Pseudonymization pipeline for shareable fixtures

[gavinbee]

Take a real scanned PDF + its REMS golden CSV, emit a committable name-swapped pair into `tests/fixtures/pseudonymized/`.

Pipeline sketch (per design doc §Test data acquisition Tier 3):

1. Identify bounding boxes for name fields on each page (Official Name, Mentor, COC, Competition Coordinator).
2. Generate pseudonyms (Faker), preserving referential integrity across appearances of the same real name within the meet.
3. Use a generative image model (inpainting / image-to-image, e.g. local Stable Diffusion + ControlNet, or a handwriting style-transfer model) to write each pseudonym onto the scan while preserving the original handwriting character — same style, ink color, contrast, paper texture, scan noise. Clean white-rectangle overlays with handwriting fonts are not acceptable because they erase the artifacts the parser must handle.
4. Rewrite the golden CSV with the pseudonym map.
5. Commit (image, golden_csv) pair into `tests/fixtures/pseudonymized/`.

Directory reserved in v1; this issue implements the pipeline later.