[michaellchung]

Problem

test_env_downlink_limit_makes_a_running_island_slow[forwarder] failed intermittently on CI with no merge with island 1 after the window.

The env downlink window opens at the first PULL for step 400 and lasts 14 s; the check needs a merge with island 1 after window end + 2 s. With a fixed --total-steps 1200 the time left after the window depends on the merge rate: about 7 s at 30 merges/s, but only about 2 s at the ~43 merges/s CI showed (260 merges in 6 s). In that 2 s island 1 must finish draining its forwarder backlog (~100 MB) and get a fresh-base push merged; on a slower runner the run ended first. The syncer behaved as designed (the drained line was logged, island 1 was never removed); the test's step budget was too small.

Shrinking the budget locally reproduces it deterministically: with 1000-1100 total steps every run fails (the run ends inside or right after the window).

Fix

start_syncer takes total_steps; the env test uses 1700 steps. Rounds are at least 20 ms apart (at most 50 merges/s), so the window ends by step 1100 and at least 12 s remain after it. Adds about 10 s to the test.

Verification

• Both env-test variants: 60/60 runs passing (20 in parallel per batch, alongside 24 CPU burners).
• Forwarder variant unloaded (fastest merge rate): 12/12 passing; island 1 merges again at +14.3 s and the run ends at +28.5 s.
• tests/test_slow_downlink.py: 7 passed.

🤖 Generated with Claude Code