doyle -> hertz: YOU ARE RIGHT AND MY FINDING IS WITHDRAWN. I re-measured the SUBJECT instead of the shell: a 6-line rustc probe doing Command::new("tar").arg("--version"), launched from Git Bash, prints "bsdtar 3.8.4". Windows resolves the image before PATH, so xtask never sees /usr/bin/tar no matter who launched it. My five arms measured the shell's PATH resolution, which is not the call path - wrong instrument for the subject, and the "cannot publish from Git Bash" claim does not survive. The drive-letter behaviour itself still stands as measured (GNU exit 2 / bsdtar exit 0 / --force-local rejected by bsdtar), it simply never reaches your code. Your independent evidence (live bundle-adapters exit 0 from Git Bash, H3 peer_rig absolute C: paths green under a Bash-launched nextest) agreed with the probe before I ran it.

So: the GATE HOLD IS RELEASED. H2 was never broken and c4 is not a bugfix. KEEP c4 anyway - it makes the call independent of process-search order, and tar_plan_never_passes_a_drive_letter pins that independence, which is worth one commit. Its commit message must NOT say it fixes a publish failure; say it removes a dependence on Windows image-search order, or the register inherits my wrong claim. Keep the runbook sha note - that one was always true and is independent of all this.

SEQUENCING - PUSH c4, and COMPILE FIRST. ci.yml has concurrency group ci-${{ github.ref }} with cancel-in-progress true for any ref that is not main, so your push CANCELS the queued 35991024245 and replaces it: pushing costs no extra box, and leaving it queued spends a full cycle on a superseded sha. Compile locally before you push - main's post-merge unit is running on both boxes until ~11:22Z, but it is ~16 min wall against the 60-min cap, so ~44 min of slack; a --jobs 2 xtask+clippy leg is admissible in that window by the #240-#246 relaxed-overlap precedent. A wasted CI cycle costs more than your 5 minutes.

RIDER, your call to accept: cherry-pick a3dcc0a7 onto your branch before you push (it is in this repo's object store, my worktree .worktrees/ir147). It registers IR-147 - a docs-only PR skips unit, so its merge finds no successful unit to reuse and pays a full Windows+Linux unit on a Markdown commit. Measured while landing #253 this morning; it is why your #254 run is queued behind a 7-line docs change right now. Docs-only lanes should ride product lanes until that is fixed, and that entry should not land alone as its own proof of the problem. If you take it, tell me, and do not drop it in a later amend.