I found that the `/get` endpoint often returns incorrect or unsynced
lyrics, while results returned by the `/search` more accurate options.
Thus I reversed the change in the previous commit to prioritize
searching first.
Adjust the base URL to perform a '/search' instead of attempting to
'/get' specific lyrics where we're unlikely to find lyrics for the
specific combination of album, artist, track names and the duration (see
https://lrclib.net/docs).
Since we receive an array of matching lyrics candidates, rank them by
their duration similarity to the item's duration, and whether they
contain synced lyrics.
Add explicit checks for lyrics texts fetched from the tested sources.
- Introduced `LyricsPage` class to represent lyrics pages for integrated
tests.
- Configured expected lyrics for each of the URLs that are being
fetched.
- Consolidated integrated tests in a new `TestLyricsSources` class.
- Mocked Google Search API to return the lyrics page under test.
Since at least one Backend requires album` and `duration` arguments
(`LRCLib`), the caller (`LyricsPlugin.fetch_item_lyrics`) must always
provide them.
Since they need to provided, we need to enforce this by defining them as
positional arguments.
Why is this important? I found that integrated `LRCLib` tests have been
passing, but they called `LRCLib.fetch` with values for `artist` and
`title` fields only, while the actual functionality *always* provides
values for `album` and `duration` fields too.
When I adjusted the test to provide values for the missing fields,
I found that it failed. This makes sense: Lib `album` and `duration`
filters are strict on LRCLib, so I was not surprised the lyrics could
not be found.
Thus I adjusted `LRCLib` backend implementation to only filter by each
of these fields when their values are truthy.
Modified `search_pairs` function in `lyrics.py` to:
* Firstly strip each of `artist`, `artist_sort` and `title` fields
* Only generate alternatives if both `artist` and `title` are not empty
* Ensure that `artist_sort` is not empty and not equal to artist (ignoring
case) before appending it to the artists
Extended tests to cover the changes.
- Consolidated multiple test cases into parameterized tests for better
readability and maintainability.
- Simplified assertions by comparing lists of actual and expected
artists/titles.
- Added `unexpected_empty_artist` marker to handle cases which
unexpectedly return an empty artist. This seems to be happen when
`artist_sort` field is empty.
- Replaced unittest.mock with pytest fixtures for better test isolation and readability.
- Simplified test cases by using parameterized tests.
- Added `requests-mock` dependency to `pyproject.toml` and `poetry.lock`.
- Removed redundant helper functions and classes.
In order to include the table name for fields in this query, use the
`field_query` method.
Since `AnyFieldQuery` is just an `OrQuery` under the hood, remove it and
construct `OrQuery` explicitly instead.
Unify query creation logic from
- queryparse.py:construct_query_part,
- Model.field_query,
- DefaultTemplateFunctions._tmpl_unique
to a single implementation under `LibModel.field_query` class method.
This method should be used for query resolution for model fields.
Ubuntu version in GitHub Actions has recently been upgraded to 24.04:
https://github.com/actions/runner-images/issues/10636)
This meant that pandoc was upgraded and it changed the way markdown is
formatted by default.
# Improve release notes formatting / changelog conversion from rst to md
During our last release, we discovered issues with changelog formatting.
This PR improves and fixes several aspects:
## Changes
- Rewrite the changelog conversion logic to be more robust and
maintainable
- Fix indentation issues with nested bullet points
- Improve handling of long section headers
- Order bullet points alphabetically within sections for better
readability
- Use Sphinx `objects.inv` to resolve references and include links to
the documentation in _Markdown_
- Add tests to prevent formatting regressions
- Add pandoc as a dependency for Ubuntu CI builds
- Ensure documentation is built before generating changelog
## Problem
A regression was introduced when adjusting the track matching logic to
use `lapjv` instead of `munkres`. The `lapjv` algorithm returns `-1` for
unmatched items, which wasn't being handled correctly in the matching
logic. This caused incorrect track assignments when importing new music.
## Solution
- Modified the mapping creation to filter out unmatched items (where
index is `-1`)
- Updated test case to properly catch this scenario
These tests depend on certain `track_length_grace` and
`track_length_max` configuration which was set by other tests in this
module.
I discovered this issue when I tried to run
`test_order_works_when_track_names_are_entirely_wrong` test only
- I found that my local configuration was read and the test failed.
I had previously tested the `munkres` -> `lapjv` replacement
extensively, so I was today surprised to find that nothing gets matched
correctly when I tried importing some new tracks.
On the other hand I now remember making a small adjustment in the logic
to make autotagging tests pass which is when I introduced a bug: I did
not realize that `lapjv` returns index '-1' for each unmatched item.
This issue did not get caught by tests because this 'unmatched' item
index '-1' anecdotally ended up pointing to the last (expected) item in
the test making it pass.
This commit adjusts the aforementioned test to catch this issue and
fixes the logic to correctly identify unmatched tracks.
