Improve multi-device sync continuity and recording-overlap retention

[yoshihito-saito]

Summary

Stable recordings could lose usable data when temporarily weak correlations or small, unlocalized lag changes were treated as synchronization failures. Retention rates also penalized differences in recording length.

This change:

• Retains weak-correlation intervals when reliable observations on both sides support the same continuous affine clock model. Inferred support is recorded separately from measured anchors.
• Uses robust model inliers for offset-step statistics and avoids unnecessary terminal segmentation when qualified observations support continuity.
• Aligns post-merge QC filtering and window duration with synchronization, and requires non-overlapping windows for repeated-failure and correction decisions.
• Calculates retention over the common raw recording overlap while preserving full-output statistics. Internal invalid intervals remain in the denominator.

The existing 1 ms tolerance, missing/repeated-data handling, and independently localized discontinuity boundaries are preserved. Versions advance to pipeline v24, sync v16, and manifest schema 20.

Validation

Reprocessed three complete recordings with synchronization, merge, duplication scanning, and post-merge QC:

Dataset	Common-overlap retention	Invalid time within overlap
WT2 Day7 indoor	100%	0 s
WT2 Day14 outdoor	99.8947%	4.4608 s
WT4 Day14 outdoor	100%	0 s

• All six pair models passed.
• Day7 recovered 16.26965 seconds of usable data before denominator normalization.
• Across 5,019 QC windows: 4,828 passes, 191 low-confidence measurements, and no reliable lag exceeding 1 ms. Maximum reliable lag: 0.35 ms.
• All 89 source files retained their sizes and modification times. Generated outputs remained local.
• No test suite was run; PC-time fitting, IMU processing, and event export were outside this validation.