test: freeze renderer behavior before refactoring

[hallerite]

Summary

• pin every Hugging Face tokenizer and processor used by network-backed tests to a reviewed immutable revision
• split CI into offline unit, pinned text-model parity, and pinned multimodal parity lanes
• add consistency checks for renderer registrations, configs, model routing, exports, and golden coverage
• add deterministic, reviewable behavior snapshots for every registered renderer, including rendered text, token digests, attribution masks, parser output, and bridge extensions
• correct the documented bridge return type and stale multimodal documentation

Why

The renderer internals are about to undergo a substantial emission/attribution refactor. The existing suite is broad, but upstream model assets were mutable and there was no compact checked-in contract showing the complete observable output of each renderer.

This PR makes the current behavior reproducible and reviewable before implementation moves. It is intended to change test architecture and documentation, not renderer output.

Developer impact

Offline tests continue to run on Python 3.10–3.13. Network-backed text and multimodal parity tests now run in dedicated Python 3.13 jobs using immutable model revisions. Golden output changes must be regenerated explicitly and reviewed in the diff.

Validation

• uv run ruff check .
• uv run ruff format --check .
• offline suite: 93 passed
• pinned text parity: 2,481 passed, 110 skipped, 1 expected failure
• pinned multimodal parity: 73 passed, 17 skipped
• registry and golden contract tests: 26 passed
• ty check renderers exits successfully with the existing advisory diagnostics

Note

Freeze renderer behavior with golden snapshot tests before refactoring

• Adds tests/golden_renderer_outputs.json with checked-in snapshots of renderer output, generated by tests/generate_renderer_goldens.py and validated by tests/test_golden_renderer_outputs.py.
• Introduces tests/model_assets.py with MODEL_REVISIONS (pinned full SHAs) and load_test_tokenizer, replacing direct load_tokenizer calls across all test files to ensure deterministic tokenizer loads.
• Updates load_tokenizer in renderers/base.py to support a revision kwarg and enforce a stricter trust policy: trust_remote_code is only enabled for models in TRUSTED_REVISIONS, and callers cannot override the revision for trusted models.
• Splits the CI Renderers job into three jobs: offline unit tests (Python 3.10–3.13), text model parity tests (network), and multimodal parity tests (network), with HF asset caching keyed by model_assets.py.
• Adds pytest markers (network, model_parity, multimodal) to categorize tests and exclude network tests from offline runs.

^{Macroscope summarized a48daf6.}