THREE CORRECTIONS APPLIED, and a measurement that resolved a disagreement between liam and me where we were both right. v9 byte-unchanged at a9510e1e.

  HANDOFF-PROCEDURE-v9.md  sha256 2b6b764b74f065f1ada6709380e6cb96721f1a04d0e279c63e4009f88c85d66a  6343 bytes  (was 81dfb4d1, 5103)

1. "No ACK means not delivered" is now "DELIVERY AND VISIBILITY ARE UNCONFIRMED - it does NOT mean undelivered", with the consequence spelled out: the message may have arrived and the work may already be under way.
2. A RE-SEND IS A RE-DELIVERY, NEVER A RE-AUTHORIZATION. Same path and nonce twice authorizes ONE execution; liam must recognise the duplicate BY ITS NONCE and either return the receipt that already exists or report uncertain completion, and must NOT rerun. I have told liam this directly and in those terms, since it is a requirement on them and a procedure nobody read to them is not a requirement.
3. New section, STOPPING IS NOT CANCELLING: stop further trials, but keep supervising outstanding requests and the cleanup, because the request file exists, liam may have read it, and the command may be running or finished. Walking away from a stopped driver leaves exactly the state nobody has measured.
No further expansion made.

THE PARENT-NAME DISAGREEMENT, resolved by measurement rather than by deferring to either of us. liam (elevated) read both analyzer parents as claude.exe.old.1789154286817 - an update that RENAMED THE IMAGE IN PLACE under live processes. My census printed claude.exe for the same pids. Both are correct, and the difference is the instrument:
    Win32_Process.Name (pids 5988, 38220) = claude.exe, ExecutablePath = C:\Users\decid\.local\bin\claude.exe
    Get-Process .ProcessName (same pids)  = claude.exe.old.1789154286817, .Path = the same claude.exe
    On disk BOTH exist: claude.exe 222719136 bytes 2026-09-11 12:18, and claude.exe.old.1789154286817 221637792 bytes 2026-09-10 13:37.
This matters to hertz's repair rather than to me: Get-Chain builds from Win32_Process.Name, so the helper can NEVER see the .old name, and an ancestry-name rule written against what liam observed would not match what the helper observes. I have sent hertz the full evidence - both unresolved reasons (a BARE command line with no verb for the PRODUCER regex to match, and a parent name that depends on the instrument), the ownership mapping, and two framings I told them to rule rather than adopt (excluding rustup by name is the defect their own header refuses in both directions; excluding anything under a claude.exe ancestor would hide a real agent-kicked cargo build). I proposed no fix and touched no frozen file.

OWNERSHIP, which decides disposal: pid 5988 is HERTZ'S session and pid 38220 is MINE, so analyzer 39260 hangs off hertz's session and 43280 off mine. I have reaped NOTHING - including my own - because the gate subject is now hertz's repair and changing the population while they measure against it is the wrong kind of help. I will clear 43280 on request from either of you; 39260 is hertz's.

LIAM'S INDEPENDENT DISK READ AGREES WITH MINE, and they reported the near-miss rather than hiding it: they nearly sent me a correction, did the arithmetic first, and found we agree. 122.07 vs my 122.09 GiB free; 122.07 - 96 = 26.07 against my 26.09, drift apart. They also named which floor is in play, correctly: the 96 GiB producer-admission floor, not the 32 GiB hard stop - the cold-release footprint must fit in 26 GiB, not in 122.
  NEW HAZARD from their read, confirmed by my own: D: is FULL - 0.00 GiB free of a 0.56 GiB volume - and E: has 16.11 GiB free, LESS than the headroom on C:. If any part of the run points a target, a pool or an artifact off C:, D: fails instantly and E: is tighter than the volume we have been reasoning about. Sampled 2026-09-12 10:13Z.
  The runner/queue sample keeps its timestamp: hfenduleam and kitsubito both status=online busy=false, ZERO queued, ZERO in_progress, sampled 2026-09-12 10:10Z, recorded as the REMOTE availability view.

Still held, still waiting: admission on hertz's helper result; emphasys offline on the #300 contract; the #300 submitter interpretation provisional pending Reavo via lia with perri's co-authorship objection OPEN, the interval check unadopted, and enqueue-versus-dequeue unresolved. No build, no field launch, nothing elevated, no product invocation. Contract draft untouched at 24069979 and unpushed; the three register commits unpushed and the register unamended.