fix(workflow): stop a re-emitted output echo from reordering the replay sequence

[chelsealong]

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

• Closes: [Bug] ADK 2.8.0: repeated HITL deadlocks on replay within one invocation (no LLM) #7027

Problem:

On a resumable Workflow running the documented review/revise HITL loop
(review → revise → review → ... → approve), a third resume can deadlock with:

RuntimeError: Replay divergence detected: Timed out waiting for sequence key 'HITL@1' to be unblocked.

Root cause: Workflow._maybe_reemit_replayed_output re-surfaces a
fast-forwarded node's output again on every turn it is replayed through, so a
resumable event stream stays complete for that turn. That re-emitted event
carries the same node path / run id (HITL@1) as the node's original,
genuine completion.

ReplayManager._scan_sequence builds the replay barrier's expected
completion order by scanning terminal events and, for a segment it has
already seen, removing the old position and appending the new one — logic
meant to update a node's position when it moves from "interrupted" to
"completed". Applied to the re-emitted echo, it instead moves HITL@1 past
Revise@1, even though Revise@1 can only ever run after HITL@1
completes. On the next resume this produces a corrupted barrier sequence
([Revise@1, HITL@1, ...]) that creates a real cycle: HITL@1's
fast-forward path waits for Revise@1's turn, but Revise@1 can't even be
scheduled until HITL@1's (fast-forwarded) task returns — a deadlock that
only times out after 15s.

Solution:

_scan_sequence now tracks which child run ids have already reached a
genuine terminal completion (an event carrying output, route, or
error_code). Once a run id is marked completed, any further terminal event
for that same id is ignored for ordering purposes, since a completed run id
cannot complete a second time — a repeat is always the re-emitted echo, not
new information. This preserves the legitimate "interrupted → completed"
reposition (the first time a node moves from pending to done) while making
the re-emitted echo a no-op for sequencing.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Added test_scan_workflow_events_sequence_ignores_reemitted_completion_echo
in tests/unittests/workflow/utils/test_replay_manager.py, which reproduces
the exact echo pattern from the issue (hitl@1 completes, revise@1
completes, then hitl@1's output is echoed again) and asserts the sequence
stays in the causally-correct order ["hitl@1", "revise@1"].

Verified the new test fails without the fix (git checkout HEAD~1 -- src/google/adk/workflow/utils/_replay_manager.py, rerun, restore):

$ python -m pytest tests/unittests/workflow/utils/test_replay_manager.py -k reemit -q
FAILED ...::test_scan_workflow_events_sequence_ignores_reemitted_completion_echo
AssertionError: assert ['revise@1', 'hitl@1'] == ['hitl@1', 'revise@1']

With the fix:

$ python -m pytest tests/unittests/workflow/utils/test_replay_manager.py -q
....................
20 passed in 1.10s

Full workflow suite:

$ python -m pytest tests/unittests/workflow -q
797 passed, 1 skipped, 5 xfailed, 44 warnings in 12.77s

Full unit test suite (unaffected):

$ python -m pytest tests/unittests -q -n auto
13940 passed, 85 skipped, 27 xfailed, 2 xpassed, 2176 warnings, 24 subtests passed in 220.78s

Manual End-to-End (E2E) Test:

Ran the standalone repro script from the issue (unmodified, google-adk from
this checkout) end to end:

Before the fix — third resume raises the reported RuntimeError and
Publish never runs:

{"phase": 0, "requests": ["review_0"]}
{"phase": 1, "requests": ["review_1"]}
{"phase": 2, "requests": ["review_2"]}
{"exception_type": "RuntimeError", "exception": "Replay divergence detected: Timed out waiting for sequence key 'HITL@1' to be unblocked.", "published": 0}

After the fix — all three reviews are processed and Publish runs exactly
once:

{"phase": 0, "requests": ["review_0"]}
{"phase": 1, "requests": ["review_1"]}
{"phase": 2, "requests": ["review_2"]}
{"phase": 3, "requests": []}

(The script's final assert published == [True] passes; no exception is
raised.)

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

This PR was prepared with AI assistance (Claude Code), including root-cause
analysis via source inspection, the fix, the added test, and this
description. All changes were verified locally as described above before
submission.