviving row deletes (index pointed to non-existent entity WARNs). Lives in the mqdb repo (fjall_backend.rs / query.rs), not here. Still pending.

The chorale AWS liveness repro (this issue) needs both A and B landed before re-running, so keeping this open until Fix B ships.

[fabracht]

Both round-2 fixes have shipped.

• Fix A (stitch) — merged to main as ac87c8a (squash of PR upsert out-of-order remote updates in cache mirror #3), stitch 0.2.1. Out-of-order Update-before-Insert deliveries now upsert the cache instead of dropping.
• Fix B (mqdb) — shipped as tag v0.8.10 (e03075e, self-heal of stale pending_sync index entries in list_from_index_ids). stitch builds green against it (97 tests).

Repro is unblocked, with one environment requirement: stitch consumes mqdb as a path dependency, so Cargo.lock cannot enforce the mqdb version — the harness must have its MQDB checkout at v0.8.10 (e03075e) or it will build against whatever mqdb is on disk. (Note: the v0.8.10 tag left mqdb-agent's crate version at 0.8.9, so the lockfile reads 0.8.9; that's expected.)

Next: run the chorale 10×200 (and 2000-task) scale test on AWS Fargate against main @ ac87c8a with MQDB @ v0.8.10. Keeping this open until liveness: PASS / tasks_terminal: 2000 is confirmed.