Tune the Jev planner, and choose from custom lists

[brynary]

Stacked on #22. The Jev planner gets finer intent, candidate, and acceptance rules, and handles lists and dropdowns built without a native <select>. One commit per change, with a re-eval after each batch.

Commits

Commit	Change
c82eba1	Fuller intent question. Ten action families, each with a description. A split vote between click and select, double-click, or press is sure when the pair reaches 0.85. The same request asks for the key, the mouse button, the checkbox end state, and whether typing ends with a suggestion. Right and middle clicks, suggestions, and a click that would undo a checkbox already in the asked state go to the model.
5e863ac	Fill text. A lone placeholder is the text to type. Otherwise a fill_value question lets Jev choose among quoted strings and placeholders. For unquoted text, a small model call that sees only the instruction reads the text, and Instruction::span requires it to be the instruction's own words. The planners share one Arc<ModelClient>.
b812e61	Candidate tiers. The action's own view, then every element with a name or text. A list over 40 is first cut to its 30 best BM25 matches. If Jev rejects the cut list, the full list is asked, split into parts of up to 254 that nominate their top three in parallel. An exact quoted name is confirmed with one small request.
7efca9f	Acceptance rules. A clear-leader rule: ≥ 0.5 confidence, none-of-these ≤ 0.1, and a leader ≥ 0.6 and 2.5× the runner-up. Held picks: an accepted pick with none-of-these > 0.5 waits while the next tier tries. Copies of one control in one row or item share their vote.
acaecb8	Fix: describe a table element by its table's caption or name and its column header. Commit 1 let "pick March 14" reach a Jev pick for the first time, and Jev clicked February 14 in all 6 runs: the two "14" buttons differed only by position.
8f9e159	Custom lists and dropdowns. When the named option shows, click it: first elements with option roles, then list items and clickable elements, confirmed by Jev. Otherwise pick the control that opens the list as step one, and click the named option in step two, which no longer always goes to the model. Each pick asks its own question (target, opener, option). Also fixes a description bug: an element with its own text no longer takes its neighbour's text as a label.
7dbaa0b, 197f63c	Eval results after commits 1–5 and after commit 6.

Results (local set, 48 tasks, -n 3)

	pass	mean ACT	median	p90	model calls	by Jev	$/task
Flash-Lite	0.99	1.14s	0.99s	1.81s	1.09	-	$0.0009
Luna	0.99	1.71s	1.30s	3.23s	1.10	-	$0.0004
Jev + Flash-Lite, #22	0.97	0.85s	0.59s	1.59s	0.49	0.58	$0.0007
Jev + Flash-Lite, commits 1–5	0.98	0.74s	0.51s	1.32s	0.38	0.76	$0.0004
Jev + Flash-Lite, this PR	1.00	0.66s	0.47s	1.13s	0.28	0.84	$0.0004
Jev + Luna, #22	1.00	1.31s	0.64s	2.94s	0.50	0.56	$0.0004
Jev + Luna, commits 1–5	1.00	1.00s	0.47s	2.38s	0.38	0.75	$0.0003
Jev + Luna, this PR	1.00	0.83s	0.47s	1.79s	0.28	0.83	$0.0002

• Both Jev variants pass every run, and Jev chooses 83–84% of actions.
• The mean ACT time falls 42% (Flash-Lite) and 51% (Luna) against the model alone.
• Jev + Luna costs half as much per task as Luna alone.
• The tasks where Jev still never decides are split intents: "cancel my membership", "show me my invoices", "reply to Grace Hopper's comment", "set the volume to 75". These fall back to the model by design; authors should name the action instead.

Tests

• Acceptance: unquoted fill text through the small model call, which never sends the page; the broad tier; copies sharing a vote; a custom listbox; and a two-step custom dropdown with no model calls.
• Unit: the intent merge, fill_value, span, splitting long lists, exact names, the acceptance rules, table places, option tiers, and neighbour labels.
• mise run check passes. SPEC §7.4 and CHANGELOG describe each change.

🤖 Generated with Claude Code