feat(plugin): add execution_arn to OperationInfo hooks

[wangyb-A]

Summary

Adds an additive execution_arn: str | None field to the shared OperationInfo
dataclass so every operation and user-function plugin hook, and the operation
maps/snapshots on the invocation hooks, can attribute an operation to its
enclosing durable execution without relying on dictionary insertion order.

This is a prerequisite API improvement for the Python Workflow Insight plugin
(PR #632), which needs a reliable per-operation execution ARN.

Design & compatibility

• New field is keyword-only, default=None, repr=False, compare=False,
  hash=False, experimental metadata — mirroring the additive-field pattern
  established in merged PR feat(plugin): operation maps on invocation hooks #629. Existing infos keep comparing, hashing, and
  reprinting exactly as before, and callers constructing OperationInfo without
  an ARN are unchanged.
• Inherited by OperationStartInfo, OperationEndInfo, UserFunctionStartInfo,
  and UserFunctionEndInfo.
• PluginExecutor stamps the active invocation's ARN (from the captured
  InvocationStartInfo) onto every direct hook: on_operation_action (START),
  on_operation_replay, terminal on_operation_update, on_child_context_end,
  on_user_function_start, and the derived user-function end
  (UserFunctionEndInfo.from_start_info preserves it).
• The OperationInfo entries in InvocationInfo.operations,
  InvocationStartInfo.updated_operations, and OperationChangeInfo
operations/updated_operations carry the ARN too. It is threaded explicitly
  through OperationInfo.from_operation and _to_operation_info_map — never
  inferred from map ordering.
• Before an invocation starts, the ARN is None (no active invocation to
  attribute to).

Tests

Added focused unit tests in plugin_test.py:

• field is keyword-only, experimental, and excluded from repr/eq/hash on the base
  class and all subclasses;
• repr/equality/hash are unchanged by the ARN;
• from_operation accepts and propagates the ARN (and defaults to None);
• every direct hook (on_operation_action, on_operation_replay,
  on_operation_update, on_child_context_end, on_user_function_start,
  user-function end) receives the active ARN;
• an A/B/A regression proving hooks follow whichever invocation is currently
  active;
• invocation start/end and OperationChangeInfo map entries all carry the ARN;
• legacy constructors without an ARN still work.

Validation

• hatch run dev-core:test — 1640 passed
• hatch run dev-core:typecheck (mypy) — Success, no issues in 77 source files
• hatch fmt --check (ruff check + format) — all checks passed, 77 files formatted

Draft — prerequisite for Workflow Insight PR #632. Not for merge until reviewed.

[github-actions]

Codex AI review

[P2] Add end-to-end coverage for ARN propagation

These tests call PluginExecutor directly, so they do not verify propagation through durable_execution, ExecutionState, background checkpointing, and replay. Repository rules require e2e tests for public API and cross-component changes. Add a tests/e2e/ suspend/replay test asserting the invocation ARN on lifecycle hooks and operation-map entries.

[github-actions]

Codex AI review

The implementation is internally consistent, but the new public plugin field lacks required end-to-end coverage.

Reviewed commit 22f5fb059f1f5d799263dde82077acc86d54d069. Workflow run

[github-actions]

Claude AI review

No actionable findings. This is a clean, well-scoped additive change that adds an experimental execution_arn: str | None field to OperationInfo and threads it through PluginExecutor's hooks and operation-map snapshots.

Correctness verified:

• The field is kw_only=True, default=None, repr=False, compare=False, hash=False, so equality, hashing, repr, and all existing positional/keyword construction are unchanged. Pre-existing field-enumeration tests match on input/output/result/error name substrings, which execution_arn does not, so they remain green.
• Subclass dataclass ordering stays valid: UserFunctionEndInfo's non-defaulted fields still follow only the parent's non-defaulted positional fields because every defaulted parent field (including the new one) is kw_only.
• Every OperationInfo/subclass emission site is covered — on_operation_action (START), on_operation_replay, terminal on_operation_update end, on_child_context_end, on_user_function_start, UserFunctionEndInfo.from_start_info, and the maps on on_invocation_start/on_invocation_end/OperationChangeInfo. No site was missed.
• _active_execution_arn correctly returns None before an invocation and follows re-entrant on_invocation_start (validated by the A/B/A test). Where OperationChangeInfo reads self._invocation_status.execution_arn directly, the preceding None guard makes it equivalent to _active_execution_arn, so stamping is consistent within a single on_operation_update call.
• No cross-package compatibility risk: only plugin.py constructs these infos; the OTel/testing/examples packages consume them, and a defaulted keyword-only field cannot break existing construction.

The added unit tests are comprehensive (field additivity/experimental metadata, keyword-only enforcement, per-hook ARN stamping, A/B/A regression, and operation-map coverage).

Residual test risk (minor, non-blocking): coverage is unit-level only. Since this adds a public API field, an e2e test in tests/e2e/ asserting the ARN propagates end-to-end through a real durable_execution() run would harden the guarantee, but the existing focused unit tests adequately cover the changed behavior for a draft/experimental field.

Reviewed commit 22f5fb059f1f5d799263dde82077acc86d54d069. Workflow run