feat(ttl): Run Script 4's CPU stress as the same systemd unit

The thermal / safety macro started mlucas-avx2 with '&', exactly as
Script B used to, so it carried the same failure: the load ended quietly
when the console went away - in the one test where a long, steady full
load is the entire point of the run.

It now uses the same mlucas-amm transient unit, and 'systemctl is-active
mlucas-amm' joins 'bgctl list' and 'jobs' in the monitor loop.

Script 4 still has no teardown section, so the matching stop stays
manual, and it is now two commands rather than one:

    blanton_tr518.sh stop
    systemctl stop mlucas-amm

Closing Tera Term no longer stops the CPU load, and 'kill $(jobs -p)'
never could reach a unit. Script A's V1.1.8 history, CLAUDE.md,
ARCHITECTURE.md and the release note all say so now; the version stays
V1.1.8.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014RetWKFZFG1ZHcitQAyhwM

docs/release-notes/v1.1.8.md

@@ -3,7 +3,7 @@
 ## ✨ New features
 
 **`mlucas` now runs as a systemd unit, so it survives a lost session**
-* The AMD CPU load is started as `systemd-run --unit=mlucas-amm --property=Type=exec --property=Restart=always --property=RestartSec=2`, instead of being backgrounded with `&` in the login shell.
+* The AMD CPU load is started as `systemd-run --unit=mlucas-amm --property=Type=exec --property=Restart=always --property=RestartSec=2`, instead of being backgrounded with `&` in the login shell. Script B and the thermal/safety macro (Script 4) both use it.
 * As a shell job it belonged to that one terminal: close Tera Term, drop the console, reconnect on a different session, and the load was gone — while `jobs` from the new session showed nothing to tell you. A multi-day soak could spend most of its length with an idle CPU and produce thermal numbers that looked fine because nothing was loading the part.
 * `Restart=always` brings it back if it exits on its own, and the log is appended (`>>`) inside the unit while being truncated (`>`) once before the unit starts, so a restart adds to the run rather than erasing it.
 * Script B now prints `systemctl is-active mlucas-amm` beside `jobs`, both in the one-shot check and in every pass of the monitor loop. One word — `active` — tells you the CPU is still loaded.
@@ -53,7 +53,7 @@ ls ~/hammer/tools/amd/mlucas-avx2   # must already be there - see below
 * **The three stress durations are deliberately different**: SSD and USB at 1 hour, the BMC USB ping at 2 days. Set each to match the soak you are actually running.
 * **`--runtime` is the job's overall limit, not a per-pass one.** `qfx5252-stress-*` stops after `RUNTIME` seconds whatever `--passes` says. For a 24 hour soak set `RUNTIME` to `86400`.
 * **`mlucas-amm` has not been run on hardware yet.** The unit resolves `~` through bash's fallback to the passwd entry, because systemd does not set `HOME` for a system service without `User=`. It should land on `/root/hammer`, the same path the interactive shell uses — confirm it on the bench, or change both paths to an absolute `/root/hammer/...`.
-* **Script 4 still backgrounds `mlucas-avx2` with `&`.** The two macros use different mechanisms now; do not copy one's teardown into the other.
+* **Script 4 uses the same unit but has no teardown.** The thermal/safety macro now starts `mlucas-amm` the same way Script B does, and it still ends at `while 1` — so stopping it means `blanton_tr518.sh stop` **and** `systemctl stop mlucas-amm`. Closing Tera Term does not stop the CPU load any more.
 * **Script 5 (reboot cycle) is still unverified on hardware**, still has no exit condition, and `wait "Booting in blind mode"` still has no `timeout` guard.
 * **`blanton_multiphase_margin.sh save` is permanent**, and must not be mixed with `margin.sh` — same rails, two access paths, no interlock.
 * **`READ_VIN` and `READ_IIN` are still not in the ADPM12200 coefficient table.**