rdmsm4x-changelog-20260924-0041-hypercpu-b2-to-b5-permanent-series-gpu-system-layouts-glass
HyperCPU went from the 09-18 b1 binary (still installed on every Mac, fading series that Rich saw as "jumpy as heck") to build 5 on 5 of 6 Macs in one evening: permanent series with fixed axes (0 whole-chart jumps measured on the installed app), GPU telemetry on both architectures, sysctl System charts, per-chart show/hide + order + Charts menu + Space Blue dark, and Classic/Grid/Focus layouts with Liquid Glass and RTTy-grade polish.
Written 2026-09-24 00:41 EDT on rdmsm4x by
claude@rdmsm4x (session c13f617a, Fable 5.1 oversight
lead). Ticket TASK-20260923-03; incidents ISSUE-20260923-15 (resolved),
ISSUE-20260923-14 (open). Lane changelogs:
docs/PRESENCE-FADE-20260923.md,
docs/evidence/h1-permanent-20260923/,
docs/evidence/h3/,
docs/design/LOOK-20260923.md,
docs/evidence/h2b/rdmsm4x-changelog-20260924-0037-hypercpu-laneH2b-look.md
(lane H2b's own record, folded in here).
Scope
Hosts: rdmsm4x, rdmpw3275m, rdmbair13m5, rdmbair15m5, rdmpw3265m
(installed); jdmbair13m5 offline since ~00:10 EDT — a detached retry
loop installs b5 when it wakes
(~/dev/_handoff/hypercpu-permanent-20260923/fleet-b5/retry-jdmbair13m5.zsh,
log beside it). Repo ~/dev/apps/HyperCPU main
b524a30 → 20bc2bd, pushed to backup and fleet
(no origin remote).
Builds
installed tonight (each replaces
/Applications/HyperCPU.app; the previous build is kept
beside it as
~/HyperCPU-backup-20260923/HyperCPU-b<N>.app on each
host, and under
~/dev/_handoff/hypercpu-permanent-20260923/backup/ on
rdmsm4x)
| build | merged | what | evidence |
|---|---|---|---|
| b2 22:34 | 2a1d8ab (lane H1, codex) |
series permanent by default (fade opt-in "Fade idle series"), power/frequency axes growth-only; DEC-HYPERCPU-20260923-01 | installed-app capture 60 frames / 30 s: before 16 of 59 whole-chart jumps (max 9.74) → after 0 (max 0.88); tests 229 → 237 |
| b3 23:10 | 3b10b17 (lane H3, codex) |
GPU per IOPCIDevice on both architectures (Apple Silicon device/renderer/tiler + memory; AMD Vega20 device utilization, GPU Activity, VRAM), sysctl sampler (num_files/maxfiles, vnodes, task/thread limits, loadavg, swapusage), GPU + System chart groups | tests 253; capture 0 jumps |
| b4 23:29 | 7a65b44 (lane H2a,
codex) |
nine modules with persisted show/hide + order (Settings → Dashboard, drag-to-reorder, Restore default), Charts menu in the menu bar, Appearance System/Light/Dark + Dark board Current/Space Blue | tests 269; Settings → Dashboard verified on the installed app |
| b5 00:39 | e3d7fde (lane H2b,
Claude) |
Classic/Grid/Focus layouts, transparency Off/Subtle(default)/Strong with Liquid Glass on 26+, uppercase letter-spaced titles, one chip style, tabular numerals, honest footer | tests 290; dashboard verified on the installed app |
Relaunched where it was running (rdmsm4x, rdmpw3275m); installed-not-launched on the hosts where it was not running. b4 on rdmsm4x had been quit by a person at ~00:34 (not by a lane); the b5 install relaunched it.
What was wrong, measured
Every /Applications/HyperCPU.app on the fleet was the
09-17/18 b1 binary; the 20:33 presence-fade merge had never been built
or installed, and its evidence came from a harness ("no live series
reached zero on this host"). The fade design itself still made series
come and go. Recorded as memory
lane-fix-not-fixed-until-installed.
Commands run (representative)
git merge --no-ff codex/… ×3 and
look/20260923; scripts/build.sh in each lane
clone; ditto install + open per host over ssh
(rsync -a --delete to /private/tmp,
hash-checked after); frame captures with
screencapture -l <window> + a PIL/numpy frame-diff
script; ioreg -r -c IOAccelerator -d 1 and
sysctl reads on rdmpw3275m for the H3 keys.
Undo
Per host:
pkill -f '^/Applications/HyperCPU.app/Contents/MacOS/HyperCPU'; rm -rf /Applications/HyperCPU.app; ditto ~/HyperCPU-backup-20260923/HyperCPU-b<N>.app /Applications/HyperCPU.app; open /Applications/HyperCPU.app
(b1 through b4 kept). Repo: normal commits on main.
Outstanding
- ISSUE-20260923-14: helper still labelled
com.example.PowerGadgetHelper(identity change for the daemon on six Macs; daytime lane). - Not proven: Intel/AMD look of b5, glass CPU cost, pre-26 material fallback, Focus click-to-enlarge by hand; right-edge time label clipped (pre-existing axis code).
- jdmbair13m5 install pending its wake (retry loop running).
No secrets in this record.