serve: hot-window time cursors resolve against the hot tail, never cold master
a time_us cursor at/after the oldest retained hot frame now starts on the hot tail at indexForCursor. previously resolveCursor sent every time cursor through resolveTimeToSeq, which only consults sealed segments — a near-now resume landed at the last sealed-segment boundary and stranded the subscriber on the cold reader, delivering read-batch bursts at block-seal cadence indefinitely (the jerky resume coral and bufo-bot hit after the 2026-08-08 proxy restart; every zat reconnect carries a time_us cursor). upstream's replay engine walks sealed segments then the active segment before handing to live; the hot tail is stream's port of that walk, so hot-first resolution is the faithful translation, not a divergence. known-open follow-up (documented): genuinely cold resumes still trail seal cadence after catch-up — the coldBatch seam handoff does not fire at the tip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>