fix(labs): persist Antigravity resume state when a turn ends early

[LHMQ878]

What's wrong

AntigravityAgent._run_async_impl renamed the harness trajectory and wrote the resume step index after the async with block, so neither happened unless the turn ran to the end of the step stream. Two ordinary paths skip it.

A caller that stops reading. The runner wraps agent generators in contextlib.aclosing (runners.py:1415), so a disconnected SSE client or a cancelled run closes this generator at a yield. __aexit__ still runs and the harness still flushes traj-<random>, but the statements after the with do not.

A harness error mid-stream. Same skip, and here the events emitted before the failure are already in the session.

Reproduced

Driving _run_async_impl with a stubbed SDK agent, on 735402d:

--- A: caller breaks after the first event ---
  files:                      ['traj-harness-random']
  renamed to traj-<cid>?      False
  resume index written?       False
  next turn would resume?     False        <-- starts a fresh conversation

--- B: harness raises after emitting step 2 (resumed turn, index was 1) ---
  events emitted to session:  ['new']
  resume index on disk:       1            <-- step 2 replays next turn

So in A the conversation is silently abandoned: the ADK session holds the turn's events, the harness is told nothing about it next time, and traj-harness-random is orphaned in save_dir permanently. In B step 2 gets recorded twice.

The fix

Move the rename and the index write into a finally, and read active_agent.conversation_id before the step loop rather than after it, so it is available when the loop does not finish.

Both writes are plain synchronous file operations, which is what makes them safe in a finally that may be running under GeneratorExit — nothing is awaited and nothing is yielded there.

The condition is resumed or harness_conversation_id, so a turn that never reached the harness at all (say Agent(config) raises) still writes nothing: there is no trajectory to rename and no step worth remembering, and a stray traj-<cid>.resume would describe a conversation that never began.

Tests

Three added to tests/unittests/labs/antigravity/test_antigravity_agent.py:

test	without the fix
test_trajectory_is_claimed_when_the_caller_stops_reading	fails — no rename, no index
test_resume_index_survives_a_harness_error_mid_turn	fails — index stays 1
test_a_turn_that_never_reaches_the_harness_records_nothing	passes

The third passes on main by construction — it is a guard against this change writing state for a conversation that does not exist, not a regression test for the bug. Calling that out rather than presenting three red-to-green tests.

Verification

• pytest tests/unittests/labs/antigravity/ → 50 passed (was 47 passed)
• without the src change, the same run is 2 failed, 48 passed
• pytest tests/unittests/labs/ tests/unittests/agents/ → 986 passed, 2 xfailed, no failures
• pyink --check → 2 files would be left unchanged; isort --check-only → exit 0

Deliberately not in this PR

_trajectory_files.save_resume_step_index truncates and then writes, so an interruption during the write leaves an empty file. load_resume_step_index treats that as -1, and with the trajectory still present the whole conversation then replays. Putting the write in a finally makes that window slightly more reachable, so it is worth a temp-file-plus-os.replace follow-up — but it needs a crash to trigger and I have no test that demonstrates it, so I left it out rather than mixing a shown bug with a reasoned one. Happy to add it here if you would rather have both together.

Follows on from #6040.