Plymouth cost audit

This is the laptop-side audit for the-screen-lights-before-linux task 5.1. It evaluates the pinned nixpkgs and the current shell configuration without enabling Plymouth in nix/hardware.nix and without building a package.

The design budget is no more than the shell’s measured 89 local riscv64 derivations / 875 MiB fetched-unpacked delta. The comparison method is the one in docs/display-environment-options.md: evaluate the complete k230 system with the existing shell enabled, then run nix build --dry-run against the candidate system. The baseline nix build --dry-run .#toplevel produced no output on this host because its paths are already present.

The reproducible candidate expression is tools/k230-plymouth-eval.nix. It composes the same k230.nix, hardware.nix, and shell.nix modules as the flake, then enables Plymouth with the built-in spinner theme. spinner is the minimal theme baseline: it is supplied by the Plymouth package itself and needs no extra theme package. Commands used from the repository root:

nix eval --impure --expr 'import ./tools/k230-plymouth-eval.nix'
nix build --dry-run --impure --expr 'import ./tools/k230-plymouth-eval.nix'

The initial evaluation (before the helper was narrowed to spinner) produced /nix/store/x4jjd44s91yrhnx40rjy9l7w9vcz85az-nixos-system-nixos-26.11.20260919.20b1ddd.drv. After the native media work changed the warm store, the exact committed helper was rerun at 2026-09-22T18:08:06Z. It produced /nix/store/yxclgzsb8v2qhg34mz4ldlzak86pdb9k-nixos-system-nixos-26.11.20260919.20b1ddd.drv. That rerun’s dry-run summary was:

76 derivations will be built
83 paths will be fetched
481.5 MiB download
897.3 MiB unpacked

The counts are unchanged between those two evaluations; only the toplevel derivation path changed. The 76 derivations include the riscv64 Plymouth build and its generated initrd units, plugins, themes, fonts, and transitive closure. The 83 fetched paths are mostly development inputs and image/font libraries. These figures are incremental against the warm store used by the shell measurement; they are not a from-scratch closure.

Applying the written rule to the required metric: Plymouth is over the recorded budget by 22.3 MiB and should take the minimal DRM splash fallback. The 76-derivation count is within 89, but the fetched-unpacked figure is the binding comparison and 897.3 MiB is greater than 875 MiB. This is a dry-run budget verdict; it does not claim a final installed closure size.

The final closure byte count remains unknown because nix build --dry-run reports fetched unpacked sizes, not the output sizes of derivations that would be built. That uncertainty does not change the dry-run verdict, since the design rule explicitly budgets the fetched-unpacked metric.

The pinned NixOS module supports the requested integration. With the current configuration it evaluates to boot.initrd.systemd.enable = true, theme spinner, the default plymouth-riscv64-unknown-linux-gnu-26.134.222 package, and no extra theme package. The module places plymouthd, its renderer and theme files, fonts, and systemd units in the initrd, starts Plymouth from initrd-switch-root.target/sysinit.target, and provides the corresponding quit/switch-root units. That establishes a supported integration path; it does not prove that this panel’s early DRM device is available or that the splash hands off without a dark frame.

No image, kernel, hardware, or Wi-Fi state was changed. Task 5.2 should not be started under the design rule; task 5.3 is the documented fallback and task 5.4 remains a hardware claim.

Task 5.3 implementation boundary

The selected fallback is a small libdrm owner in the main system. It opens the primary node, enables universal-plane enumeration, finds a connected connector, preferred mode, CRTC, and matching primary plane, then creates a dumb buffer. The source asset is immutable logo.xrgb in the Nix store. The K230 primary plane does not advertise XR24, so the default path converts its B,G,R,X bytes to RG16; AR24 is available only when the discovered primary plane advertises it. There is no animation because this program only bridges the static stage-1 frame until the shell begins.

On shell handoff the owner drops DRM master on SIGUSR1 but retains its file descriptor and framebuffer. It polls the CRTC until a non-zero successor framebuffer replaces its own, then removes only that replaced framebuffer and its dumb allocation, without issuing a CRTC clear. Until a successor appears it retains the active objects; service stop is the only other exit path. This arrangement avoids treating close-time framebuffer lifetime as a handoff mechanism. It still does not prove that the panel will remain lit: DRM object lifetime after file release and Sway’s first modeset are driver-dependent. Task 5.4 must film the transition before any continuity claim is made. The service is present only when both k230.panelConsole is false and the shell is enabled; the daily no-logo console image keeps its current path. A shell restart accepts an already dropped master or an inactive optional owner, since neither can retain DRM master, but refuses to start after its bounded wait if an active owner never reaches scanout.

The splash-enabled toplevel build passed on 2026-09-22 with k230.panelConsole = false: /nix/store/m39vc96iaw7s5vlyxmlxa9a027kkizag-nixos-system-nixos-26.11.20260919.20b1ddd. Build and installed unit evidence records the exact command, the built unit, its immutable program, and the shell’s release hook. Task 5.3’s build claim is complete; physical handoff task 5.4 remains open.

A manual physical diagnostic now exercises the owner and its retained-buffer handoff. One warm-boot trial rendered Sway correctly; a controlled repeat wrapped the logo during the owner’s first modeset and left Sway physically wrong, despite successful DRM calls and a correct native screenshot. See phase recordings and register evidence. This establishes the owner protocol running on the board, not reliable visual handoff or automatic service startup. Task 5.4 remains open.