fix(panel): stream-interval 8 default — interval 1 was the warm-turn collapse

The "native MTP collapses on warm turns" hunt (40.4 -> 17.4 -> 14.8 t/s
run-over-run in the app) ended at the delivery path, not the engine:
with stream-interval 1 the renderer re-renders a growing markdown
document per token, TCP backpressure stalls the engine's emit loop,
and the MTP runtime cost gate reads the stall as speculative-decoding
cost (measured 98.2ms/token "MTP cost") and falls back to AR.

Exoneration was systematic: the identical engine, prompts, and restore
depths over the raw API hold 26-31 t/s at turn 4 with 1537 tokens
restored, on both the L1 and L2 lanes, with fence and prefetch toggles
changing nothing. Switching the app session to stream-interval 8
recovered warm turn 3 from 14.8 to 28.9 t/s with no engine change.

Ship interval 8: new sessions default to it, the create/merge
sanitizers lift the legacy value 1 (any other explicit choice is
kept), and buildArgs emits 8 for saved rows still carrying 1 so
existing sessions are covered without a rewrite. dots3 sessions have
shipped interval 8 all along for the same reason.