Report whether durability/projection affinity engaged (GH-3785)

[jeremydmiller]

The GH-3785 cross-family affinity join (#3805) is deliberately fail-silent — a durability agent that cannot be matched to its database's projection owner falls back to the even spread, because a miss is never wrong, only not-better. That is the right runtime behavior and an unusable diagnostic story: a join that never fires because the two descriptor pipelines spell the same database differently looks exactly like the feature working, minus the benefit.

Short of joining pg_stat_activity against the assignment table on a production cluster, there was no way to tell the two apart. That matters right now: #3785 is one confirming measurement away from closing, and the measurement it is waiting on is precisely "did the join engage against this cluster's 512 databases and its particular server naming."

What this adds

DurabilityAffinityPreference wraps the preference function and counts both outcomes:

KnownDatabases	databases with event-subscription agents in the grid — something to follow
Considered	durability agents whose URI actually names a database (the null store and the composite marker are not databases, and would otherwise read as permanent misses)
Matched	of those, how many resolved to their database's projection owner

Logged only when the numbers move. Assignment is re-evaluated on every health check, and a settled cluster reports the same figures forever — an unconditional line would be noise at exactly the cadence that makes it unreadable. A change means a deploy, a rebalance, or a database arriving or leaving.

Three outcomes, three behaviors:

• Matched > 0 → Information, once: co-located {Matched} of {Considered} durability agents ... across {KnownDatabases} databases
• Matched == 0 with KnownDatabases and Considered both non-zero → Warning. There are projection agents, there are database-bearing durability agents, and not one joined. On a multi-database store that is a spelling divergence, not a coincidence — and it names the consequence (each of those databases attracts two nodes' connection pools instead of one).
• KnownDatabases == 0 → silent. An application with no projections has nothing to co-locate with; that is not a finding.

Testing

7 new CoreTests, each mutation-tested against a matching defect — counting non-database URIs as considered, dropping the change gate, not escalating the fail-silent case, warning when there was nothing to follow. All four mutants killed, one test each, no overlap.

The real-stores Marten test now also asserts every database matched. That test exists because the join spans two descriptor pipelines (Weasel vs Marten), so it is the right place to pin the counters — it is the only test where a spelling divergence could actually occur.

• CoreTests.Runtime.Agents — 293/293
• durability_projection_affinity_real_stores — passes against real Postgres tenant databases, 3 of 3 matched
• Full wolverine.slnx Release build clean

Behavior is unchanged: this is counting and logging only.

Related: #3785, #3805

🤖 Generated with Claude Code

https://claude.ai/code/session_01WHAuhdWS3XeAk16swV9G8m