[zhongkechen]

Issue #, if available

Fixes #860.

Description of changes

An externally completed callback could resume with terminal state but no
UpdatedOperationIds entry, so its first completion was treated as replay and
the terminal plugin hook could be missing before dependent code ran.

Record the callback operation ID after accepted success, failure, and timeout
transitions, under each path's existing state lock. The existing helper keeps
insertion idempotent. Results, diagnostics, sequence/token behavior, and rejected
completion validation remain unchanged.

Design

• Design note (rule 3): not needed; this corrects existing completion metadata
  through existing public APIs and adds no API or option.
• Simpler options considered (rule 2): reuse _record_updated_operation, as wait,
  retry, and chained-invoke completion already do. A separate delivery ledger or
  changes to plugin hook scheduling are unnecessary.
• Root cause (rule 5): callback completion touched the operation's sequence and
  replaced its state but omitted the existing updated-ID bookkeeping call.
• Net change: three source lines, no new public API members.

Validation

• First commit adds regressions: 12 failed, 6 passed against unmodified main.
  The failures demonstrate missing pending IDs and missing terminal hooks before
  the dependent step. Already-pending-ID controls pass on the baseline.
• After the fix: 18 focused tests passed.
• hatch run dev-testing:test -q --tb=short: 1766 passed. The full
  suite ran with loopback access after the sandbox denied web-runner binds.
• hatch run dev-testing:typecheck: no issues in 137 source files.
• Package hatch fmt --check: passed.
• hatch run dev-testing:js-examples --js-dir ... --no-build: 123 passed,
  zero retries/failures, four files without cloud tests out of 127 selected,
  using the existing build at 3deb8389531d35642e2d4657a69575522dc7a7ea.
• Python 3.13.17, CI-pinned Hatch 1.16.5, and Node.js 22.

Unit tests cover success/failure/timeout, idempotent ID insertion, unknown and
already-terminal rejection, unchanged unrelated callbacks, sequence/token
accounting, both stores, and input construction through both invokers. Public
handler tests cover suspension/resume on memory and filesystem stores, snapshot
visibility, the first terminal hook's non-replay flag, status/result/error, and
delivery before the dependent step. A saved step executes once.

No existing test expectations change.

Assumptions

• Reporting external callback terminal transitions as updated operations is the
  service-parity expectation specified in [Testing]: Record callback terminal completions as updated operations #860 and carried forward from fix(testing): match replay state to the service #795.
  Deployed Lambda qualification is unverified here; tests exercise the local
  runner and persistence paths.
• Timeout integration uses the runner's scheduled callback deadline with real
  time; it does not fabricate a completion hook or operation snapshot.

Decisions not covered by the issue or the design note

• Capture the terminal hooks already delivered when the dependent step begins,
  then assert after execution. This proves ordering without turning an expected
  baseline assertion failure into step retries.
• Stop the regression after first delivery. Later barriers and input
  acknowledgment remain independently scoped to [Testing]: Acknowledge invocation input without losing pending updates #861; empty errors remain [Testing]: Preserve empty callback error history without changing runtime errors #862.

Checklist

• Every new branch has a test that fails without it (rule 4).
• The first commit adds failing regressions (rule 5).
• No existing test expectations were changed.

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

No actionable correctness or approach findings after reviewing the complete diff against CONTRIBUTING rules 1–9. Tests were not executed during this read-only review; deployed Lambda parity remains explicitly unverified in the PR.

Reviewed commit 18e88c2c2061cd3238992717925dcbca9fe3205f. Workflow run