test(gate): add unit tests for the evaluation gate decision core

[nankingjing]

What

Adds tests/test_gate.py — 22 unit tests for skillopt/evaluation/gate.py, the optimizer's validation-gate decision core (select_gate_score, evaluate_gate, GateResult).

Why

skillopt/evaluation/gate.py is the accept/reject decision at the heart of the ReflACT training loop — the analog of validation-based model selection / early stopping. It decides whether a candidate skill becomes the new current skill and whether it is promoted to best-so-far. Despite being pure, side-effect-free logic (the trainer owns all I/O), it currently has no dedicated test coverage — it is only reached indirectly through engine/trainer.py, so a regression in the gate's branching or metric math would ship silently.

This is in the spirit of #94 (bring evaluate_gate under direct test), applied to the main optimizer's gate.

Coverage

select_gate_score:

• hard / soft / default-metric projection
• mixed weighting (default weight, custom weight, w=0→hard, w=1→soft)
• out-of-range mixed_weight clamping to [0, 1] (both directions)
• return type is float; unknown metric raises ValueError

evaluate_gate — all three actions and their exact resulting state:

• accept_new_best (beats current and best; best_step advances to global_step)
• accept (beats current, not best; best-so-far incl. best_step preserved) — the branch only reachable when current_score < best_score
• reject (does not beat current; all state preserved)
• strict-> semantics: a tie with current_score is rejected (no lateral moves); a tie with best_score above current is an accept, not a new best
• metric routing via cand_soft (soft/mixed), and that the default hard metric ignores cand_soft

GateResult: field wiring and frozen-dataclass immutability.

Plus a behavioral invariant test documenting that when s_cur == s_best at the start (the trainer's initialization), the two stay locked together and the accept branch is never taken — capturing the design property discussed in #113.

Verification

Tests are pure-logic (no LLM, rollout, or network). Ran the full file against the unmodified gate.py: 22 passed. Style matches the existing suite (tests/test_scoring.py, tests/test_types.py): class-grouped, pytest.approx for float math, docstrings per group.

No production code changed — tests only.