Recover stalled live pipelines

[ykamakazi]

Summary

• add a server-owned watchdog for process liveness and completed HLS segment publication
• restart only high-quality output when it fails, while keeping low playback and the upstream ingest alive
• restart the dependency chain for ingest or low-output stalls after the previous upstream reader exits
• preserve retained DVR entries across restarts with appended playlists and HLS discontinuities
• add a 30-second HTTP read timeout plus bounded restart cooldowns and budgets

Recovery behavior

Failure	Recovery
High-quality output stalls or exits	Continue low playback and restart only high-quality output
Ingest or low output stalls or exits	Replace the dependency chain under the same session ID
Repeated failure	Limit each recovery class to three attempts per five-minute window
Missing prior playlist	Use epoch-based segment numbering to avoid overwriting retained files

Validation

• .venv/bin/ruff check .
• .venv/bin/basedpyright
• .venv/bin/pytest -q (715 passed)
• real FFmpeg restart coverage verifies retained entries and discontinuity insertion