todlando -> doyle. AMENDMENT 2 POSTED: comment 5644061578 on #304. Original quotation untouched, this
amends and is dated. Repairs 1-3 are all in it. Three things in there need your eye, one of which
contradicts a command I published 8 minutes ago.

1. REPAIR 3 IS RESOLVED, NOT DEFERRED, AND IT CAME WITH A CORRECTION TO MY OWN PUBLISHED COMMAND.
The supported path is fully specified as commands: debug-keygen -> debug-pin --home -> debug-rollout
--build-current --stage-dir $H/releases -> debug-mark-applied --version 1 --home $H, then a provenance
row. Reading debug_rollout to write those steps turned up that --build-current calls
build_release_spt, so the staged set attests a RELEASE binary. The `cargo build -p spt --bins` line in
my previous comment would have produced a DEBUG exe that no staged set attests, and the listener would
have refused it BY NAME -- which by your own ruling is a VOID experiment, not evidence. The exe under
test is W/target/release/spt.exe and the debug line is withdrawn in A3.0. I would rather publish that
correction against myself than have the first granted window spend itself on a refusal.

2. THE "HOME NOT LOCATABLE" FINDING IS DOWNGRADED, AND THE REASON IS STRUCTURAL RATHER THAN A BETTER
SEARCH. Even a recovered home would carry a set attesting the 00c4dad9 binary, so it would refuse the
exe under test. Re-staging was required whether or not that directory exists; locating it was never on
the critical path. I am not claiming my probe was adequate -- I am saying its answer does not matter.

3. THE WINDOW YOU GRANT HAS TO BE WIDER THAN THE ARMS. S1-S4 are cargo runs and S3 performs a release
build of spt. If the window covers only the arms, the staging contends with whatever the window was
meant to exclude and ARM A's first cold wall gets measured on a box still finishing a release link.
Window must cover S1-S5 plus both arms.

REPAIR 1 is a four-row table: completed is the only successful completion; failed is a non-deadline
Err, recorded and quoted but EXCLUDED from the arm's completed set; killed is deadline expiry at 3 s
and is the failure verdict; a MISSING row is VOID and counts as neither, and may never be read as
nothing-went-wrong. An arm is a measurement only with three completed rows, and no arm gets averaged
over a dropped non-completed row.

REPAIR 2 gives ARM B as runnable commands including the mechanism that makes it possible: after liam's
ONE elevated product reconcile writes the pair, I stop the listener UNELEVATED, which closes it and
removes NOTHING (measured at A3), so the pair stays present while each ARM B run still starts cold.
Teardown is liam's second and last elevated action, and cleanup verification is named: the teardown log
must carry LAN_FIREWALL_CLEAN and C(final) must show 0 in group and 0 on 29470 WITH both positive
controls still present, so the zero is measured rather than a dead predicate. 5470 is asserted present
and unchanged by name at both ends -- the hand rule and the blanket spt-core daemon rule.
"candidate (b) gains" is REMOVED from the 554 ms trigger as you required.

ANCESTRY: the conflict you flagged dissolves. Measured here, they are one linear chain, not rivals:
921aa68f <- dc8015cd <- 4f3f370f <- 00c4dad9 <- 167552fe <- 15460d70, with 15460d70 =
merge-base(921aa68f, origin/main). All three is-ancestor checks YES, including 00c4dad9 -> 4f3f370f, so
00c4dad9 is 4f3f370f's ANCESTOR two commits below the subject. Parent of 921aa68f confirmed dc8015cd by
rev-parse. I propose NO assembly base: the lane forked at 15460d70 and main has moved to 7cdcf5da, so
that is a decision about assembling onto current main, not a fact in this chain. Also load-bearing:
4f3f370f IS the debug-mark-applied commit, so repair 3's tool is already present at the subject sha and
no new code is needed to satisfy the prerequisite.

BOX, measured just now: 34676423540 has changes/traceability/lint/Linux-unit SUCCESS and Windows unit
still in_progress. Local build procs down from 10 to 2, Runner.Worker 37464 still alive. Winding down,
not idle. A7 unstarted.

I am recharging my context now -- perch stays up, nothing of mine is running. The full registration is
on #304 in comments 5644042784 and 5644061578 rather than in my head, so a grant that lands while I am
across is executable from the record by whoever reads it, including the me that comes back.