DocumentQueryExecutionContextFactory still has a static final CosmosQueryRequestOptionsAccessor (qryOptAccessor) that is initialized during this class’ <clinit> via getCosmosQueryRequestOptionsAccessor(). This reintroduces the class-initialization-time accessor resolution pattern this PR is trying to eliminate and can put initializeAllAccessors() back on a <clinit> path. Please remove this static accessor field and resolve the accessor lazily at the point of use (e.g., inline getter call or a method-local variable).

CosmosQueryRequestOptionsBase still initializes thresholdsAccessor as a static final field by calling getCosmosDiagnosticsThresholdsAccessor() during class initialization. This is the same pattern that can trigger accessor initialization during <clinit> and can reintroduce the deadlock this PR is addressing. Please remove this static field and resolve the accessor lazily (inline call or method-local caching) where it’s needed.

CosmosQueryRequestOptionsImpl still initializes thresholdsAccessor as a static final field by calling getCosmosDiagnosticsThresholdsAccessor() during <clinit>. This keeps accessor resolution on a class-initialization path and can reintroduce the deadlock scenario. Please remove this static field and resolve the accessor lazily at the call sites (optionally cache in a method-local variable).

The JavaDoc for allAccessorClassesMustHaveStaticInitializerBlock refers to ensureClassInitialized() and claims the getter triggers Class.forName() → <clinit> and will fail if static { initialize(); } is missing. In the current ImplementationBridgeHelpers implementation, getters call initializeAllAccessors() (not a targeted Class.forName()), so this test can pass even when a class doesn’t register its accessor during <clinit> (because initializeAllAccessors() can set accessors directly). Please adjust the test (or the infrastructure) so it actually enforces the intended contract.

concurrentAccessorInitializationShouldNotDeadlock is marked as a unit test but forks multiple child JVMs (invocationCount=5 and an inner runs=3, i.e., 15 JVM forks per test execution). This is likely to add substantial runtime/overhead to the unit test suite and increase flakiness on constrained CI agents. Consider moving this to a heavier-weight test group (or reducing the fork count) while keeping the regression coverage.

These inlined accessor calls create very long lines (likely exceeding the 120-character Checkstyle LineLength limit) and are harder to read. Consider storing the accessor in a method-local variable (safe w.r.t. <clinit> deadlock) and/or wrapping the call chain across lines to satisfy the repository’s Checkstyle line-length rule (e.g., eng/lintingconfigs/checkstyle/track2/checkstyle.xml).

This line is extremely long due to nested inline accessor calls and is likely to violate the 120-character Checkstyle LineLength rule. Please wrap this across multiple lines and/or use a method-local variable for the accessors to keep the code readable and compliant.

The endSpan call inlines another getCosmosDiagnosticsContextAccessor() invocation, making the line very long and repeating the accessor lookup. Please store the accessor (or isEmptyCompletion result) in a local variable and/or wrap the call across lines to avoid Checkstyle LineLength violations and reduce repeated volatile reads.

The changelog entry for the <clinit> deadlock fix points to PR #48667, but this PR is described as an alternative approach. Please update the reference to point to the correct PR (this one) or otherwise clarify which change actually delivered the fix in this release line.