Interactive auto-labeling's in-flight feedback is in the wrong place and has no delay threshold

[JArmandoAnaya]

The suggest tool arms with S, and a click on the canvas asks a segmentation model to propose a
shape for whatever sits under the point, through POST /inference/suggest. It shipped as slice 3b
of the inference work, in the pull request that added the interactive tool and its ephemeral
preview (#451).

While that request is in flight the editor does say something, and this is worth stating plainly
because it changes what needs building: SuggestPanel has an asking reading that renders a
spinning icon, the title "Looking at that…", and the sentence "The first click on a frame is the
slow one — refining after it is quick." So the feedback exists. Two things about it are wrong.

It is in the wrong place. That card lives in the notice column at the top-right of the editor.
The person who just clicked is looking at the point they clicked, in the middle of the picture,
where nothing happens at all — no marker changes, no cursor changes. On a cold start, which is
model load plus inference on the CPU and can run well over a second, the canvas is silent for the
whole wait.

It has no delay threshold. The card flips the instant the session's status becomes asking, so
a warm response — which is most of them, because a segmenter reads the image once and answers every
later click from that reading — produces a brief flash of a different card and then flips back. The
refine gesture is the worst case: the panel is showing "A shape for …", the click replaces it with
"Looking at that…" for a fraction of a second, and it comes back. The cold-start sentence has the
same problem from the other end. It is true of the first click on a frame and it is shown on every
click, including the fast ones it is not describing.

The result is that the two states a person actually needs to tell apart — this is working and
this is broken — are not distinguished where they are looking, and the reassurance that is
offered arrives when it is not needed.

What this asks for is one timing rule governing both surfaces: nothing at all below a threshold, an
indicator anchored at the click point above it, and the cold-start sentence held back until the
wait is long enough to be about a cold start.

Two constraints are worth recording before anyone picks this up, because both are deliberate and
neither should be undone in passing. Refining while a request is in flight is supported on purpose
— withPoint bumps a serial and keeps the previous preview so the canvas never blanks, and a stale
answer is dropped by that serial rather than by cancelling anything — so an indicator must
re-anchor rather than block the click. And the annotator package ships headless, with no Tailwind,
no design tokens and no stylesheet, so anything drawn on the canvas is drawn in the package's own
SVG vocabulary.

[JArmandoAnaya]

Decision (Armando, 2026-08-11): Interactive auto-labeling gets an in-flight
loading indicator anchored at the click point — supersedes any prior direction.

• Single timer, two thresholds: nothing before 200 ms; at 200 ms a pulsing halo
  appears around the click-point marker and the canvas cursor switches to
  progress; at ~1.5 s the hint card adds one prose line explaining that the
  first suggestion takes longer (model preparation / cold start).
• Once shown, the halo stays visible a minimum of 250 ms even if the response
  arrives earlier, to prevent a flicker on ~200–250 ms responses.
• The halo is a quiet neutral and is never brand coral. No third brand site
  is being added.
• The 1.5 s escalation is prose, not a bar: the wait is indeterminate and an
  indeterminate Progress renders as an empty track (no-bar-over-empty-track
  rule in DESIGN.md).
• prefers-reduced-motion: the halo renders static at 60% opacity instead of
  pulsing.
• Refining stays available while a request is in flight; the halo re-anchors to
  the newest point rather than blocking the click. Esc clears the indicator
  along with the points, which is the take-back the tool already has.
• Errors surface as prose in the hint card through the shared refusal
  vocabulary — the indicator never spins forever.
• The delay/min-visible logic is a pure, React-free state machine with its own
  fake-timer unit tests.

Three points were amended against what the editor already does, and the
amendments are recorded here rather than discovered later.

The escalation line already exists, so it is moved rather than written. The
panel's asking card carries "The first click on a frame is the slow one —
refining after it is quick." unconditionally today. That sentence is exactly the
cold-start explanation the 1.5 s threshold is for, so it goes behind the
threshold. No second sentence saying the same thing is added.

Refine clicks are not ignored. withPoint bumps a serial and keeps the
previous preview so the canvas never blanks while the next answer is in flight,
and answered drops any response whose serial is stale. Ignoring the click
would undo a deliberate design and make part of that serial machinery
unreachable, and the gesture table in docs/ui.md documents refine-as-you-click
as how the tool is used.

Nothing is aborted. There is no AbortController anywhere in the frontend.
The behaviour the abort was for — a discarded or superseded answer must not
repaint the canvas — is what the serial already guarantees, which is why
cleared() keeps the serial counting rather than resetting it.

One further note on where the halo lives. @visionset/annotator ships headless:
no Tailwind, no design tokens and no stylesheet at all, which is what lets an
embedder restyle around it. So the halo is drawn in the package's own SVG
vocabulary, as a named constant beside SUGGESTION_OPACITY, and pulses through
SVG animation rather than a CSS keyframe. It cannot name the primary token,
and it does not need to: what the brand rule actually forbids is a third coral
site, and this adds none.

[JArmandoAnaya]

Decision (Armando, 2026-08-11): The 200 ms show-delay is removed — supersedes
the two-threshold timing in the prior decision on this issue.

• The in-flight halo appears immediately when a point-suggest request is
  dispatched. Dogfooding shows inference never resolves under 200 ms, so the
  delay suppressed nothing and only added timer logic.
• The 250 ms minimum-visible floor stays: it is the anti-flicker guard, and
  with the delay gone it is the only one.
• The ~1.5 s prose escalation, Esc cancellation, the reduced-motion variant, and
  prose error surfacing are unchanged.

One bullet of the dispatched form named "in-flight click ignoring" among the
things left unchanged, and it is dropped rather than carried, because there is no
such behaviour to leave alone. The prior decision on this issue settled the
opposite: refining while a request is out stays available, the halo re-anchors to
the newest click, and a superseded answer is dropped by the session's serial. That
is unchanged by this decision too — it simply was never a rule about ignoring
clicks.

Two consequences worth stating, since removing a threshold changes what the
remaining ones mean:

The escalation counts 1500 ms from the dispatch, not 1300 ms from a halo that
now appears at zero. The old arrangement put the sentence on screen at 1500 ms
after the click, and that is the number worth preserving — it is a claim about how
long a person has been waiting, not about how long a ring has been drawn.

The floor counts from the dispatch as well, so hide fires no earlier than
250 ms after the request left, and immediately on any answer after that. With the
delay gone the floor is the only thing standing between a fast answer and a halo
that appears for two frames, which is why it is not removed alongside it.

The delay is removed rather than set to zero. A configuration knob for a
behaviour nobody wants is the same logic kept in costume, so the timing machine
loses its "started but not yet visible" state entirely and can no longer represent
it.