refactor(flake): add user registry and multi-user host support

Separate user identity (data) from the reusable modules, and let a host
declare any number of users instead of exactly one.

- users/registry.nix: per-user identity (name, email, groups, authorized
  and signing keys) as the single source of identity; no user data is
  hardcoded in the modules.
- mkHost takes a `users` set keyed by username; per-user identity is
  injected into each home config via the `identity` module arg
  (extraSpecialArgs is per-host, so it cannot carry per-user data).
- modules/users.nix builds accounts from the registry; modules/ssh.nix no
  longer defines authorized keys (the registry owns them); home/git.nix
  and home/desktop.nix read `identity`; users/emmathorpe/work.nix drops
  its now-redundant git identity override.
- Restructure the tree: users/, home/, modules/, hosts/, lib/ replace the
  former lyrathorpe/ and system/ layout.
- Add standalone homeConfigurations (the portable subset: shell, git,
  editor, claude) and an exported homeModules output for use on machines
  not managed by this flake, or as an input to other flakes.

Behaviour-preserving for existing hosts: lyrathorpe-mbp and
emmathorpe-edaas evaluate to identical derivations; lyrathorpe-t400,
lyrathorpe-macpro31 and lyrathorpe-rpi5 differ only by de-duplicating a
repeated authorized_keys entry. Fixes the SSH authorized-key leak (one
user's key was applied to every account), the hardcoded default git
identity, and the hardcoded EDaaS linger setting.

home/claude/memory/git_commit_signing.md

@@ -0,0 +1,22 @@
+---
+name: git-commit-signing
+description: "Commits sign in-sandbox via ssh-agent — needs `allowAllUnixSockets: true` in settings, plus pubkey inlined in user.signingkey."
+metadata:
+  node_type: memory
+  type: feedback
+  originSessionId: a223254b-6bee-435f-ac39-e3cedf064893
+---
+
+Lyra's git is configured to SSH-sign commits (`commit.gpgsign=true`, `gpg.format=ssh`). The sandbox masks `~/.ssh/*` (read-denied; the files appear as char devices backed by `/dev/null`), so git cannot read a file-based `user.signingkey` and ssh-keygen cannot read the private key directly. Signing in-sandbox therefore requires routing through ssh-agent over the agent's unix socket.
+
+**Working setup (as of 2026-06-02):**
+
+1. NixOS / home-manager runs an ssh-agent so `/run/user/1000/ssh-agent` exists and `SSH_AUTH_SOCK` is exported into the sandbox env.
+2. `~/.claude/settings.json` has `sandbox.network.allowAllUnixSockets: true` to let the sandbox `connect()` to that socket. On Linux/WSL2 this is the ONLY available switch — the per-path `sandbox.network.allowUnixSockets` array is macOS-only because the seccomp filter cannot inspect socket paths. Tradeoff: every unix socket on the host (including `/var/run/docker.sock` if present, DBus, etc.) becomes reachable from sandboxed commands.
+3. `user.signingkey` set to the inlined pubkey: `git config --global user.signingkey "key::$(cat ~/.ssh/id_ed25519.pub)"`. Must run with DOUBLE quotes outside the sandbox so `$(...)` expands; single quotes or running it from inside the sandbox stores literal garbage (`cat ~/.ssh/id_ed25519.pub` reads `/dev/null` in-sandbox).
+
+**Why:** removes the per-commit `! git commit ...` friction; private key stays in the agent, never enters the sandbox.
+
+**How to apply:** Commit normally with `git commit`. If signing fails with `Couldn't load public key`, check (a) `git config --get user.signingkey` starts with `key::ssh-ed25519 AAAA...` (not literal `$(...)`), (b) `ssh-add -l` from in-sandbox lists keys (if it says "Operation not permitted", the sandbox config didn't take effect — restart Claude Code), (c) the ssh-agent on the host actually has the key loaded (`ssh-add -l` outside the sandbox). Do NOT use `--no-gpg-sign` to bypass — the repo's `ReleaseWorkflow-Commit` check enforces signed commits.
+
+Related: [[git-network-ops]], [[git-conventions]].