Harden test logistics

[benjipeng]

What changed

• move every root-level test into the unit, checkpoint, runtime, or integration tier
• split mixed fast and local-artifact coverage into dependency-appropriate files
• make default pytest collection fast and deterministic while validating all opt-in tiers in CI
• add strict pytest configuration, line and branch coverage, and a 72.8% baseline floor
• document test placement, stubs, golden fixtures, required gates, and coverage policy

Why

Root-level tests were outside the normal unit CI command, default collection had stale import failures, and the checkpoint CI step could report success after skipping all tests because model artifacts are not present. This made green CI an incomplete signal.

Impact

The default suite now covers all artifact-free tests. Expensive tiers remain opt-in and are collection-checked in CI. No files under src/ changed, so runtime and public API behavior are unchanged.

Validation

• pytest tests/unit/ -q: 1071 passed
• CI-equivalent custom-Metal-disabled coverage run: 1065 passed, 6 explicitly skipped, 72.81% line and branch coverage
• all opt-in tiers: 152 tests collected
• focused Ruff check and format check passed
• git diff --check, workflow YAML parsing, and pyproject TOML parsing passed
• root-level test files: 0
• production source changes: 0