targets: an area follows the selection unless you pin it

[zmaril]

Stacked on #117.

You were right that #109 was hard to tell apart from nothing. Here's why: the file browser could not open anything.

TreeView { children_of }     // ← no on_pick

on_pick was wired in exactly one place — the target-picker modal. So clicking a file selected the row inside that one editor and no other area heard. The only way to point a code viewer at a file was the chip and the modal. Which made #109 nearly untestable too: it marks the row an area is already open on, and almost nothing ever opened one.

The model

Blender's, because it's the right one: an editor follows the active thing, and a pin freezes it. Pin is opt-out.

select { kind: "file", value: "src/main.rs" }

Writes the target on every unpinned area whose editor points at that kind. Not resolved at draw time — written into the layout, so the document still says exactly what each area shows and every reader of it (the target chip, get_state, a saved workspace) needs no new concept. set_pinned { id, pinned } is the freeze.

Both are bus commands, so both got MCP tools — the parity test asked for them by name.

target_kind is not target_noun

The noun is the word shown to a person. The kind is what an area points at.

editor	kind
code, diff	file
files	folder
run	run
chart	chart

code and diff share one, so picking a file drives both — that's the Code and Changes arrangements finally working. The browser deliberately does not share it: its target is the folder it's rooted at, and re-rooting it onto the file you just clicked would collapse the tree you clicked it in. There's a test asserting the browser stays put.

Details worth knowing

• The pin is composed into every targeting header, not declared fourteen times — it belongs to the act of looking at something, not to any one editor. Uses the header-item slot from editors: the area header carries the editor's own controls #114.
• The half you split off keeps what it was showing. Splitting is how you get a second view of this thing; a fresh half that followed the next selection away would be the opposite of the gesture.
• select is navigational, so a click in a list doesn't bury the splits and joins on the undo stack.
• Selecting a kind nothing shows is not an error — picking a file with no viewer open is a reasonable thing to do. A malformed call still is.
• pinned is serde(default), so every layout saved before it reads as unpinned, which is what those layouts did. The test uses a real pre-pin document, because a default coming out the other way would silently freeze every existing workbench.

Trying it

Open the Code workspace and click a file. Then Changes — same click, both panes. Pin the code viewer, click another file, watch only the diff move.

161 tests pass (5 new); fmt clean, clippy clean under -D warnings, docs/reference.md regenerated with the two new commands and tools.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XjCx5QszLXroGYh1YBxybw

[github-actions]

powderman preview

https://powderman-pr-120.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-120.fly.dev/mcp"

Updated on every push. Destroyed when this PR closes.