feat(speaker): add linear decay to speaker attribution

[odsod]

Summary

• Speaker detection relies on Google Meet's CSS indicator which only fires in brief 2-3s pulses with median 6.7s gaps — covering ~24% of actual speaking time
• Whisper segments falling in indicator-off gaps get no attribution (55% of sys lines unattributed today)
• Add configurable linear decay: when indicator goes off, speaker weight decays 1.0 → 0.0 over decayDurationS seconds
• Coverage() integrates the decay weight function (trapezoid area) instead of binary on/off
• DecayDuration=0 preserves existing binary behavior (no-op path)
• Expected to raise sys attribution from ~45% to ~75-80%

Test plan

• Existing binary coverage tests pass unchanged
• 7 new decay-specific tests (mid-decay, beyond-decay, reactivation, two-speaker stagger, threshold filtering, pre-window decay, zero-equivalence)
• mise run build passes (lint, test, tidy, binary)
• Manual: run recorder with "decayDurationS": 10 during next meeting, compare attribution rate