feat(installer): add ZCode install target

[modoojunko]

What

Adds zcode as an install target (codegraph install -t zcode, --print-config zcode, multiselect, --target=all/auto), wiring the same three layers Claude Code gets: MCP server entry, the UserPromptSubmit prompt-hook, and the marker-fenced instructions block.

Why

ZCode is a desktop coding agent with an MCP client, an AGENTS.md instruction surface, and a hooks runner — but codegraph install had no target for it, and a Claude-style MCP snippet pasted into ZCode's config silently does nothing (shape mismatch below). All config surfaces described here were verified against a live ZCode install.

How ZCode differs from the other JSON targets

1. MCP servers are nested under mcp.servers in ~/.zcode/cli/config.json (global) / <repo>/.zcode/config.json (local) — not the top-level mcpServers key Claude/Cursor use. Both scopes auto-connect at session start.

   {
     "mcp": {
       "servers": {
         "codegraph": { "command": "codegraph", "args": ["serve", "--mcp"] }
       }
     }
   }

2. Server entries carry no type key. ZCode's bundled servers (node_repl, computer-use) and a verified-working codegraph entry are bare { command, args }; tolerance for type: "stdio" is unverified, so install writes the verified shape — and normalizes a hand-copied Claude-style entry down to it.

3. Config-file hooks need hooks.enabled: true (plugin-contributed hooks auto-enable the hook runner; config-file hooks don't — without the flag the hook is parsed but never fires). Hook entries are also process-shaped — { type: "process", command, args, timeoutMs }, not Claude's { type: "command", command: "<shell string>" } — so the platform handling moves from the command string (codegraph.cmd prompt-hook, Windows: UserPromptSubmit hook written as "codegraph prompt-hook" fails in Git Bash (missing .cmd extension) #1466) into the argv array (command: "codegraph.cmd" + args: ["prompt-hook"]). The installer's prompt-hook question is now offered when ZCode is a target, not just Claude.

4. Instructions go to ~/.zcode/AGENTS.md / <repo>/AGENTS.md — ZCode reads AGENTS.md (user scope first, then workspace narrows it), so the same conditional marker block works unmodified in both scopes.

Uninstall semantics

Surgical, mirroring the Claude target: only mcp.servers.codegraph is removed (siblings, plus unrelated top-level keys like plugins, survive); only process entries matching codegraph[.cmd] + args includes "prompt-hook" are dropped, with empty groups/events pruned. hooks.enabled is deliberately left in place on uninstall — its provenance is unknowable and a lingering enabled: true with no config-file hooks is inert.

Testing

• The existing target contract suite covers zcode automatically (install/idempotency/sibling-preservation/uninstall/printConfig); the sibling-preservation case gained a zcode branch for the nested shape.
• New ZCode-specific tests: nested MCP shape + no type key, hooks.enabled + process-shaped hook, byte-identical re-runs, opt-out round-trip, Claude-style entry normalization, sibling-hook preservation across install/uninstall, cross-platform command spelling, local-scope paths, and printConfig output shape.
• npx vitest run on the three installer suites: 272 passed; tsc --noEmit clean; full npm run build succeeds. Live-checked install -t zcode -l global -y against an isolated $HOME: all three layers written exactly as documented above.