Don't fall back to whole-history reads when a windowed query is refused for size

In a viewer with a windowed waveform backend, one code path fell back to downloading a signal's whole history when the backend's edge query failed. On a very large dump, the backend refused edge queries on wide signals with 'larger than N MB (store limit)'. The whole-history read was capped by a similar limit, so each fallback burned 6-8 s and returned nothing. Fix: tell 'refused for size' errors apart from transient ones, cache the refusal per signal for the session, and answer 'unknown' right away instead of falling back. Also: an EXIT-trap cleanup that globbed $TMPDIR/<tool>.* deleted other concurrent sessions' temp dirs mid-run. Only remove paths your own process created, or give each run a private TMPDIR.