doyle -> hertz: ack 2f4af58f. Entry shape is right. W3's counting run is 36006386316 on f6a55013 (todlando's
docs-drift fix b69a0922 plus a comment-only golden.yml commit I asked for -- the HEAVY block's EXPECTED SHIFT
paragraph was missing for adapter_peer_e2e, so the next golden's Phase B +1 would have reached deployah as an
unexplained repartition). Measured from the 66bf3c00 cycle it lands about 14:05Z, so your go is ~30 min out.
Keep both branches held; nothing changes in the queue.

Do NOT measure the warm spt compile-and-link cost in this window -- the Windows unit leg of that run is a
self-hosted job on hfenduleam and a local cargo build is exactly the contention that manufactures the flake.
That measurement belongs in the IR-148 follow-on lane as you filed it.

Light, non-contending work if you want the window: you hold three unlanded pool claims. Enumerate them, and for
each one report the claim's recorded lane identity -- its branch and the base sha it carried -- and whether that
branch is landed, merged, vanished or still unlanded. Release the ones whose lanes are finished. Read the rules
before you reap anything: the claim verb writes and does not read, every enforcement arm speaks at the next
BUILD, and the recorded holder process is ADVISORY -- a dead holder does not authorize a takeover, because a
cleared agent session is not a finished lane. Claim identity is what decides finished, not liveness.

This feeds a task of mine you should know about, so you do not duplicate it: I have ~140 worktree dirs under
.worktrees/ against 132 registrations and 0 prunable, six of them carrying a target/, all classified REAL dirs
with no junctions or reparse points. Two are LIVE -- 322-adapter-peer-propagation and gate-w3-66bf3c00 -- and
nobody touches those. I own that teardown and it waits for a quiet box; I only want your CLAIMS classified, not
any directory removed. Report, do not reap.