test: add unit tests for skillopt.optimizer.scheduler (edit-budget schedulers)

[nankingjing]

Summary

Add comprehensive unit tests for skillopt/optimizer/scheduler.py — the edit-budget schedulers that control how many skill edits are allowed per optimisation step (analogous to gradient clipping / learning-rate annealing in neural-network training).

This module had zero test coverage despite being a core part of the training loop. It is purely deterministic math with no LLM or network dependencies, making it ideal for precise unit tests.

Coverage

• ConstantScheduler — step/state_dict/resume/arbitrary-step behaviour
• LinearScheduler — decay sequence, monotonicity, bounds, resume consistency, degenerate cases (total_steps=0/1, max_lr==min_lr)
• CosineScheduler — cosine annealing shape, midpoint exact value, early/late flat regions, monotonicity, bounds
• AutonomousScheduler — always returns NO_LIMIT, state dict round-trip
• build_scheduler — factory creates correct types, unknown mode raises ValueError
• LRScheduler — abstract base cannot be instantiated directly

Style

Matches existing test conventions from tests/test_scoring.py and tests/test_types.py:

• from __future__ import annotations
• Class-based organisation with descriptive method names
• pytest with pytest.raises for error paths

Verification

All test expectations are derived from the source formulas by reading — no execution was performed (tests cannot be run in this environment). Every arithmetic assertion has a comment deriving the expected value from the formula in scheduler.py.

Co-Authored-By: Claude noreply@anthropic.com