commands: a command says whether it applies, and menus grey what does not

[zmaril]

Stacked on #120. First of the Blender-review sequence.

Blender's UI is four ideas: operators, properties, poll(), and context. We had the first two and neither of the last two — every surface offered every command unconditionally. "Close area" sat live in the menu of the only area and failed on the click; the whole workspace.* family stayed live with one workspace.

The field

pub struct Command {
    ...
    /// Whether this could run against the workbench as it is — Blender's
    /// `poll()`. The params may be `Null`: a surface deciding whether to
    /// *offer* a command asks before it has any.
    pub poll: fn(&Workspaces, &Value) -> bool,
    pub run: fn(&mut Workspaces, &Value) -> Result<()>,
}

Most are always — there is no workbench where splitting an area makes no sense, which is exactly why it needs no poll of its own. The ones that aren't:

poll	commands
many_areas	join, join_into, ratio, swap
many_workspaces	workspace.close / cycle / switch / move

run checks it too

Not just the chrome. The chrome isn't the only caller — an agent reaches the same registry — and a command that can't apply should say so once, in one place, rather than failing differently depending on who asked. It refuses before touching anything, so a refused command is never a half-applied one.

Grey, not gone

A row that disappears teaches nothing; a row that's there and dim says this exists, not now. MenuItem::when(available), disabled on the wire type (ts-rs regenerated), and the shim draws it dim and inert.

The detail that matters: disabled rows are excluded from the list the arrows walk — the same list Enter fires and the letter accelerators match. Leaving them in would have let the keyboard run what the pointer refuses.

The library asks the host through a new can callback, because the host holds the registry. A host that offers no answer gets every row live, which is what these menus did before there was a poll to ask.

What I did not do, deliberately

The command palette is still a hardcoded list of five entries plus workspace switches. It never enumerated the registry at all — that's not a design decision anyone made, it's a stub. But enumerating it needs commands to declare their parameters, since most can't run without an area id or a direction, and that's its own piece of work rather than a line in this one. Flagging it rather than half-doing it.

Four new tests, including that run refuses what poll refuses (a poll the chrome respects and run doesn't is a poll an agent walks straight through) and that available shrinks with the workbench.

165 tests pass; fmt clean, clippy clean under -D warnings, shims and docs/reference.md regenerated.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XjCx5QszLXroGYh1YBxybw

[github-actions]

powderman preview

https://powderman-pr-121.fly.dev

The workbench, live: split/join/swap areas, the F3 palette, the widgets, themes, tooltips, the status bar.
The fleet and run execution are inert here (a Fly machine has no herdr) — this is for driving the interface.

On branches with the MCP server, point an agent at it:

claude mcp add --transport http powderman-preview "https://powderman-pr-121.fly.dev/mcp"

Updated on every push. Destroyed when this PR closes.