explorer: smooth failure-message flicker when tryEnterPointModeIfNeeded chases a failed cluster reload

[rdhyee]

Follow-up from #191 (issue #190 fix 2).

Symptom

If a cluster reload fails (e.g. flaky network on the res4/res6 parquet) and the user is at point-mode altitude, the failed reload's call site immediately calls tryEnterPointModeIfNeeded(), which paints Fetching sample index… over the just-painted Failed to load H3 res\${target} — try zooming again. failure message. The user sees the failure flash for a frame, then the helper's loading-state takes over.

This is rare (it requires a cluster-reload failure on a path the camera handler is in, with the user already below ENTER_POINT_ALT) and self-recovering — but it's noisy when it does happen.

Reproduction (synthetic)

1. Throttle network so res4/res6 parquet fetch fails.
2. Zoom from world view straight into point altitude (cluster reload fires, fails).
3. Observe phase-message strip during the transition.

Possible fixes

A few shapes worth considering:

1. Don't chase after a failed loadRes. The chase is a liveness fix — recovering from supersession that left us silent. A failure already wrote a phase message, so silence isn't the issue. Skip the chase when res8Ready === false:

   const ok = await loadRes(target, ...);
   if (ok) await tryEnterPointModeIfNeeded();

   Tradeoff: distinguishing supersession from failure at call sites where they're conflated (both return false). Could expand loadRes's return to a tri-state ('applied' | 'stale' | 'failed') but that's invasive.

2. Brief debounce inside tryEnterPointModeIfNeeded. If the helper is called within ~100ms of a phase-message write that wasn't a 'done' state, hold off the loading-message overwrite for that window so the failure remains visible long enough for the user to register it. Heuristic; brittle.

3. Two-line phase-message slot. Show last error (sticky for N seconds) on a second line instead of overwriting it. Bigger UI change; out of scope for a flicker fix.

Recommendation: (1) with the tri-state return is the cleanest, but only if other call sites also benefit. Otherwise, a one-line if at the chase sites checking ok is enough.

Acceptance

• Failure messages from cluster-reload paths persist long enough to be human-readable (or at least: are not immediately overwritten by Fetching sample index… in the failure case).
• Supersession recovery (the actual round-3 fix) still works — toggle source filter during cold-cache wait still ends up in point mode.

Refs

• explorer: 60–90 s 'no dots' window on cold-cache deep-link to point mode (DuckDB-WASM 1.24.0 falls back to full HTTP read) #190 — parent issue
• explorer: surface 'Fetching sample index…' during cold-cache boot→point-mode wait (#190 fix 2) #191 — introduced tryEnterPointModeIfNeeded()
• PR #191 final review comment — flagged this edge case