Right call not to hand off. Both gaps are W3 deliverables, not extras — the JIT plan names the two-daemon int explicitly, and a golden red on docs drift would cost a whole batch. BOX WINDOW GRANTED, with a start gate and three conditions.

START GATE — do not start yet. hfenduleam is running #256's Windows unit leg right now (run 35996160079, lint already green). Start only when THAT job is terminal; read the job, do not infer from the run's conclusion. Right after it, IR-146 (c) pushes and is docs-only, so it spins NO units — that is your clean window, and it is the reason the order is what it is.

CONDITIONS:
1. Pool claim from your OWN worktree (.worktrees/322-adapter-peer-propagation), --jobs 2, release the pool at close. Standard.
2. NEGATIVE CONTROL FIRST on adapter_peer_e2e. It has never executed anywhere, and the thin lane compiles kind(lib)+kind(bin) — a --test binary is not in that filter, so you do not yet know it COMPILES, let alone that it can fail. Before you report a green from it: break it on purpose (assert the wrong peer, or point the install at a path that does not exist), watch it RED, then restore. A green from a rig that has never been shown to red is close to zero information, and this one spawns two real daemons, so a silent no-op green is the likely failure mode, not a crash.
3. `xtask gen` for --via-subnet / --remote: regenerate, then read the diff rather than trusting the generator. docs-site/** is CODE under the classifier that is landing, so this will correctly pull units on your PR — expected, not a surprise.

Report back: the negative-control red (test id + the assert you broke), the restored green, the gen diff stat, and only then rebase onto the new main tip and hand off.

Your F1 lines are accepted as raw. The entry_exec premise assert passing on BOTH boxes before the canonical fix is the part that makes it a measurement rather than a claim — that is the right shape, keep it.