Give find and resolve coherent configuration snapshots without lock-held discovery #536

Tracking plan: #528
Priority: P2. Evidence: source-confirmed ownership inconsistency; add deterministic race tests before claiming a reproduced user failure.

Problem

Refresh uses a transient locator graph and a configuration-generation snapshot. Find/resolve still use the long-lived shared graph while configure mutates its locators one by one before publishing the new generation. Those operations can therefore observe partially applied configuration even though refresh has a coherent boundary.

Additionally, find passes a chained configuration read/clone expression directly into discovery; the temporary read guard can live through that call, retaining the configuration lock during filesystem work.

Sources: configure prepare/mutate/publish, resolve, find, refresh-state contract.

Scope

• First make find's configuration snapshot an explicit local value and release its guard before discovery.
• Define one coherent request-snapshot boundary for configure/refresh/find/resolve. Prefer constructing the replacement configuration/locator graph off-lock and atomically publishing it, rather than extending broad locks across I/O.
• Separate deliberately shared performance caches from configured input and correctness-critical discovery state. Preserve current generation-gated notifications, locator ordering, scoped refresh-state sync, and coalescing semantics.
• Replace unnecessary mutable-graph rollback complexity only where the new ownership model demonstrably removes the failure mode; do not bundle unrelated rewrites.

Acceptance criteria

• Barrier/channel-driven tests prove find and resolve see either the old or new configuration, never a mix across locators.
• Slow discovery/configuration fixtures do not hold a global configuration lock over filesystem or subprocess I/O.
• Concurrent configures publish in a defined order; failures leave the previous usable snapshot intact.
• Full/workspace/kind-filtered refresh sync, stale notification suppression, joined refresh replies, and manager fidelity retain existing behavior.
• In-flight cache invalidation/clear and configuration changes have explicit tested semantics.
• No regression in client-observed latency or long-lived cache behavior, and locator-state documentation includes the actual lifetime of every locator's mutable state.

Dependencies and prior work

Use #531 metrics and the #533 long-lived/concurrency workload before broad ownership changes. This establishes the ownership foundation for #539 and #540; the complete order is in #528. #385 fixed configure versus refresh isolation; #461 removed locator I/O from the configuration write lock. This issue extends consistency to find/resolve and must not regress either fix.