Returns focus when a Plan picker closes

[johnnyt]

Refs: se-9cpa

What

Closing a picker on the Plan page removed the control that held the focus, and a keyboard fell back to the document - the same loss the pickers already fix on opening with JS.focus_first/0. Every way a picker closes now names the control the focus goes to:

The picker closes by	Focus goes to
Cancel, or a refused pick, in a row's picker	that row's "+"
Cancel, or a refused pick, in a slot's picker	the row of the block whose slot it was
a pick that lands	the new step's row, the first new one on the page for a recipe

The page adds a hidden plan-focus-return-<n> element on every close whose phx-mounted is JS.focus/1 naming that control, so the command runs once the patch has drawn it; each close adds a new element, so the command runs again. The row's "+" and sentence buttons gain ids built from the block id (plan-add-, plan-step-). The selector quotes the id ([id="..."]) because a block id is opaque. The design is written in the Plan page's moduledoc, beside the passage on focus when a picker opens, and in the comment above the panel's picker. assets/js/app.js is unchanged.

Evidence

• New LiveView tests in describe "focus when a picker closes", one per closing path: Cancel in a row's picker; Cancel in a slot's picker; a type pick in a row's picker; a recipe pick (the core "deadline" recipe, whose first new row heads the group's body rather than the armed gap); a pick in a slot's picker; a refused type pick and a refused recipe pick in a row's picker; a refused pick in a slot's picker; each close adding its own element; a close with no picker open moving nothing. Each asserts the hidden element's phx-mounted is JS.focus/1 on the control it names, and the path-specific ones assert that control is on the page. All run on the patron registration document.
• Each new test carries a one-line sabotage note. Mutations of the changed lines were run one at a time, each restored byte-equal from a copy before the next: every one turned at least one of the new tests red on an assertion.
• Full mix quality green on the committed tree: 644 of 644 tests, 85.0% coverage, Credo and Dialyzer clean. The commit was made on the byte-identical tree the gate ran on.

Review

Tier: gate (a reference-host page; no public function, option or wire shape a host reads; under the size threshold). In-turn review: the diff was re-read against the bead and its dated notes, which name Cancel, a pick and a refused pick. The closing handlers go through close_picker/1 ("insert-close" and both "insert" refusal branches) or commit_pick/2 (both "insert" success branches); a read-only page still refuses all of them in its write-gate clause before either runs. The timing claim in the moduledoc (the command runs after the patch has drawn its target) was checked against phoenix_live_view 1.2.12's client: phx-mounted is run from the patch's after-added callbacks, which fire once the morph has finished. The moduledoc's "refused pick" sentence was checked against the three refusal arms: an unresolved gap or type, a recipe that does not land, and a session refusal.

Provenance

• Focus moves from the server through JS.focus/1 on a freshly added element rather than through a client hook, so the page needs no JavaScript of its own. Decided under the bead's acceptance by the worker, 2026-09-30.
• After Cancel in a slot's picker the focus goes to the row of the slot's block rather than restoring the selection that arming the slot cleared; what the page shows after Cancel is unchanged.
• After a recipe pick the focus goes to the first row on the page whose block is new, because a recipe does not always write at the armed gap.
• The session-refusal arm of commit_pick/2 has no test of its own: no payload tried reaches it (every palette type was tried at one row's gap in the patron registration document, and every one landed). It sends the focus where Cancel does.