Replace video corpus with episodic fixtures

[brianmeyer]

Summary

• replace the tiny UAT video clips with compact episodic-memory fixtures and richer transcript sidecars
• add sidecar search text plus related image/document metadata so video benchmarks can exercise parent-memory, transcript, and related-artifact retrieval
• update text-to-video benchmark prompts, corpus docs, release docs, and video-quality UAT to match current embed/hybrid modes and deterministic CI behavior

Validation

• python3 -m pytest -q tests/test_video_corpus.py tests/test_video_sidecars.py tests/test_cross_modal_benchmark_defs.py
• python3 benchmarks/cross_modal_ablation.py --dry-run
• bash tests/uat/test_video_quality.sh
• python3 -m pytest -q
• .venv/bin/python -m pip wheel . -w /tmp/recallforge-wheel

Notes

• python3 -m build and .venv/bin/python -m twine check were not available locally because build/twine are not installed; pip wheel succeeded with an isolated build.
• Research references: Ego4D episodic-memory benchmark and Pillow ImageDraw docs for caption rendering.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d1393d90fd

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Use distinct image files for each captioned cooking scene

cooking_tutorial.mp4 reuses food_pasta_dish.png for scenes 1 and 3, but _burn_text_onto_image writes captioned frames to a temp filename keyed only by source.stem; when the same source appears twice with different scene_captions, the later render overwrites the earlier one, so scene 1 ends up showing scene 3’s caption. This makes the generated fixture inconsistent with its intended timeline/transcript and can skew video-retrieval benchmarks that rely on scene text cues.

Useful? React with 👍 / 👎.