Add ProjectionCoordinator + distributor layer for hot-cold daemon coo…

[jeremydmiller]

…rdination

Closes #83.

Polecat could spin up the async projection daemon via PolecatDaemonHostedService (single-node, no coordination), but had no path to safely run multiple Polecat nodes against the same database — every node would try to drive every shard. This adds the multi-node-aware coordinator that mirrors Marten's hot-cold distributor design, adapted to SQL Server idioms.

What's added

• SqlServerAppLock — Polecat's analog to Marten's Weasel.Postgresql.AdvisoryLock. Wraps sp_getapplock / sp_releaseapplock with @LockOwner = 'Session', holding a dedicated SqlConnection for the lifetime of the lock owner so the Session-scoped locks persist until we explicitly release or the connection drops. Stale-handle detection clears local tracking when the connection is found broken so the coordinator's HasLock() check stays honest.

• IProjectionDistributor + ProjectionSet — the per-cycle "(database × shards) with a lock id" grouping the coordinator polls.

• Three concrete distributors:

  • SoloProjectionDistributor — single-node fall-through, no locks (test / explicit-opt-out path; not auto-selected by the coordinator).
  • SingleTenantProjectionDistributor — one set per shard, deterministic lock id from db.Identifier:schema:shard.Identity so every node in the deployment races for the same sp_getapplock resource.
  • MultiTenantedProjectionDistributor — one set per database, all shards grouped behind one per-database lock so a tenant's projections never split across nodes.

• ProjectionCoordinator — the concrete IProjectionCoordinator implementation. The ExecuteAsync loop is a slot-for-slot mirror of Marten's ProjectionCoordinator.executeAsync:

  • Build the current distribution
  • For each set: if we hold the lock, ensure agents are running; if not, try to attain it (start agents on success, stop them on failure to recover from a lost lock)
  • Sleep based on agent pause status Picks SingleTenantProjectionDistributor for DatabaseCardinality.Single and MultiTenantedProjectionDistributor for StaticMultiple (mirrors Marten's Tenancy is DefaultTenancy ? Single : MultiTenanted choice). Plus ProjectionCoordinator<T> for ancillary multi-store DI registration.

• DI wiring on PolecatConfigurationExpression: AddProjectionCoordinator(DaemonMode) and the typed AddProjectionCoordinator<T>(DaemonMode). Mutually exclusive with AddAsyncDaemon — pick one.

Tests

src/Polecat.Tests/Daemon/sql_server_app_lock_tests.cs — 6 cases for the lock primitive, all passing against SQL Server 2025 docker on net9.0:

• Only one of two SqlServerAppLock instances acquires a given id
• Releasing lets a waiting instance acquire
• Disposing one auto-releases its session locks (Session-scope)
• Multiple distinct lock ids are independent
• TryAttainLockAsync is idempotent when the id is already owned
• Releasing an unknown id is a no-op

The full coordinator loop (leadership election + agent lifecycle) follows Marten's well-tested implementation slot-for-slot; the lock primitive is the genuinely Polecat-new code path and gets dedicated coverage.