Physical K230 shell memory, host date 2026-09-22

Source transcript: docs/evidence/shell-memory.txt
Image: /nix/store/s50h4h5awhx9479nhflayif890z2r6c0-k230-sd-image.img
Kernel: 6.6.36; stock Sway1.12, Pixman RGB565, 568x1232
State: idle terminal after confirmed reboot, with no camera/screencopy sampler
or Monitor workload running. Keyboard visibility changed by RTMIN signal;
this measurement does not claim a finger tap. The board clock is unset.

                                 keyboard hidden    keyboard shown
Session PSS (sum, KiB)            37500              40950
Session PSS (MiB)                 36.62              39.99
shell.service MemoryCurrent(B)   50073600           53559296
shell.service MemoryCurrent(MiB)  47.75              51.08
Linux usable memory(MiB)         935                935
Linux available memory(MiB)      760                759
Swap(MiB)                        0                  0

PSS is summed across the shell user's graphical/session processes and
proportionally accounts for shared pages. Individual process RSS values are
in the transcript; summing RSS would double-count shared pages. The cgroup
MemoryCurrent also accounts for memory charged to the service beyond the
process PSS, so the two values are not interchangeable. These are sequential
snapshots, not an atomic whole-system measurement or a peak-load benchmark.

CPU-side frame-work measurement, host date 2026-09-22

Verdict: Sway/Pixman is viable for this terminal-and-touch-menu shell. The
20-second measured interval with the normal 15pt portrait terminal scrolling
and the 400px keyboard visible yielded 254 successful instrumented output commits:
  median 14.0037405 ms; p95 15.585629 ms; p99 22.735222 ms
  min 0.714408 ms; max 23.253370 ms; failed commits 0.
These durations do not imply an FPS, physical refresh, or touch-latency result.
The output reports 52.190 Hz (about 19.16 ms per refresh), but this probe does
not wait for display completion and the workload does not demand every refresh.

Method and exact evidence:
- Full journal: docs/evidence/shell-frame-journal.txt (includes startup and two
  preliminary runs using Foot's smaller default font, excluded from statistics).
- Final launch marker: shell-frame-portrait-start.txt, uptime 1587.38 s.
  Fixed analysis interval [1589.38, 1609.38), excluding the first two seconds.
  shell-frame-portrait-end.txt records collection after the workload.
- Per-sample data and nearest-rank percentile method:
  shell-frame-timing/measurement.json. Select journal lines whose monotonic
  timestamp lies in that interval and parse
  build_state_plus_commit_submission_ns; divide by 1e6 for milliseconds.
- Workload source: shell-frame-timing/scroll.sh: 400 numbered text lines, a
  requested 50 ms sleep after each line. Process creation and scheduling add
  delay, so this is a nominal 20-line/s producer, not a measured 20 Hz rate.
  Foot uses the image's 15pt terminal configuration. No wlfps overlay is used.
- Keyboard shown with SIGUSR2; workload launched through Sway IPC. Screenshot
  shell-frame-timing/portrait-scroll-keyboard.png proves the visible layout.
  The adjacent camera JPEG shows the physical terminal but its oblique framing
  does not show the entire keyboard. These are automated controls, not finger
  accuracy evidence. One Grim screenshot and serial retrieval occurred during
  the analysis interval; this is not an instrumentation-free benchmark.
- k230.shell.frameTiming selects the Nix-built diagnostic Sway package and
  SWAY_K230_CPU_FRAME_TIMING=1. The same binary and environment were installed
  temporarily through a runtime systemd override on the existing verified image,
  without rebuilding/flashing a diagnostic image. shell-frame-confirm.txt proves
  environment, executable hash, service invocation, and 568x1232 mode.
  Diagnostic package: /nix/store/l1v7llidvivsmy0s7g9lf314kvr82m6b-sway-1.12
  Actual executable: /nix/store/jwnazrqchp75n155mfwmbks3y348nw6b-sway-unwrapped-riscv64-unknown-linux-gnu-1.12/bin/sway
  MD5 on host and board: 8137c3ffa57ef99148da756323f8165a.
- Patch: nix/patches/sway-k230-cpu-frame-timing.patch. CLOCK_MONOTONIC wall
  time starts immediately before wlroots scene-output state initialization /
  build and ends after output commit returns. It includes scene/Pixman work,
  KMS submission and scheduling delay. It is neither process CPU accounting,
  presentation timing, scanout, nor end-to-end responsiveness. Sway debug logs
  and the diagnostic logging itself may affect timing.
- All journalled builds/commits succeeded in this diagnostic session. The
  service was then restored to its stock executable, with no runtime drop-in;
  shell-frame-restore.txt records the active service and native output mode.

The remaining usability gates are deliberate real-glass accuracy/control tests
and unattended cable-free boot. Neither is established by these measurements.

Nix option evaluation (no diagnostic image flash):
{"compositor":"/nix/store/l1v7llidvivsmy0s7g9lf314kvr82m6b-sway-1.12","environment":"1"}


Representative journal lines (minimum, middle, maximum; full log cited above):
[ 1589.504125] nixos sway-frame-timing[874]: 00:02:22.927 [INFO] [sway/desktop/output.c:342] K230_CPU_FRAME_WALL output=DSI-1 build_state_plus_commit_submission_ns=714408 committed=1
[ 1598.940200] nixos sway-frame-timing[874]: 00:02:32.367 [INFO] [sway/desktop/output.c:342] K230_CPU_FRAME_WALL output=DSI-1 build_state_plus_commit_submission_ns=14005185 committed=1
[ 1599.974114] nixos sway-frame-timing[874]: 00:02:33.396 [INFO] [sway/desktop/output.c:342] K230_CPU_FRAME_WALL output=DSI-1 build_state_plus_commit_submission_ns=23253370 committed=1

Host validation after committing measurement evidence:
  openspec validate --all: 14 passed, 0 failed.
  tools/blob-scan.py --no-vendor: all 74 binary files accounted for.
  scripts/build_site.py: 55 pages, 1,099,242 bytes, 10.77 seconds; passed.