B1.4 (deferred spike): Evaluate one database-wait evidence shape

[pedrosakuma]

Parent, status and trigger

Parent: #919, Phase 2 / B1.4. Deferred catalog feasibility spike, separate from network waiting and consistent with #681's independently scoped catalog work.

Pick up after B1.2 (#921) and explicit maintainer prioritization. Reuse network-spike findings (#922) when useful, but do not make them a mandatory dependency.

Evidence and question

The four manifests in tests/DotnetDiagnostics.ScenarioEvaluation.Tests/Scenarios/ contain no dedicated database-wait case. Connection-pool guidance in docs/investigation-playbooks.md is not itself evidence of reproducible diagnostic distinguishability.

Can one selected database mechanism be distinguished from generic network or ThreadPool waiting using existing tools?

Bounded scope and acceptance

• Select one provider and one mechanism (for example, pool acquisition waiting OR server-side execution waiting), with an isolated local dependency and healthy control.
• Define which existing outputs can support attribution and which cannot, including explicit dependency prerequisites.
• Bound setup, load, observations, repetitions, retained evidence and teardown.
• Preserve competing generic-wait explanations and credit inconclusion when the collected evidence cannot distinguish them.
• Produce deterministic replay evidence and a bounded live feasibility trial on one topology. Missing database prerequisites are unavailability, never a passing no-op.
• Apply B1 blinding, provenance and human-review conventions.
• Publish GO/NO-GO with a narrow catalog-integration follow-up only for GO. An evidence gap must not silently become a new collector project.

No provider matrix, combined pool/lock/query catalog, external database, SQL optimization benchmark or new MCP tool. Do not require production target instrumentation to make the case pass.

Validation and closure evidence

Reuse scenario manifest/replay tests and the isolated runner, extending them only after feasibility is demonstrated. Preserve provider/topology versions, configuration, actual observations, limitations and failures.