[kPsarakis]

Fixes #105.

Stack: 1 of 3. This PR → #107 (tied scores in 1:1 selection) → #109 (metric input validation and docs).

Problem

Metric.name() returned only the class name, so get_metrics() stored F1Score(one_to_one=True) and F1Score(one_to_one=False) under the same key and one score overwrote the other. METRICS_ALL returned 6 of its 9 scores. Which variant survived depended on set iteration order: in one run "Precision" held the one_to_one=False value.

Change

• Metric.name() appends every dataclass field that differs from its default, e.g. "F1Score(one_to_one=False)". Metrics built with default arguments keep their old names ("F1Score", "PrecisionTop10Percent"), so METRICS_CORE output, the README and the benchmark scripts are unchanged.
• PrecisionTopNPercent keeps n in the name and appends only the other fields: "PrecisionTop70Percent(one_to_one=False)".
• get_metrics() computes equal metric instances once, and raises ValueError when two different metrics report the same name (e.g. custom name() overrides), instead of silently overwriting.

I didn't use str(self) as suggested in the issue, because it would rename every key, including the defaults ("F1Score" → "F1Score(one_to_one=True)").

Breaking change

Code that reads a non-default metric by its bare class name now gets a different key: get_metrics(gt, {F1Score(one_to_one=False)}) returns "F1Score(one_to_one=False)" instead of "F1Score". This is recorded with a migration note in the changelog's Unreleased section, so it's better suited to a minor release than a patch.

Tests

• Updated the 4 assertions that expected bare names for one_to_one=False variants.
• Added tests for metric names, the issue's reproduction, METRICS_ALL returning every score, duplicates computed once, and the name-clash error.
• Added tests that differently configured custom metrics (as in docs/metrics.md) get separate keys, and that instances differing only in a repr=False field raise.
• Full suite: 311 passed locally.

CI fixes (separate commits)

Both are unrelated to #105, but this PR's CI needed them, and they sit at the bottom of the stack so #107 and #109 get them too.

• macOS tests: every macOS job failed on TestJaccardEmbeddingIntegration with UserWarning: Skipping device Apple Paravirtual device that does not support Metal 2.0. torch emits it when transformers probes for MPS on GitHub's macOS runners, and filterwarnings = ["error"] turns it into a failure. It appeared since the 2026-10-14 master run, alongside torch 2.14.1 → 2.15.0 and a new macos-26-arm64 image, so master would fail the same way now. pyproject.toml ignores that one message.
• Docs builds: all Documentation runs shared one pages concurrency group with cancel-in-progress, so a push to one PR cancelled the docs builds of the other open PRs. PR runs now get a group per branch; master pushes and manual runs, which deploy, still share one.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.24%. Comparing base (f9d7234) to head (957414b).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #106      +/-   ##
==========================================
+ Coverage   95.17%   95.24%   +0.06%     
==========================================
  Files          53       53              
  Lines        2820     2839      +19     
  Branches      440      446       +6     
==========================================
+ Hits         2684     2704      +20     
  Misses         83       83              
+ Partials       53       52       -1     

Flag	Coverage Δ
unit	95.24% <100.00%> (+0.06%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
valentine/algorithms/matcher_results.py	98.12% <100.00%> (+0.03%)	⬆️
valentine/metrics/base_metric.py	96.77% <100.00%> (+3.02%)	⬆️
valentine/metrics/metrics.py	89.69% <100.00%> (+1.14%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.