resync on broadcast lag and map mqdb conflicts to typed error

[fabracht]

Context

Follow-up to the chorale AWS scale test (10×200 tasks) which surfaced liveness: FAIL against the embedded mqdb backend. Root cause split across two crates:

• mqdb side (already shipped in mqdb-agent 0.8.8 / mqdb-core 0.7.3): Database::update/delete CAS conflicts could stall single-process embedded callers and silently drop writes on the fjall backend. mqdb added bounded internal retry-on-conflict (32 attempts, field-level last-writer-wins) and an atomic fjall commit. No stitch change is required for the stall itself — this PR is the stitch-side hardening + the remaining liveness gap.
• stitch side (this PR).

Changes

• Typed conflict mapping (memory_store.rs): map mqdb Error::Conflict → stitch's existing Error::Conflict { entity, id } on the two CAS paths (update, delete), so the residual backstop conflicts (mqdb cap-exhaustion / vault path) are classifiable via is_conflict() instead of falling into an opaque Mqdb error.
• Resync on broadcast lag (store.rs, config.rs): the internal mutation_loop previously dropped inbound remote deliveries on RecvError::Lagged and continued, causing local divergence + missed scheduler wakeups under high throughput (Finding 2). It now runs an authoritative sync_root_entity_list re-fetch when connected. Default event_channel_capacity raised 1024 → 4096 for high-write scopes.
• No more blanket silent drops (store.rs): handle_remote_mutation replaced .unwrap_or_default() and the let _ = cache-mirror writes with warn-level logging on genuine failures; intentional LWW/scope drops still return Ok(false) quietly.
• Regression test (tests/memory_store.rs): concurrent_updates_to_same_record_converge_without_conflict — 16 concurrent same-key updates on a multi-thread runtime must all return Ok, converge to a last-writer value, and preserve the untouched field (field-level merge).
• Cargo.lock: mqdb-agent 0.8.8 / mqdb-core 0.7.3.

Notes

• Resync is inline in mutation_loop (the background loops hold strong Arc<StoreInner> refs, so spawning would add an Arc cycle) — natural backpressure, nothing new to track.
• sync_root_entity_list recovers root + child entities for the connected user, matching existing reconnect coverage. Top-level wildcard entities aren't re-fetched here — a pre-existing reconnect gap, not introduced by this PR.

Verification

• cargo clippy -p stitch --all-targets clean
• cargo test -p stitch green (91 tests, incl. the new concurrency test)
• End-to-end chorale scale repro (liveness: PASS) needs the AWS Fargate harness and was not run locally.