test: split slow tests, cache CI deps, add pre-push gate

[bartzbeielstein]

Context

The pytest suite has grown to 1,727 tests across 151 files and takes ~113 s locally with -n auto; on GitHub it's worse because no dependency caching exists, so every run reinstalls all deps including the CPU-only torch wheel. There was also no way to skip heavy tests (no markers) and no enforcement of running tests before pushing.

This PR accelerates the suite locally and in CI without removing any test coverage, and adds a local pre-push gate.

What changed

Slow/fast split (coverage preserved)

• New tests/conftest.py central registry tags the 109 heaviest end-to-end tests @pytest.mark.slow (derived from pytest --durations). No tests deleted — the full suite still runs everything.
• pyproject.toml: registered the slow marker + addopts=["--durations=15"].

Run	Tests	Wall-clock (local)
Full suite	1,727	~113 s
Fast gate (-m "not slow")	1,618	~48 s

CI acceleration (ci.yml, release.yml)

• Dependency caching via setup-uv enable-cache (existing pinned action SHA kept) — torch & friends stop re-downloading every run.
• Installs switched uv pip install --system → uv sync (no more double-install with uv run).
• Fast required PR gate (-m "not slow") + new test-full job running the whole suite nightly (cron) and on workflow_dispatch. release.yml now runs the fast gate as a pre-publish check instead of the full suite.

Force tests before push

• New .pre-commit-config.yaml: pre-push hook runs uv run pytest -m "not slow" -n auto and blocks the push on failure; commit-stage ruff + black --check. Added pre-commit to dev deps.
• Install per-clone (documented in README): uv run pre-commit install --hook-type pre-push --hook-type pre-commit.

Notes for reviewers

• The slow set lives in one reviewable place (tests/conftest.py) and is matched by node id (file / class / test granularity); refresh it from --durations as the suite evolves. New slow tests can also just use @pytest.mark.slow.
• uv.lock includes a pre-existing lockfile refresh that was already in the working tree (Dependabot transitive bumps) plus the new pre-commit dependency tree.
• Real CI timing improvements will be visible on this PR's first run (warm cache benefits subsequent runs).

Verification

• Fast suite: 1,618 passed in ~48 s; slow set: 109 tests deselected.
• ruff check + black --check clean on new files; all workflow/precommit YAML parses.
• Pre-commit hooks install and validate; commit-stage hooks pass (verified on the commit in this PR).

🤖 Generated with Claude Code