fix(stereo-split): detect localized timestamp bursts

[shark0F0497]

Summary

Follow-up to #173.

• detect burst-and-stall patterns inside rolling 300-gap windows instead of diluting them across the full recording
• repair files whose beginning is bursty even when the remaining timeline is healthy
• preserve the existing requirement that a window must contain both at least 20% short gaps and a long stall
• keep healthy-file no-op behavior unchanged

Root cause

capture (4).mcap contained 14 long stalls and a bursty prefix, but its full-file short-gap ratio was 17.82%, just below the previous 20% threshold. The healthy tail therefore caused the entire file to be classified as healthy.

Verification

• 34/34 stereo-split tests pass
• new pure-computation regression test passes in 0.001s
• real capture (4).mcap now reports log_bursty,publish_bursty
• first 1,000 output frames: max gap reduced from 2137.928 ms to 32.020 ms
• complete output: P50 16.003 ms, P99 31.992 ms, max 32.076 ms
• log/publish retain one constant offset from the embedded header across all 8,261 frames

Review

• Standards review: no hard violations
• Spec review: no findings