Enrichment looks up albums whose tags are poor and writes what it proves into the library only — the tracks columns, plus a record of which fields it set in track_meta_arricchita. Since 2.3.1 the music file is never opened for writing, so its mtime never changes and the next scan does not see it as modified.
Who is a candidate
A file row (path IS NOT NULL) with no mb_recording_id, and at least one of: artist «Artista sconosciuto», album «Album sconosciuto», no cover, or source = 'catalogo'; and whose enrich_status is NULL, or needs-review/error older than RITENTA_INCERTI_MS = 7 days, or no-match older than RITENTA_NESSUNO_MS = 30 days. ok and undone are final. Candidates are grouped by album_key, most recently added first, LOTTO = 12 groups per pass; each group is read whole, because the track count is evidence.
Phases
candidati() lock read a batch of album groups
decidi() no lock search MusicBrainz / iTunes / Deezer, decide
applica() no lock download covers into the cover store
registra() lock write rows, provenance, aggregates (one transaction)
The decision (pure, in rymembra-domain)
Every decision has three outcomes: Applica writes, DaRivedere (needs-review) writes nothing and retries in a week, Nessuno (no-match) writes nothing and retries in a month. An unknown axis (duration, artist) is neutral when ranking but never counts as agreement when deciding.
Album route (named album, more than one track): album_distance weighs album title 3, artist 3, track count 2, missing 0.9, unmatched 0.6, year 1, per-track title 3, duration 2, index 1. decide_album applies only if album title ≥ TITOLO_FORTE 0.85, artist ≥ ARTISTA_FORTE 0.80, exact track count, at least 4 of 5 tracks with a strong title, no track off by more than VETO_DURATA_SEC = 15 s, no variant mismatch, and distance ≤ DISTANZA_APPLICA 0.12; ≤ 0.35 is DaRivedere.
Track route: candidates from three independent sources (MusicBrainz, iTunes, Deezer) are scored 0.4·title + 0.3·artist + 0.3·duration. Below SOGLIA_CANDIDATO 0.65 → Nessuno. Applica requires a strong title, an anchor (artist ≥ 0.80 or duration within 4 s), consensus of two different declared sources on the same recording, no contradiction (artist or album similarity < 0.35) and no variant mismatch.
Variants (live, remix, karaoke, acoustic, …) are compared, not stripped: two different variant masks are a veto, so a karaoke base of the same length cannot take the original's place.
plan_write then fills empty or fallback fields always, replaces existing values only on Applica, never erases a value the candidate does not know, never replaces a genre, and only adds identifiers (never over a different one). Fields whose provenance is a manual correction are never rewritten.
Network and cache
Requests go through rymembra_meta::Fornitori (20 s timeout), paced per service — MusicBrainz 1.1 s between requests, Cover Art Archive 250 ms, iTunes 350 ms, Deezer 220 ms — with a circuit breaker; see Rate limits. Responses are cached in the enrich_cache table (search 7 days, release 30 days, «not found» 3 days) through a secondary connection. If no source answers, the group is non_raggiungibile: nothing is written and nothing remembered. Covers come from the Cover Art Archive (front-500, release group, then up to 3 releases), iTunes (600 px) or Deezer, at most 8 MiB, magic-byte checked, stored with CoverSource::CoverArtArchive / Provider.
The thread
rymembra-arricchimento: first pass after 90 s, then every 30 min, and after scans/downloads once a 2 min burst is over. It never takes the player lock and holds the library lock only per group. enrich.auto (default on) switches it; events arricchimento:stato, :avanzamento, :esito report progress and the counts applied / held back / without match.
Forgetting and the legacy exit
arricchimento_annulla → enrich::dimentica: re-reads tags from the untouched files, clears what enrichment set, sets status undone, re-applies manual corrections, and switches enrich.auto off so the next pass does not redo it. arricchimento_riporta_nei_file is the only code in Rymembra that still writes tags into files: it restores, from enrich_undo, the tags that versions up to 2.3.0 overwrote. It is a temporary exit that the CHANGELOG says will be removed.
The effective metadata of a row is the result of three layers, applied in this order:
raw tags → track_meta_arricchita → track_overrides
(what the file says) (what a catalog found) (what you decided)
Raw tags are written by the scan (library::update_track) every time it sees the file changed.
Enrichment writes its fields and, in track_meta_arricchita, the value the file had underneath (Sotto). After an external re-tag, each enriched field is kept only if the file still has the value it replaced (arricchimento_sopra_i_tag) — a new tag from another program wins.
Corrections (metadati_correggi → incerti::correggi) are stored as JSON in track_overrides and re-applied after every rewrite of the row by incerti::riapplica. A correction recomputes track_key (other devices must find the track under its corrected name) but nevercontent_key.
Corrections never touch the file, so its mtime is stable and it is not re-read every scan. If the file disappears, the correction is parked in correzioni_orfane by content_key for 90 days and restored when the file comes back (Removal). Profile imports apply corrections before anything else, so the rest of the imported data finds tracks under their corrected keys.
meta_origine / meta_problemi hold per-field provenance (Tag, TagRiparato, Percorso, Ripiego, Manuale) as JSON, and meta_salute a summary used by the review list (metadati_incerti). As noted in Tags, the scan does not yet fill these columns, so the review list is empty; corrections work regardless.
One LRC parser exists, rymembra_domain::testo; the window receives already-parsed lines (TestoIpc: lines with ms, optional per-word times, secondary lines such as translations, or a plain text, or «instrumental») and never sees raw LRC.
The source chain
a sidecar beside the file — <name>.a2.lrc (word timing) then <name>.lrc — re-read every time so edits in a text editor show up; bytes are decoded with codifica::leggi_byte (UTF-8, then windows-1252/1251, Shift_JIS, GBK);
synced tags in the file: SYLT (ID3v2), SYNCEDLYRICS (Vorbis/APE), read on demand, not by the scan;
the lyrics row, keyed by track_key (so two formats of the same song share it, and moving the file keeps it): downloaded or hand-timed lyrics;
the plain tag, tracks.lyrics.
What is found in 1, 2 and 4 is copied into 3, which acts as a cache. Fonte records where it came from: sidecar, tag, lrclib, mano (timed by hand, never overwritten) or none.
LRCLIB
When nothing local exists, Rymembra asks LRCLIB — public, no key. A request is sent only for the track playing and only while a lyrics panel is open (testi_pannello counts open panels), unless you start «Fill the library» (testi_riempi, batches of LOTTO = 50, stoppable). testi.rete switches the network off entirely. Requests are paced at one per 250 ms by the LRCLIB cadence. The exact get (artist, title, album, duration in seconds) is tried first, then a search, retrying after 500 and 1,500 ms; up to 20 candidates are scored by testo::scegli: duration veto (15 s), title veto below SOGLIA_TITOLO 0.55, accept above SOGLIA_SCELTA 0.66, with a bonus for synced text. A track with no answer is retried after RIPROVA_MS = 14 days, or 1 day after a network failure.
testo_candidati lists all answers with their first lines; testo_scegli pins one (lyrics.scelta), testo_rifiuta / «none of these» records rejections (lyrics.scartati) so they are never proposed again. While a track plays, precarica_prossimo (thread rymembra-testo-dopo) fetches the next track's lyrics if a panel is open.
Fit and display
verifica_durata returns Buona, Sospetta (text ends more than CODA_SOSPETTA_MS = 75 s before the file) or Fuori (times beyond the end by more than 10 s). The active line is found by binary search; a line lights ANTICIPO_MS = 120 ms early (eye travel), and progress within it is computed from the real time. An [offset:] tag is preserved, not applied. Two buttons shift a constant drift (testo_scarto, stored in lyrics.offset_ms), and the output latency setting moves the reported position (Output). Testo.tsx writes progress into a CSS variable through a ref, per line and per word, and shows three filling dots for instrumental gaps of 3 s or more.
Writing
Lyrics live in the database. A .lrc beside the track is written only if testi.accanto is on (default off). testo_pubblica can publish hand-timed lyrics to LRCLIB, which requires solving its proof-of-work challenge (TENTATIVI_MASSIMI = 50,000,000 nonces).
Sincronizza.tsx is a tap editor: paste the lines, play, press Space at the start of each line; optionally a second pass taps each word (single-word lines are skipped, their time is known). Human taps are late by a reaction time that varies, so the backend fixes them against the audio.
Onsets
rymembra_play::attacchi decodes the whole file at FREQUENZA = 22,050 Hz mono and computes spectral flux (increases only) between bins 14 and 186 (≈ 300–4,000 Hz, where voice attacks live), frames of 1,024 with hop 512 (≈ 23 ms resolution). A frame is an onset when its flux exceeds 1.5× the local mean over ±20 frames, with at least 50 ms between onsets. The onsets never cross IPC.
Snapping
testo_aggancia passes taps and onsets to the pure testo::aggancia:
each tap snaps to the nearest onset within FINESTRA_PRIMA_MS = 250 ms before or FINESTRA_DOPO_MS = 120 ms after — asymmetric because tapping is late, not early;
an onset can be used by one tap only, so two close lines never collapse onto the same time;
the median of (tap − onset) is the person's reaction latency in that session, subtracted from taps that found no onset (median, because one tap two seconds late would ruin a mean);
the result is kept monotonic.
Without onsets (silent file, undecodable) the taps are returned unchanged. The result is saved with testo_salva as source mano; with word times it is extended LRC, exported as .a2.lrc beside the ordinary .lrc when sidecars are enabled.