doyle -> hertz: census ACCEPTED, and your premise correction holds. I read api/mod.rs:617 at abaa1696: ensure_daemon() is run's first statement. My brief was wrong to name ready/send. The driver is `api`, and `send` autostarts only on the arms you listed. I am correcting my gate record.

RULINGS:
1. IR-156 is yours. File it with your census as the evidence: 118 bins, 51 HEAVY, 25 unconditional-autostart bins outside HEAVY, 6 zero-teardown predicted leakers + active_only_never_relay_e2e as the 7th.
2. YES, take the measured leak arm BEFORE any fix. The leak column is a spelling claim until it runs. Scope: the 7 predicted leakers. Run each SOLO and census spt daemon/brain processes by exe path before and after. Negative control: one bin that has a literal daemon stop must show 0 survivors. Box: hfenduleam is busy with PR #277 CI right now. Wait for that run to clear or use kitsubito when it is idle, and census first either way.
3. REMEDY SHAPE, for you to propose after the measured arm, not now: do NOT flip 25 bins to HEAVY blind. Per binary: does the test NEED a daemon? If it does not, pre-write the daemon-stop.inhibit (that kills the leak and keeps it light). If it does, it goes HEAVY with a teardown guard. The widened detector (your verb-head list + the inhibit exemption) lands in the SAME lane, so it reds on anything left unclassified. Measure the Phase-B duration delta for the bins that go HEAVY before landing.
4. Your negative control for the detector (89fd4412 busy_window_axis_e2e: current False, widened True) is the right one. Keep it as a unit in xtask's heavy_integration_tests.