Editor polish: asset-bounded suggest clicks, confidence only in preview, selected-only labels

[JArmandoAnaya]

Three editor-surface corrections, all frontend. No wire changes: openapi.json and the
generated client stay exactly where they are.

1. A suggest click outside the asset is a no-op

AnnotatorCanvas's input surface is the whole pane, deliberately — "a press in the margin
around the picture reaches the machine with the negative or past-the-edge coordinate it
deserves" — and imagePoint returns that coordinate unclamped. That is right for drawing,
where the interaction machine and the kernel both have an opinion about an out-of-frame
corner. It is wrong for a prompt point: the stage surround is not the asset, and a click
there has nothing for a segmenter to segment.

With the suggest tool armed, hit-test the press against the asset's rendered rectangle
before it becomes a prompt. Outside: nothing happens — no point recorded, no request sent,
no preview change. Positive and negative (alt) clicks alike. Correct under zoom and pan.
Clicks inside the asset behave exactly as they do today, including on top of an existing
shape.

2. Confidence renders only on the live suggestion preview

The number is a decision aid for accept-or-reject, and it stops being one the moment the
shape is accepted. So it stays on the preview overlay and leaves the editor everywhere
else. This narrows what shipped when a reviewer gained the ability to see which labels a
model drew and how sure it was (#512), which implemented criterion 2 of the work surfacing
model provenance and confidence in the review loop (#417):

• Canvas labels — class name only, no percentage.
• Panel rows — the figure goes; the Sparkles marker stays, and its tooltip carries the
  model_ref alone.
• The preview — unchanged, with confidenceLabel as the interactive suggest tool
  shipped it (feat(annotator): interactive suggest tool with ephemeral preview (#424 slice 3b) #451).

confidence and model_ref are unchanged on the wire and unchanged in storage. The number's
home is the batch review loop, where unattended pre-labeling is reviewed at volume (#425).

confidencePercent stays exported — it is the shared spelling the preview consumes and the
review surfaces will.

3. The canvas label renders only while its shape is selected

DESIGN.md states the rule under Annotation shape rendering (canvas): the class label
renders only while selected, anchored at the first vertex, never a pointer target. The React
adapter renders ShapeLabel unconditionally, so every shape on a busy frame carries its
class name at all times. Restore the rule at the shipped metrics. Unselected shapes: the box
alone. The list panel is the full inventory.

[JArmandoAnaya]

Decision (Armando, 2026-08-10):
(1) With the suggest tool armed, a click outside the asset's bounds is a complete
no-op — positive and negative (alt) clicks alike; no prompt recorded, no request sent.
(2) Confidence renders only on the live suggestion preview, where it aids the
accept/reject decision. Once accepted, the number leaves the editor entirely: no
percentages on canvas boxes, in the annotation list, or in tooltips. The sparkles
marker remains the sole provenance signal; its tooltip carries the model_ref only.
Confidence and model_ref stay stored on the wire unchanged — the number's home is the
batch review loop (#425), where it will be named. This partially supersedes #417's
criterion 2 surface as shipped by #512 — deliberate.
(3) The canvas label rule is restored as written: the class label renders only while
the shape is selected; the canvas stays clean otherwise. The list panel is the full
inventory.
cf. #417, #425, #451, #512.