fix: persist advance_in_progress to prevent cross-process session skip - #114

Summary

Fixes the premature track skip where a session started via vesper ask would advance to the next track within ~3 seconds.

Root cause

A cross-process worker race. vesper ask is a separate CLI process that clears Cider's queue and starts a track. During that ~1-2s startup, Cider reports is_playing=false. The long-lived server's background worker (separate process) sees the stop and advances prematurely. Two gaps allowed it:

• advance_in_progress was in-memory only — never persisted to the DB, so the server worker couldn't see the CLI was mid-advance.
• last_advance_at was persisted at the end of track play, so the cooldown check read a stale timestamp during the startup gap.

Fix (two layers)

1. Persist advance_in_progress to the session_runtime table (new column + schema migration). Set/persist it True at the start of _play_session_track, clear/persist it False at the end and in the exception handler. _effective_session_runtime reads the persisted value as authoritative so a stale in-memory flag can't override another process's in-flight advance.

2. Min-play-duration backstop (SESSION_MIN_PLAY_SECONDS = 10.0): don't auto-advance if the current track has played less than the threshold. This catches Cider's noisy playback reporting during buffering/startup independent of the cross-process coordination.

Also persists last_advance_at at the start of the advance (not the end) so the cooldown is measured from when the advance began.

Test commands run

uv run python -m compileall vesper tests
uv run pytest -q

All 285 tests pass (+4 new).

Note

This addresses the premature skip only. The separate "Lizzo cluster" issue (a single descriptive vibe term matching one artist's catalog with no artist diversification) is tracked for a follow-up.