feat(plugin): prepare propagation metadata

[zhongkechen]

Refs #751. Coordinated groundwork: aws/aws-durable-execution-sdk-js#957 and aws/aws-durable-execution-sdk-java#764.

This model-blocked draft prepares the SDK-owned propagation contract, synchronous collector and pure OTel producers. It does not transmit metadata or implement the full chained-invoke feature. Public Lambda models currently lack ChainedInvokeOptions.XAmznTraceId and DistributedMapOptions; production invoke START and wire serialization remain unchanged.

The core contract adds frozen PropagationInput (execution ARN, operation ID, optional parent operation ID, target function name) and frozen PropagationMetadata with optional x_amzn_trace_id. The optional synchronous hook defaults to no contribution. Collection is ordered and first-non-null wins; equal values do not conflict, while different later values warn with both plugin identities and a running conflict count. Ordinary hook/getter/result/diagnostic failures are isolated, accidental coroutine returns are closed and rejected, and existing BaseException cancellation/control semantics remain unchanged.

Both OTel views encode canonical trace identity, the actual operation span ID (including an invocation-view continuation), and the resolved sampled/unsampled decision. Before span creation they derive the existing deterministic logical operation ID. Producers create no span solely for an ID, reject inactive/mismatched ownership and cannot return stale metadata after failed setup. Tracer/provider/resource ownership is unchanged.

Backward compatibility is additive: existing supported core ranges and provider API version 1 remain unchanged. OTel modules do not eagerly import missing new core types. On older cores without the contract, existing plugin loading/tracing continues and the new producer returns no contribution. A coordinated capable core is required only for the new typed metadata/collector capability. Runtime type introspection remains safe in either case, and SDK-owned types are used when available.

Validation: 1,716 core tests and 267 OTel tests pass; both package type and formatter/linter checks pass, as does git diff --check. New unit/component cases cover immutable contracts, optional-hook compatibility, merge/conflict diagnostics, synchronous/error/cancellation behavior and real OTel exporters. Root/Parent/Sampled match operation contexts without extra spans or ambient mutation. Actual installed core 2.0.0 plus this OTel wheel passes 14 valid-registration, wait/resume, success/failure and constructor-correlation cases. An OTel-only layer over separately installed core 2.0.1 passes the same 14 cases and pip check, with no core package in the layer; import paths and type introspection are verified. No AWS calls were needed.

Before this can become a complete feature:

1. Supported public client models and generated serialization for the missing fields.
2. Backend capability rollout.
3. Consume the hook on real invoke START, with replay and failed-checkpoint integration coverage.
4. Deployed topology/propagation validation.

Integration prerequisite: rebase onto the coordinated minor fixes #752/#753/#754, retaining exclusivity metadata and compatible registration cleanup. Version/release coordination must preserve existing valid combinations and advertise the new capability separately; do not raise the broad core minimum merely to force rejection of old installations. Keep this PR draft while model/backend dependencies block completion. The full feature issue remains open; no unsupported wire members, serializer bypasses or new fanout/HTTP APIs are included.