2026.05.09

[tsirysndr]

[2026.05.09]

Fixed

• DSP compressor divide-by-zero crash on x86_64 (SIGFPE in get_att_rls_coeff) — added release > 0 guard in compressor_update() mirroring the existing attack > 0 guard; ARM64 silently returned 0 on integer divide-by-zero while x86_64 faulted; also added function-level guards in get_att_rls_coeff and get_lpf_coeff for zero rc/fs/rc_units parameters, and an early fs <= 0 return in compressor_update for uninitialised output frequency
• Startup hang on second+ launch — FTS5 backfill WHERE NOT EXISTS (SELECT 1 FROM fts_table f WHERE f.id = t.id) forced an O(N) full scan per row (O(N²) total) because id is UNINDEXED in FTS5; replaced all four backfill INSERTs with an uncorrelated WHERE NOT EXISTS (SELECT 1 FROM fts_table) which SQLite short-circuits at the first row (O(1) for non-empty tables)
• Library startup blocked indefinitely on repeated runs — SQLx hangs when re-executing CREATE VIRTUAL TABLE IF NOT EXISTS on an existing FTS5 virtual table; fixed by checking sqlite_master before the migration and skipping it entirely if track_fts already exists; same guard added for dedupe_genres (checks UNIQUE constraint on genre table)
• FTS5 and dedupe_genres migrations ran in slow DELETE journal mode — PRAGMA journal_mode=WAL was set only after all migrations; moved to SqliteConnectOptions::journal_mode(Wal) so WAL is active from the first connection
• FTS5 index migration moved to a background tokio::spawn task so startup is non-blocking; dedupe_genres (schema DDL) remains synchronous with an O(1) skip guard
• cpal PCM sink: audible silence gap at the start of every track on Linux — sink_dma_start() previously stored the first chunk in pcm_data/pcm_size and then called pthread_create, leaving the ring empty for the 1–5 ms thread-creation window; fixed by pushing the first chunk synchronously via pcm_cpal_push() before spawning the writer thread so the ring is pre-filled when running=true is set; the writer thread now picks up from chunk 2 onwards; also added !r.running early-exit to the f32 cpal callback (mirrors the existing i16 guard) and reset resampler state (cur_valid = false, phase = 0) in pcm_cpal_start() to prevent interpolation artefacts from the tail of the previous track

Full Changelog: 2026.05.06...2026.05.09

---

This discussion was created from the release 2026.05.09.