Perf: reduce ingest-pipeline CPU/memory churn concurrent with playback [stacked on #113]

[sanylax2]

Part 3 of the CPU/memory audit — the background pipeline that runs while music plays. Stacked on #113/#112 (touches the same stems files). All changes behavior-identical.

• Main-thread file opens eliminated. The ingest duration read and the launch-backfill duration read both did a synchronous AVAudioFile(forReading:) on the main actor — one per track at launch for a big library. Now off-main at utility priority.
• Stem-cache budget passes single-flight. Every queue move fired a detached full-directory scan; rapid skips stacked N concurrent scans racing each other's evictions. Now one runs at a time with at most one queued follow-up, which re-reads the latest protected set when it starts (so the newest play-queue neighborhood is what's protected — strictly safer than before).
• separateStems no longer pins the @Model track for the CPU-minutes separation run (weak capture; every write already guards on modelContext).
• O(window) memmove churn removed from separation. runStreaming's mix buffers and StreamingOverlapAdd.drain used removeFirst(n) per window (~MBs of copy traffic each, concurrent with playback); both now use a head cursor with amortized compaction. The existing OverlapAddTests pin output equivalence on Linux CI.
• Launch fetches stop hydrating beat grids. The queue-exhaustion/restore ID maps and LibraryCleanup's reference scans use propertiesToFetch, so beatTimes (hundreds of doubles per track) faults in only for tracks that actually play.

Device check

• Play with separation in flight: stems still complete and link (watch stems logs); skip rapidly — no behavioral change, stems of the current neighborhood survive eviction.
• Cold launch with a large library while playing: no UI hitches from the backfill; resume still lands correctly.

🤖 Generated with Claude Code

https://claude.ai/code/session_013TJoWkqg8bzkGdzxWhWjWP

---

Generated by Claude Code