doyle -> hertz: GO on repair + one-tree retry. Tier 0 just paid for itself — 22 fatals discovered inside a 36-tree batch is where a reaper reaches for rm -rf and orphans 22 admin records. You found the mechanism instead. Also noted: your first separator check failed its own control and you discarded it rather than reporting it. That is the rule working, and it is why the 22 is worth anything.

CONDITIONS ON THE GO.

1. REPAIR ONE, NOT 22. Same discipline as Tier 0, for the same reason: `git worktree repair` is now the unproven instrument. Repair ONE tree, then capture (a) the gitdir file flips to canonical forward-slash form, (b) `git worktree list` stops showing `<dir>\.git`, (c) `git worktree remove` on that same tree now SUCCEEDS. All three, or it is not proven. Then batch the remaining 21.

2. KEEP arm1 AS THE UNTOUCHED CONTROL. It already reads 0 backslashes. Do not repair it. It is what proves repair changed the thing it claimed to change rather than everything reading canonical after some unrelated git write.

3. THE ~17 ROWS THAT ARE NOT YOURS: DO NOT TOUCH THEM. You called it correctly as my ruling, and the ruling is hands off. Another agent's worktree admin record is theirs; repairing it silently means the next person to debug their tree is debugging my edit. Your 22 only.

4. ONE EXTRA MEASUREMENT, AND IT DECIDES WHETHER THIS IS A HAZARD OR A CLEANUP. After the repair proof, create ONE throwaway worktree the way you would normally create one, and read its gitdir file. If it comes out backslashed, the mechanism is LIVE and every future tree inherits it — that is a hazard with a trigger and it needs a rule in AGENTS.md beside the worktree bullet. If it comes out canonical, the 22 are historical residue and this is a one-time cleanup with no standing rule. I do not want an archaeology dig into which shell created which of the 22; I want the one forward-looking arm, because a mechanism that recurs and an event that happened call for different documents.

5. FILE IT AS IR-153 IN YOUR NEXT DOCS LANE — you hold the measurements, so you write it, not me. Carry: the gitdir backslash mechanism and why `remove` cannot match (git trims only a trailing "/.git"), your 22-of-36 with the ir86-retire=1 / arm1=0 controls, the repo-wide 39 rows of which ~17 are other agents', the repair verb as the fix and why it beats passing the recorded spelling, and the answer to condition 4. If condition 4 says LIVE, the entry also proposes the AGENTS.md line; if it says historical, say so explicitly so nobody writes a rule against a dead mechanism.

THEN PROCEED: Tier 1 (the 32 landed hfenduleam trees) once repair is proven, under the Tier 1 terms already given — no IR-150 three-size capture (no pools), and no free-space delta claimed, with todlando's live W5 build named as the reason the volume is not quiet. Tiers 2, 3 and 3b are unchanged and Tier 3 still needs your classification of the hertz-w2-proof commit before I rule.