feat(eval): add model-quality benchmark fixtures and metrics

[zachshallbetter]

What does this change?

The repository benchmarks runtime speed in benchmarks/bench_inference.py (PR #221), measuring FPS, latency, and VRAM scaling. It doesn't measure whether an inference result is correct.

Latency isn't key quality. It's a separate measurement.

Without a quality harness, evaluating candidate changes relies on eyeballing frames. Eyeballing doesn't scale across subtle edge cases. We're already seeing the limits of visual inspection in open discussions:

• [Bug]: Inconsistent quality #184 shows frame drops when the alpha hint drifts.
• Feature Request: Support for non-green chroma key backgrounds (blue, magenta) #195 shows edge damage on chroma collisions.
• Possible improvements on the data augmentation pipeline #129 proposes backing augmentations without a way to measure regressions.

This PR adds a standalone, reproducible model-quality evaluation harness in corridorkey.eval. It measures alpha error, straight foreground recovery, edge boundary error, and physical recomposition. It doesn't touch the training pipeline.

Area	Change
corridorkey/eval/cases.py	MattingEvalSample dataclass contract and dataset loaders. Validates dimensions, channel counts, and spatial alignment.
corridorkey/eval/metrics.py	Four core quality metrics: alpha_mae, foreground_mae (where A > 0.01), boundary_alpha_mae (0.01 < A < 0.99 plus morphological band), and recomposition_mae (C_hat = alpha_hat * F_hat + (1 - alpha_hat) * B).
corridorkey/eval/fixtures.py	Generates 7 canonical diagnostic cases on demand: opaque_subject, soft_hair, motion_blur, green_foreground, blue_foreground, bad_hint_eroded, and bad_hint_dilated. Zero binary files tracked in git.
corridorkey/eval/runner.py	Engine adapter supporting CorridorKeyEngine, PyTorch modules, and callables/mocks. Aggregates global means and tag breakdowns.
corridorkey/eval/report.py	Formats structured JSON benchmark receipts and terminal summary tables.
corridorkey/eval/__main__.py	CLI entry point: python -m corridorkey.eval and console script corridorkey-eval.
pyproject.toml	Packages corridorkey and backend into wheel targets; adds corridorkey-eval script.
.gitignore	Ignores generated benchmark datasets, eval receipts (eval*.json), and cache artifacts.
tests/eval/	24 unit and integration tests covering metrics, loaders, reporting, and CLI.

How was it tested?

1. Unit and integration suite:

   uv run pytest tests/eval/ -v

   24 passed in 3.46s across metric math, boundary masking, sample contracts, and CLI execution.

2. CLI execution:

   uv run corridorkey-eval --checkpoint mock --dataset /tmp/ck_eval --generate-fixtures --output /tmp/ck_eval/receipt.json

   Generates fixtures on the fly, evaluates all 7 cases, prints the terminal summary table, and writes a valid JSON receipt.

3. Wheel build and packaging:

   uv build --wheel
   unzip -l dist/corridorkey-1.0.0-py3-none-any.whl

   Confirmed corridorkey, corridorkey/eval, backend, and CorridorKeyModule are properly packaged at the root.

4. Backwards compatibility:
   Existing inference pipelines, wizard prompts, and CLI options are completely untouched.

Checklist

• uv run pytest passes
• uv run ruff check passes
• uv run ruff format --check passes