fix(otel): export sampled fallback roots early

[zhongkechen]

Fixes #748.

Related timing requirement: aws/aws-durable-execution-sdk-js#938.

Create and end DurableExecutionRoot at invocation Start whenever a sampled invocation resolves an SDK-owned fallback ancestor. Complete remote parents are untouched; DROP and RECORD_ONLY invocations emit no anchor. The anchor uses its existing ARN-scoped ID and backend start timestamp for both start and end. It is a zero-duration identity span whose descendants can end later; Workflow carries execution duration and terminal outcome.

Only the resolved sampling decision is reused for the anchor. Its SDK attributes are durable.execution.arn and durable.execution.synthetic_root; variable sampler attributes and tracestate are omitted from this repeated identity. Normal durable spans retain the full resolved sampling result. Configured provider resources and processors are preserved. Re-export on later sampled invocations can recover a lost copy, but export/flush failure or termination before flushing does not provide a delivery guarantee.

Each SDK span receives a single-use sampling-intent delivery, consumed before processors run. A processor forwarding parent_context therefore uses its configured sampling policy and normal ancestry instead of replaying private SDK sampling metadata. Implicit-current behavior and public ID-generator scope semantics are unchanged. Repeated post-bind tracer checks are removed while the existing SDK/deferred-provider gate remains. Shared non-null timestamp conversion preserves each plugin's existing None policy.

The branch includes main's approved worker lifecycle and MicroVM worker package. OTel 1.1 requires core 2.1; compatibility validation uses built immutable artifacts. The removed handler_context API and marker are not reintroduced.

Validation of the review follow-up:

• Five focused regressions fail on the original runtime while two implicit-current controls pass; the same seven pass after the minimal fix.
• All 668 workspace OTel tests pass. Installed wheels pass 668 tests plus 12 public version-pair controls on API/SDK 1.45.1 and again on the actual supported minimum API/SDK 1.20.0. The validation environment is restored afterward.
• New coverage includes changing sampler metadata across retry/resume, real processor re-entry, atomic consumption across forwarded contexts, distinct providers and start times, and decorated-handler wait/resume with completed-step suppression. Original provider initialization, status, replay and SDK-key collision controls remain.
• All 49 core/OTel Python source files in the tested wheels match the committed sources. Types (37 OTel files), Hatch lint/format, diff and commit checks pass. The unmodified main MicroVM package passes 154 tests and its 11-file type check.

Current follow-up head: e3c2d41. All current-head whole workflows have succeeded, including package, cloud and OpenTelemetry Conformance Tests run 38430650225. The human Changes Requested review remains for re-review.

[yaythomas]

Thanks for this — the "resolve the root sampling result before descendants" wiring, the ARN-scoped canonical trace ID, and the override of SDK-owned attributes against sampler attributes (test_root_identity_wins_over_sampler_attributes_before_processors) all look correct and well-tested.

Two parity-blockers against JS #966 as merged, one test-shape suggestion, and a few nits/README tweaks. The parity concern is specifically about the shape of the anchor and the sampling intent reaching processors, where JS deliberately picked a narrower contract than the current PR.

I may have mistraced some of this, happy to be corrected on any point! :-)

[yaythomas]

suggestion — shared-trace test uses one provider and one start time

The shared-trace test (two executions with the same propagated trace ID) uses one TracerProvider and the same execution_start_time for both. JS and Java test the same scenario with distinct providers and distinct start times, which is where the ARN-scoped canonical-trace-id logic actually earns its keep (recovery across regions, cold-start resource drift, start-time skew).

Could we add a variant that gives each execution its own provider and start_time?

[zhongkechen]

Added the variant in e3c2d41. Each execution has its own TracerProvider/resource and backend start time while sharing the propagated trace ID. Both views assert the shared trace ID, distinct ARN-scoped anchor IDs, each execution's zero-duration timestamp, and its own provider resource. The existing single-provider and recovery/resource-change cases remain.

[github-actions]

Codex AI review

Found one compatibility blocker affecting the documented OpenTelemetry 1.20 minimum; the added tests do not exercise that lower bound.

Reviewed commit e3c2d419814b35e26c9ff7575b9db236a1c762aa. Workflow run

[yaythomas]

beaut, ty!