README and ARCHITECTURE were missing the dedupe command and claimed the tool
never deletes anything, but dedupe --yes removes duplicate playlist entries.
lyrics_analysis.json field list (tone/description) didn't match the actual
schema (register/concreteness/stance/summary/resonance/key_line/confidence).
lyrics_text.json is also written by report/apply as a side effect of language
detection, not only by the lyrics command. Sub-genre count corrected 34->36.
with_retries caught bare Exception with a type whitelist, masking real bugs
as network errors under retry/backoff - narrowed back to NetworkError.
fetch_api_languages was writing full lyrics text for every language into
lyrics_text.json via its track_supplement side channel, breaking the
documented RU-only invariant and letting analyze feed foreign-language
lyrics through the Russian-only prompt - now gated on normalize_api_language.
dedupe_playlists logged the found duplicate count as "removed" even when
a mid-loop network/API error cut deletion short.
THEME_PATTERN's relaxed regex allowed 2-char themes, contradicting its own
test; current_themes was hoisted out of the analyze loop so later prompts
never saw themes found earlier in the same run. Both reverted to intended
behavior, docs updated to match.
--extra alone now adds every extra column like before. Individual
--timestamp/--duration/--version/--album/--artist flags let you pick just
one group instead, and combine freely.
--extra now takes group names (timestamp/duration/version/album/artist/all)
so you can add just what you need, e.g. --extra timestamp. Also split
README command lists into core vs. extra sections for readability.
CSV can now keep source playlist order instead of forced playlist grouping,
and optionally include added_at/duration_ms/release_date/artist rating from
the API that were previously fetched but discarded.
report/apply/digest can now point at any Yandex Music playlist via a
share link instead of only the account's likes. Fetches the playlist's
tracks and artist genres straight into memory (no disk cache, nothing
persisted) and feeds them through the existing build_rows/render_digest/
apply_classification pipeline. apply supports --label to keep a foreign
playlist's tracks in separate genre playlists instead of merging them
into your own. Closes a gap where UnauthorizedError went uncaught anywhere.
Aggregates liked tracks into top artists, genres, decades and top albums
with fixed size ceilings instead of a per-track dump, so it can be pasted
into an LLM chat regardless of library size. Fully offline - reads only
the existing cache, no Client/network needed. Adds album_title/year to
the track cache (one-time re-fetch) to support the decade and album views.
README/ARCHITECTURE now bilingual (EN default, .ru.md cross-linked) per public-repo convention; added MIT license and Credits/License sections; tightened .gitignore against stray AI-tool files.
