feat(evaluation): add generic deterministic metrics (WBP 2.3)

[lzongren]

What & why

Implements WBP task 2.3 (generic metrics) for the Transform Gym evaluation framework. Before this, the only built-in scorers were assertion_pass_rate and llm_judge (plus credits/usage). This adds three generic, deterministic, zero-Bedrock metrics that resolve by name through MetricRegistry, alongside the existing two.

Metric	What it scores	Per-test config (metadata)
tool_usage	expected_tools were called and forbidden_tools were avoided	expected_tools, forbidden_tools
error_handling	transcript is free of a surfaced framework error (default marker ERROR:) — error absence, not recovery quality	error_markers (optional override)
completeness	run reached its designed end (finished before exhausting max_turns) and final output is non-empty	completion_markers (optional override)

All three score 0–10 and pass at ≥ 7.0, matching assertion_pass_rate so they combine cleanly in one run.

Design decisions

• Config via TestCase.metadata, not a schema change. The metrics read free-form metadata keys, so no TestCase/JSON schema churn.
• Abstain when there's no signal. tool_usage and completeness return 10.0/pass when unconfigured (no expected_tools/forbidden_tools; no usable turn_count/max_turns or markers), mirroring assertion_pass_rate's abstain convention — so they're safe to enable globally and only "activate" when there's something to grade. They are not added to the default metrics list (["assertion_pass_rate"]), so this PR cannot regress existing runs.
• completeness keys off the turn budget, not __DONE__. The ACP engine consumes its __DONE__ sentinel (execution/runner.py breaks before appending it to the transcript), so a marker-based check would score the inverse of completion. Instead, completeness compares turn_count < max_turns — a naturally-completed run stops below the ceiling; a truncated one hits it. This required surfacing turn_count onto ExecutionResult (populated from EvalResult.turn_count in ACPAgent, emitted in results.json). completion_markers remains an optional override that takes precedence when set.
• error_handling default is narrow on purpose. Only ERROR: is matched by default — the one prefix the framework actually emits (runner.py/engine.py; the curated sample asserts transcript_not_contains: "ERROR:" by the same convention). A broader set (Exception:/FATAL/panic:/Traceback) false-positives on ordinary agent prose; teams that want stricter markers supply error_markers.

Notes for reviewers

• Shared metadata coercion lives in metrics/_metadata.py (coerce_str_list drops non-string entries, so a stray null can't become a phantom tool/marker).
• tool_usage reads the tool name key only, matching how assertion_pass_rate grades tool_called.
• ExecutionResult gained turn_count: int | None = None (appended last, defaulted — all construction sites unaffected; JSON-serializable).
• README documents the metric table, a metadata example, and a custom-metric snippet.

This change was self-reviewed adversarially before raising: an earlier revision had completeness depending on the (stripped) __DONE__ sentinel and error_handling using the broad marker set — both were caught and fixed, with the boundary math (turn_count < max_turns, no off-by-one) re-verified against runner.py.

Testing

• Full eval suite passes (309 tests; +14 vs. base), ruff clean (E/F/I).
• New/updated tests: test_tool_usage.py, test_error_handling.py, test_completeness.py (turn-budget semantics + marker override), test_metadata.py (coercion contract), plus registry built-in/resolve tests.