Correctness AI check: collection of possible issues and related PRs

[csviri]

An audit of the codebase turned up 15 correctness bugs plus one cosmetic nit.
Each is fixed by its own PR, linked below; the reasoning, impact and (where
present) the regression test live in the individual commit messages.

SpotBugs (effort=Max, threshold=Medium) reports zero findings across all
modules, so none of these are reachable by the existing static analysis.

Confirmed with regression tests

Each of these ships a test that fails without the fix and passes with it.

Area	Bug
ExternalResourceCachingEventSource	Configuring only a generic filter throws NPE on add/delete/update — the branch guard is || but the body dereferences the specific filter. Affects all polling and inbound event sources.
ExternalResourceCachingEventSource	handleRecentResourceUpdate NPEs when the primary is cached but the secondary id is not.
AbstractDependentResource / bulk	A BulkDependentResource that is not a Creator/Updater still gets create/update attempted, failing with ClassCastException. The creatable()/updatable() hooks added for this were never wired up.
LinearRateLimiter	isLimited returns time elapsed instead of time remaining, so rate-limited resources are rescheduled almost immediately (burst of timer events) and then wait roughly a full extra period.
MicrometerMetricsV2	processing.started.latency is registered against a boxed Long that nothing holds a strong reference to, so the gauge always reports NaN.
PollingEventSource	The timer is a non-daemon Timer created in a final field, so it keeps the JVM alive and the event source cannot be restarted (IllegalStateException: Timer already cancelled).
ExternalResourceCachingEventSource	Cache reads are unsynchronized while every mutator is synchronized, and getSecondaryResources looks the primary up twice (TOCTOU NPE).
EventProcessor	When the primary is gone from cache at execution time the submission is never completed: the resource stays flagged as under processing (never reconciled again) and the in-flight metrics gauges drift.
ExecutorServiceManager	The scheduled executor is never shut down; on interrupt the helper pool leaks and started stays true, leaving a "started" operator whose executors are all terminated.

Fix only

Covered by integration tests only, or needing clock/HTTP-layer control that the
unit tests are not set up for. Each commit says which.

Area	Bug
KubernetesResourceFetcher	Optional.orElse eagerly evaluates the cluster-scoped fallback, so every namespaced bounded-cache refresh issues a second, wrong-scope API request.
AbstractExternalDependentResource	selectTargetSecondaryResource calls desired() directly instead of getOrComputeDesired(), breaking the once-per-reconciliation contract.
PrimaryUpdateAndCacheUtils	pollLocalCache builds its deadline from LocalTime, which wraps at midnight — a 10s window each day where the poll times out instantly.
ResourceOperations	addFinalizer's precondition NPEs if the resource is deleted mid-retry; removeFinalizer already guards for this.
Controller	The constructor's null-Metrics guard is bypassed on the last line, which calls the getter again.
EventSourceManager	Cosmetic: a log message has one placeholder too many, rendering a literal {}. (Originally reported as a lost stack trace — that was wrong, SLF4J extracts a trailing Throwable regardless of placeholder count.)
assorted	ManagedInformerEventSource.get can throw on a null last-sync resource version or an unwatched namespace; ExpectationResult.name() throws a bare NPE; an unused parameter and a misleading comment in ResourceOperations.

Notes

• The branches are based on a commit slightly behind main and may need a rebase.
• The PRs are opened as drafts; they reference this issue with "Part of", not a
  closing keyword, so merging one will not close this issue.

Pull requests

• fix: NPE in external resource caching event source when only a generic filter is set #3518
• fix: NPE in handleRecentResourceUpdate for an unknown secondary resource #3519
• fix: honour bulk dependent resource capabilities during reconciliation #3520
• fix: report remaining, not elapsed, time from LinearRateLimiter #3521
• fix: keep a strong reference to the processing started latency gauge #3522
• fix: create the PollingEventSource timer on start and make it a daemon #3523
• fix: read the external resource cache under the event source monitor #3524
• fix: complete the submission when the primary is missing from cache at execution time #3525
• fix: always release executors and reset state when stopping the executor manager #3526
• fix: do not eagerly evaluate the cluster-scoped fallback in KubernetesResourceFetcher #3527
• fix: use the cached desired state when selecting an external secondary resource #3528
• fix: use a monotonic clock for the cache polling timeout #3529
• fix: do not NPE when adding a finalizer to a resource deleted mid-retry #3530
• fix: use the null-guarded metrics field in the Controller constructor #3531
• fix: remove dangling placeholder from event source close warning #3532
• fix: assorted robustness and clarity fixes #3533