feat(metrics): identical-call rate and one-to-one call pairing (1.9.0)

[yotambraun]

Why

Found while evaluating a real research agent (bytedance/deer-flow) with Agent Eval Flow + Toolscore:

1. redundant_rate counts calls beyond the expected per-tool count, so an agent stuck repeating one call and an agent running many different searches look the same.
2. Argument and side-effect checks matched each expected call to the first same-tool call at the same or a later position. One missing, extra or reordered call shifted every later comparison, and correct calls scored 0.

What changes

• Added identical_count / identical_rate in efficiency_metrics: exact repeats (same tool, same arguments, key order ignored). Additive; composite score unchanged.
• Fixed argument F1 and side-effect validation now pair expected and actual calls one-to-one per tool by best argument match (exact assignment up to 12 calls per tool, greedy above). No tools expected and none called now gives argument_f1 1.0.

Case	argument_f1 before	after
One earlier call missing, the other two correct	0.00	0.67
Two same-tool calls in swapped order	0.00	1.00
Wrong attempt, then the correct call	0.00	1.00 (still penalised by redundant_rate, sequence_accuracy)
No tools expected, none called	0.00 (score 0.70)	1.00 (score 1.00)

• Fixed the Claude LLM judge with current anthropic SDKs (1.8 removed temperature from messages.create, so it raised TypeError; this also made CI's mypy step fail). temperature is passed only when the SDK accepts it.
• Docs: README "What's new in 1.9", CHANGELOG [1.9.0], user-guide sections on call pairing and loop detection; docs/changelog.rst now renders CHANGELOG.md (it was a stale 0.1.0 copy).

Release impact

feat: commit → semantic-release cuts 1.9.0 on merge. Scores can rise for existing baselines and snapshots; the CHANGELOG and README say to re-approve after upgrading.

Verification

• 14 new tests, each seen failing before its change; full suite 884 passed
• ruff check, ruff format --check, mypy toolscore clean with all extras installed (anthropic 1.8.0)
• Sphinx build renders the new sections and the 1.9.0 changelog

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 73.33333% with 16 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
toolscore/metrics/arguments.py	74.46%	11 Missing and 1 partial ⚠️
toolscore/metrics/efficiency.py	50.00%	2 Missing ⚠️
toolscore/metrics/llm_judge.py	80.00%	1 Missing ⚠️
toolscore/metrics/side_effects.py	75.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!