feat(otel): add durable sampling coordination

[ayushiahjolia]

Issue #, if available: #674

Description of changes:

• Second PR in the stack for reparenting durable OTel spans onto one shared execution trace.
• Adds the sampling machinery that lets a single decision be resolved once per invocation and applied to every durable span. Not yet wired into the plugins, so behavior is unchanged.
• Next in stack: wire both plugins onto the shared execution trace and sampling.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Codex AI review

One resource-lifetime issue found. Residual risk: OTel 1.20 compatibility is simulated rather than tested against that version.

Reviewed commit daabf76aeb9dc838dfed7d8b77a04d73d817fb21. Workflow run

[github-actions]

Claude AI review

This PR adds a self-contained OTel sampling-coordination module (durable_sampling.py) plus a thorough unit test suite. Per the description it is not yet wired into the plugins, so runtime behavior is unchanged.

I reviewed the sampling decision logic, the OTel version-compatibility shim, trace-state propagation, attribute merging, context-key storage, and the tests. The implementation is careful and well-covered:

• resolve_sampling_result correctly prioritizes an explicit backend decision, then a same-trace ambient span, then delegates to the configured sampler with a fresh root context. trace_state is preserved only for same-trace ambient spans (matching the documented contract), and preserving it on DROP is intentional and correct for downstream propagation.
• _delegate_accepts_trace_state / _delegate_should_sample use a signature probe instead of a call-time try/except, correctly avoiding masking a TypeError from a sampler body, and handle **kwargs-style and pre-1.21 (links-terminated) signatures. All branches are tested, including the raising-body case.
• DurableSampler merges span/intent attributes (intent wins), preserves the intent's trace_state, and is idempotent on install. install_on_tracer mirrors the existing DeterministicIdGenerator.install_on_tracer pattern and correctly uses the public SDK Tracer.sampler attribute.

No blocking findings. No correctness, determinism, serialization, or concurrency defects warranting an inline comment were confirmed.

Residual test risk / minor, non-blocking observations:

• The module is exercised only in isolation; plugin integration (parent-context intent storage, and consistency between resolve_sampling_result's UNDECIDED-on-canonical-trace path and execution_trace_context._trace_flags's root_sampled() path) is deferred to the next stacked PR and not yet integration-tested. See packages/aws-durable-execution-sdk-python-otel/src/aws_durable_execution_sdk_python_otel/durable_sampling.py:127.
• _delegate_accepts_trace_state caches on the delegate's bound method via functools.lru_cache(maxsize=None), holding a strong reference to each distinct sampler for process lifetime and raising TypeError for an unhashable delegate; all standard OTel samplers are hashable, so this is theoretical (durable_sampling.py:152).
• test_durable_sampler_merges_span_and_intent_attributes asserts the intent wins on a shared key but does not assert span-only keys survive the merge.

Reviewed commit daabf76aeb9dc838dfed7d8b77a04d73d817fb21. Workflow run

[ayushiahjolia]

/ai review