Add opt-in LISTEN/NOTIFY long polling - #41

Place long-poll notification DDL after the pending 0027-0029 migrations, preserving migration 0028's prune_executions_batch cleanup and companion diff baseline. Update upgrade tests and documentation without stacking the prerequisite PRs.

Include SQLSTATE and the full context chain in DDL retry warnings, test those diagnostics, and document draining writers and retrying ApplyAll when the fixed contention budget expires. Validated the intended predecessor sequence in an isolated source copy. Addresses 41-C1, 41-N3, and 41-N4.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Add a 16+16 dispatcher, 64-client profile with sampled notification-commit and synchronous-replication wait time. Exercise workflow claims and saturated orchestration, activity, and session renewals with verified synchronous-standby and pool-capacity preconditions.

Document notification-bearing operations and reproducible benchmark methodology without publishing performance results from uncontrolled runs. Include timer transaction control in idle SQL counts. Addresses review item 41-N2.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Keep the database message in ordinary migration error display while retaining the SQLx source for SQLSTATE retry classification. Add a red/green regression that also proves failed migration statements leave the ledger, stored function, and durable rows unchanged.

State that migration 0030 must not ship before the 0027-0029 prerequisites, including the dependency of #51 on #32. Addresses review items 41-N6 and 41-N7.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Keep timer scans single-flight, but queue a fresh catch-up for both queues when a reconnect adopts an older snapshot. Run scan transactions READ ONLY and verify rejected writes preserve durable state and connection defaults.

Replace SQL-lock timing gates with controlled acquisition gates, verify successful commits rather than idle backends, and wait through reconnect re-entry before checking the scan count. Add the disconnect-gap regression and benchmark statistics-role preflight/documentation. Addresses 41-N8 and all 41-N9 nits; no performance figures are published.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Keep close-on-drop armed for migration failures, unconfirmed unlocks, and cancellation, but explicitly return the healthy connection to SQLx after success and confirmed advisory-lock release. This leaves short-poll startup with its original warm connection.

Add backend-identity and immediate-idle assertions for migration/no-op/activation success, plus failed-unlock and cancellation safety regressions. Update the migration lifecycle documentation. Addresses review item 41-N10.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Match #51's associated load_migrations and parse_version functions, including the failed-unlock regression added by #41. This keeps the standalone branch buildable and avoids E0599 when the prerequisite changes are combined.

Retain b33a3f2's connection-reuse behavior without changing pool sizing or tuning for the upstream barrier-based test. Addresses review item 41-N11.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Guard migration connections with RAII: unsuccessful exits mark the connection for closure before its field is dropped, while successful migration and confirmed unlock allow SQLx's normal asynchronous pool return.

Keep the same-backend and advisory-lock regressions, remove the immediate-idle assertion, and document asynchronous return. No pool sizing changes or timing-test workarounds.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Keep both delayed items enqueued up front, but spawn the second fetch without asserting that it is still waiting. A slow dispatcher may start that fetch after the second item is already eligible.

Exercise both queues with immediate and deliberately late second fetches. The added late-start case reproduces the old assertion failure and passes after the fix; the shared waiting helper and production code remain unchanged. Addresses review item 41-N12.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>