feat(otel): Parent Workflow span onto shared execution trace - #647

Codex AI review

Found three issues affecting sampling policy, RECORD_ONLY propagation, and public API compatibility. Static inspection only, as requested.

Reviewed commit 8a88e75b8b72855f5aa3ab9ff6bea9888b3af08b. Workflow run

Claude AI review

Review: feat(otel) — Parent Workflow span onto shared execution trace (#647)

Reviewed the full SHA-anchored diff (27 files, +2473/−270) against the base revision, focusing on the trace-topology change, sampling precedence, cross-classloader bridging, checkpoint/replay semantics, and the local test runner.

No blocking issues found. The change is coherent and unusually well-tested. Key correctness properties I verified:

• Trace stability across reinvocations. ExecutionTraceContext.canonicalTraceId/resolve derive a stable ancestor (valid remote parent → else ARN/start-time synthetic root) and deliberately never adopt the per-invocation ambient span, so the execution keeps one trace ID. LocalDurableTestRunner now pins executionName/executionOperationId and the ARN's last segment across run() calls, matching that invariant.
• Sampling is resolved once and applied consistently. resolveSamplingResult follows explicit upstream Sampled → same-trace ambient bit → configured sampler (once) → documented default-sampled; DurableSampler returns the precomputed SamplingResult verbatim (preserving RECORD_ONLY) and the execution ancestor's flags are derived from the same decision, so a parent-based sampler can't silently drop the tree.
• Cross-classloader decision bridge is scoped correctly. The thread-ID-keyed system property is opened/closed synchronously around each startSpan() (sampler runs on the same thread), with prior-value restore and test cleanup — no leak or cross-thread bleed.
• Replay/large-result path is safe. The stored EXECUTION operation shares the fresh op's stable id, so the merge collapses to one op (guarded against >1) with no duplicate-key collision; covered by the new 7MB replay test. Continuation/replay operation spans link back to the deterministic initial op id in [operation, Workflow] order.
• Validity hardening. ExtractedContext tri-state sampling with null→UNDECIDED normalization (incl. Jackson legacy deserialization), and TraceId.isValid/SpanId.isValid rejection of all-zero Root/Parent, are correctly threaded through resolution.

Residual risk (not defects): The nullable-traceId branch in ExecutionOtelPlugin.resolveParentContext is currently unreachable only because executionTrace is set before tracingEnabled and never reset; if that ordering invariant ever changes, SpanContext.create would receive a null trace ID. Consider a defensive fallback to the canonical/derived trace ID there for robustness. Otherwise, verify the OTel conformance workflow (pinned to a new ref in this PR) passes against the backend-parented topology assertions, since that behavior isn't exercised by the in-repo unit/integration suite.

Reviewed commit 8a88e75b8b72855f5aa3ab9ff6bea9888b3af08b. Workflow run