feat(fleet): rebuild the dashboard's start and status logic

[outofcoffee]

Rebuilds the fleet dashboard's start reporting, refresh ordering and tile rendering so a node has one description of its state rather than two that can contradict each other, and records the CLI's UX conventions in a spec of their own.

Summary

• A start reports a fleet.StartPhase — what it is doing, when that began, when the next attempt is due — instead of a line of text. Each transition replaces the value outright, so a situation the start has moved on from cannot be left on a tile. This is what the reported defect was: instance no-capacity; retrying in 120s sitting beside a refresh reporting the node running.
• fleet.RenderPhase(phase, now) is the only place a phase becomes text, and it reads the clock at draw time. A wait counts down, a boot counts up, and no rendered number can be left at a value a transition has passed. spinloop remote start renders the same phases to stderr, so the tile and the CLI cannot word one situation differently.
• Every reading carries the time its own call returned (fleet.NodeResult.At, stamped in the fan-out), and the board draws a reading only when it was taken later than the one on screen. This replaces the fastGen/slowGen counters, which ordered rounds against each other but not against an action completing — so a round issued before a start finished and landing after it would repaint the node's pre-action state.
• A reading that ages past three of its node's own intervals shows its age and drops to the unknown health tier, rather than being drawn identically to one just taken.
• A node with an action in flight is read on the short interval whatever its kind, and returns to its own cadence once the action settles. Its neighbours keep theirs throughout.
• One function, dashNodeView, produces both a panel's lines and its health tier from arguments alone. Every pairing of a start's phase against a reading is now enumerable in a test, including the one that produced this defect.
• Tiles gained a coloured header bar, an in-flight verb gained a spinner, and the board gained a title bar of its own. The selected panel's border is the brand mint rather than the amber it shared with the attention health tier.
• New cli-ux spec: one accent used only for the tool's own chrome and never for a state, a stdout a program consumes carrying nothing else, decoration only where there is a terminal to draw it on, errors that name the fix, help as a lowercase imperative phrase, British spelling, what a long operation must keep saying, confirmation before a destructive action, and the full-screen affordances a web page has no equivalent for.
• cmd/spinloop/palette.go holds the colours and the spinner every surface draws from. The ten braille frames were declared in both fleet.go and dashboard_model.go, and the five state colours in both fleet.go and the dashboard's health switch; each is now one definition.

Implementation details

Why a value rather than a better-worded line. The quick fix in 03718af corrected the reported symptom by writing a line at the transition that had been missed. That is correct only while every write site covers every transition, and an omission produces no error and no test failure — the output is a well-formed line carrying an out-of-date value. A phase holds one value per start and each write replaces the previous one, so a superseded situation is not retained by construction.

Why timestamps rather than a per-node counter. A counter orders readings equally well, but carries no age, so showing how old a reading is would need a second field beside it. At supplies both the ordering and the age.

The read cadence is per node, not per group. The group busy-guard stays — no second round starts over one still in flight — but each node carries its own next-read time, which is how an acted-on cloud environment is polled every 2s while its neighbours stay on 60s. The round's context deadline remains its kind's interval, so shortening the cadence during an action does not shorten what a signed control-plane call is given to answer in.

internal/remote is untouched. remote.Start's existing progress and onState callbacks already supply everything the phases are built from; fleet.StartPhases maps that pair onto the phase stream, and both the dashboard and the CLI drive it.

Where the shared palette sits in the history. The rename that consolidated the colours and the spinner happened in place across three files, so the commit that introduces palette.go is the one carrying the dashboard work rather than the one adding the spec — splitting it the other way would leave the first commit unable to compile, and fleet.go's state colours colliding with the shared set.

One behaviour-visible wording change outside the dashboard. spinloop remote start's stderr lines are reworded, since they now render phases. Nothing parses them — the eval-able exports go to stdout and are unchanged — and the CLI tests that assert on them are updated.