fix(annotator): the suggest tool survives a class switch; a select option is two lines

[JArmandoAnaya]

Closes #472.

Two things a person had to work around: a select that squashed the label it was showing, and a
tool that switched itself off when you picked a class to use it on. Frontend only — no wire, no
route, no kernel, no docker.

A select option is two lines when it is an identifier plus its facts

SelectItem takes a meta. The children stay the identifier at the label role; meta goes
underneath at the meta role in muted-foreground. Radix renders the selected item's own
ItemText into the trigger, so the closed control and the open list are the same two lines by
construction — there is no second copy of the layout to keep in step, which is the reason this
is a primitive variant rather than one screen's styling.

The trigger is min-h-9 with py-1 instead of a fixed h-9. A one-line option measures exactly
36px still (22.4px of text plus 8px of padding plus the border is under the floor), so every
select that shipped before this is unmoved; a two-line one grows. Nothing truncates and nothing
ellipsises — half a model id is not a model id, so a long one wraps.

Custom model… gets a meta line too, so no row in the list is a different height from its
neighbours.

The styleguide gains the specimen and DESIGN.md gains the rule, next to the geometry picker's
grouping rule it sits beside.

The suggest tool survives a class switch

This is #451's behaviour, deliberately reversed, and that is the part worth reading. Slice 3b
held the mode beside the active class and discarded it whenever that class moved, on D2's
"switching tools discards" — moving the active class being how this build spells switching
tools. Directed (Armando, 2026-08-09): arming is a decision about how to work and picking a class
is the next thing somebody does, so a class switch ends what is pending and not the session.

withClass in core/interaction/suggestion.ts is the whole rule, and it is one function because
the three readings are one transition over a different argument:

The class moves to	What happens
another suggestible class, nothing pending	swapped in; the next click asks under it, in its geometry and its colour
another suggestible class, a preview showing	the preview and its clicks are discarded; the tool stays armed
a class that can hold nothing	the session parks: labelClass is null, no press is diverted, the armed intent is kept
— asset switch —	disarms and discards, through the key={asset.id} remount, unchanged

The preview is discarded because it was answered under the old class's allowed_geometries;
accepting it under the new one could write a kind that class does not admit, which is the same
argument D3 makes about which classes may be asked in the first place.

The serial keeps counting across all three, for cleared's reason: an answer to the ask that
was in flight when the class moved must not repaint a preview belonging to a class nobody is on.
withClass returns the state by identity when the class did not really move, so the page's effect
folds through unconditionally without a render or a discarded preview per re-selection.

Parked is a state with a sentence, not a control going quiet

• The strip button is dimmed with its reason and stays lit — both are true at once: the tool
  is armed, and it cannot act. Hidden is the schema answer (D3's third case, unchanged);
  disabled-with-reason is the class answer, because that one comes back the moment the active
  class moves, and a button that vanished and returned as somebody worked down the class list
  would be describing a project that keeps changing.
• The panel gains a sixth reading: it names the class, says what to pick, says the tool is still
  armed, and carries the way out — since the control it would otherwise be is the dimmed one. No
  Esc chip beside it, because the chord is a substitution the canvas makes while something is
  pending and a parked session has nothing pending; printing the key would be printing one that
  does nothing.
• Parked outranks the connection blocker. "Getting the model ready…" over a class that could
  never hold the answer reports progress towards something that is not going to happen and hides
  the one choice the person can change.
• The canvas is handed null while parked. The class that parked it may still be drawable — a
  lane is a polyline — and a tool that swallowed those presses would have stopped being parked
  and started being broken. There is a test that draws one.

Tests

Core (suggestion.test.ts, +10): identity when the class did not move; armed survives the swap;
the preview discards; the serial keeps counting and the in-flight answer is dropped; parking;
re-arming on the class that unparked it, not the one left behind; parked-to-parked is identity;
nothing is written without a class. paint.test.ts: a parked session draws nothing.

Adapter/product: suggestFlow.test.tsx gains the whole class-switch block driven through
AnnotationPage with a stubbed route table — including the asset switch, which needed a second
frame in the fixture, and a polyline class, which is what parks the tool. suggestPanel.test.tsx
gains the parked reading, its way out, its precedence over the blocker, and the no-class-at-all
sentence. toolPalette.test.tsx gains dimmed-with-reason and the press it refuses.
primitives.test.tsx gains the two-line option; inference.test.tsx gains it on the real form,
trigger and list; styleguide.spec.ts measures in a browser what jsdom cannot — the one-line
trigger on 36px and the two-line one taller.

The pinned test that moved. suggestFlow.test.tsx › disarms when another tool moves the
active class (D2) is gone, replaced by disarms when it is pressed again plus the block above.
It encoded the rule this PR was asked to change, and it is named here so the removal is a decision
rather than a disappearance.

Mutation verification

Each mutation applied to the committed tree, its anchor asserted present before and the
replacement after, run, then reverted by its exact edit with a clean tree confirmed.

Mutation	Result
The page's effect goes back to setSession(null) on a class switch	6 red in suggestFlow.test.tsx, led by stays armed on a class switch
withClass keeps the pending state — { ...state, labelClass } instead of { ...cleared(state), labelClass }	2 red in suggestion.test.ts + 1 red in suggestFlow.test.tsx, all discards a preview…
SelectTrigger back to a fixed h-9	1 red: grows rather than clipping, and leaves a one-line option where it was
The canvas is handed the parked session (suggesting, not diverting)	2 red, led by leaves the canvas alone while parked

The two the issue asked to be named are the first two rows.

Gate

Run in stages under the harness's ~10-minute command ceiling; pytest split by the directories
ls tests/ reported at run time. The machine was free (load average 1.45 on 20 cores), so this is
the full gate split and not the saturation fallback. Every exit code verbatim.

Stage	Exit
pytest tests/architecture	0
pytest tests/cli	0
pytest tests/examples	0
pytest tests/formats	0
pytest tests/inference	0
pytest tests/jobs	0
pytest tests/kernel	0
pytest tests/mcp	0
pytest tests/packaging	0
pytest tests/server tests/test_versioning.py	0
ruff check .	0
ruff format --check .	0
mypy src/visionset	0
lint-imports (4 kept, 0 broken)	0
pnpm -r build	0
pnpm test (node script gates + vitest)	0
pnpm -r lint (incl. all three annotator boundary gates)	0
check.sh generated (openapi drift, client drift, mcp reference, version sync)	0
playwright test (annotator + app e2e, chromium)	0 — 234 passed, no flakes
playwright test -c playwright.cycle.config.ts	0 — 1 passed

Frontend totals: annotator 914 passed (34 files), ui-core 844 passed (46 files).
tests/scripts holds only .mjs node gates, which run under pnpm test, so it is absent above
rather than reported as exit 5. No baseline-proof exception was used or needed.

Boundaries

Nothing in frontend/annotator/src/core/ gained a React import, a DOM global or a DOM type in a
signature — the mode change lives in the core because that is where the mode lives, and all three
gates are green. No route, tab, screen, nav entry or entry point moves, so the sitemap is
unchanged. openapi.json and the generated client are untouched, and the drift gates say so.

Found, not fixed

• The connection list is still fetched the moment the tool is armed, which now includes a session
  that parks before it ever asks. It is one request against a workspace-scoped list, and
  un-parking is one click away, so prefetching is the better of the two — recorded because the
  "a job nobody suggests on makes no inference request at all" discipline it sits under is worth
  keeping visible.
• #466 (the four inference request models do not forbid unknown fields) is untouched.