ci: move workflow to .gitea/workflows so Gitea picks it up

Gitea only scans .gitea/workflows and .github/workflows; it does not scan
.forgejo/workflows (that path is Forgejo-only). The CI was therefore
dormant on a Gitea instance regardless of whether Actions was enabled.
git mv to .gitea/workflows/ (scanned by both Gitea and Forgejo). Updated
the live references in STRUCTURE.md (Gitea + act_runner), AGENT.md, and
today's ROADMAP entry; left the dated historical Shipped entries as the
changelog they are.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/ROADMAP.md

@@ -147,7 +147,7 @@ Pillar is **done** when every component has a closed `wave/qa-<component>` PR an
 
 (Move items here when they land — keep them brief, link the commit/PR.)
 
-- _2026-05-29_ — **Pillar 7: opt-in toggle eval matrix in CI.** `nix flake check --no-build` only evaluates the four default `nixosConfigurations`/`homeConfigurations`, so any eval-time bug that *only* fires when a `nomarchy.*` toggle is flipped sailed straight through — exactly how the vscode option rename and the impermanence systemd-stage-1 assertion both reached `main` undetected. New `bin/utils/nomarchy-eval-matrix` layers each opt-in scenario (16 toggles: impermanence single/multi, laptop/desktop presets, accessibility, gaming, hybridGPU, hibernation, docker, fwupd, overrides, panel position, keymap variant, toggles-off — plus home + system halves) onto `nixosConfigurations.default` via `extendModules` and forces `system.build.toplevel.drvPath` (which forces the assertion checks); it also forces all 21 palettes through the system-side Plymouth `base00`→RGB-float math that the home-only `allThemeVariants` never reaches, and both standalone `homeConfigurations`. Per-scenario results come from `builtins.tryEval` so one failure doesn't mask the rest — verified the harness goes red on both a failed assertion (re-injected `boot.initrd.postDeviceCommands`) and a nonexistent-option ref, and green on the current tree (all 39 scenarios pass). Wired into `.forgejo/workflows/check.yml` as a step after `flake check`, with `jq` supplied via `nix shell` so it doesn't depend on the runner image. `docs/AGENT.md` §5.5 now instructs: when you add or rename an opt-in `nomarchy.*` option, add a matching scenario to the script's `toggleScenarios` attrset.
+- _2026-05-29_ — **Pillar 7: opt-in toggle eval matrix in CI.** `nix flake check --no-build` only evaluates the four default `nixosConfigurations`/`homeConfigurations`, so any eval-time bug that *only* fires when a `nomarchy.*` toggle is flipped sailed straight through — exactly how the vscode option rename and the impermanence systemd-stage-1 assertion both reached `main` undetected. New `bin/utils/nomarchy-eval-matrix` layers each opt-in scenario (16 toggles: impermanence single/multi, laptop/desktop presets, accessibility, gaming, hybridGPU, hibernation, docker, fwupd, overrides, panel position, keymap variant, toggles-off — plus home + system halves) onto `nixosConfigurations.default` via `extendModules` and forces `system.build.toplevel.drvPath` (which forces the assertion checks); it also forces all 21 palettes through the system-side Plymouth `base00`→RGB-float math that the home-only `allThemeVariants` never reaches, and both standalone `homeConfigurations`. Per-scenario results come from `builtins.tryEval` so one failure doesn't mask the rest — verified the harness goes red on both a failed assertion (re-injected `boot.initrd.postDeviceCommands`) and a nonexistent-option ref, and green on the current tree (all 39 scenarios pass). Wired into `.gitea/workflows/check.yml` as a step after `flake check`, with `jq` supplied via `nix shell` so it doesn't depend on the runner image. `docs/AGENT.md` §5.5 now instructs: when you add or rename an opt-in `nomarchy.*` option, add a matching scenario to the script's `toggleScenarios` attrset.
 
 - _2026-05-29_ — **Impermanence rollback no longer fails to build.** `core/system/impermanence.nix` implemented the Erase-Your-Darlings root wipe via `boot.initrd.postDeviceCommands`, but `themes/engine/plymouth.nix:63` sets `boot.initrd.systemd.enable = true` distro-wide (Plymouth is imported unconditionally from `core/system/default.nix`), and systemd stage-1 initrd hard-rejects `postDeviceCommands` with a failed assertion. Net effect: **every** install that flipped `nomarchy.system.impermanence.enable = true` — including any user who picked impermanence at the installer prompt — produced a config that wouldn't evaluate, let alone boot. Converted the rollback into a `boot.initrd.systemd.services.nomarchy-rollback` oneshot unit ordered `after = systemd-cryptsetup@${mainLuksName}.service` and `before = sysroot.mount`, with `boot.initrd.systemd.initrdBin` pulling in `btrfs-progs` / `coreutils` / `util-linux` / `findutils` (the systemd initrd doesn't ship them by default, unlike the old scripted initrd). Script body (timestamped `@` → `old_roots` move, 30-day recursive subvolume GC, `root-blank` → `@` snapshot) is unchanged. Verified the assertion is gone and `nixosConfigurations.default` + `impermanence.enable` evaluates fully via `extendModules`. **Runtime-verification caveat:** the boot-time wipe semantics (the `[[ ]]`/function/IFS shell idioms and the cleanup loop under the systemd initrd shell, plus correct ordering vs the LUKS mapping) can only be confirmed by a real wipe-boot cycle on an impermanence install — fold into the Pillar 8 punch-list. `docs/TROUBLESHOOTING.md` persistence-block line reference updated (46-72 → 91-120).