feat!: add PE-A evidence retrieval and agent evaluation

[tulayha]

Related issue

None.

Summary

• Replace the product sound-retrieval path with PE-A-Frame Small frame embeddings, bounded overlapping inference sections, and distinct 10-second evidence windows. Existing sound indexes must be rebuilt because the embedding schema and distance metric changed.
• Add shared local-answer routing through the CLI, application layer, managed Ollama runtime, and Desktop, including separate one-request router and planner policies.
• Make the Promptfoo agent ablation reproducible across VidXP, direct-local, and clean-user conditions; preserve auditable reports, call/token/latency accounting, machine provenance, and lightweight result artifacts.
• Add modality-specific gates and research-adoption documentation covering the measured model choices, product defaults, limitations, and paper-facing claims.

Diff-size note: 183,770 of the 201,021 added lines (91.4%) are sanitized, retained Promptfoo JSON runs. The remaining additions are 12,064 implementation/harness lines, 2,656 documentation lines, 2,120 test lines, and 411 configuration/other lines; the JSON occupies 8.2 MiB in a checkout and approximately 1.11 MiB in Git-compressed storage.

Validation

• ./benchmarks/codex-mcp/run pilot — 81/81 isolated agent cases completed and were scorable. VidXP Success@3 was 15/27 versus 18/27 for direct-local; VidXP used 20.6% fewer tokens and was 18.3% faster. The quality gate remains explicitly failed.
• ./benchmarks/codex-mcp/run slm — 54/54 local-SLM cases completed and were scorable. Router/planner Success@3 was 15/27 and 19/27, with one local model request per case and no external-provider cost.
• ./benchmarks/codex-mcp/run slm-smoke — router and planner each returned three bounded candidates and passed Success@1/Success@3 on the development task.
• PE-A product-path smoke on mac-m2-01 — indexed 1,896 unique frames from a 75.81-second video; the two checked sound queries ranked their target 10-second windows first.
• AEGBench frozen 50-recording/149-query component gate — PE-A exceeded FineLAP on frame AUROC, frame AP, top-point accuracy, mean IoU, and R-IoU@0.5.
• Fresh Python 3.14 base-wheel environment with NumPy absent: python utils/verify_runtime.py minimal --vidxp-executable <venv>/bin/vidxp completed successfully; both vidxp --help and vidxp benchmark --help loaded without optional scientific dependencies.
• uv run --no-sync pytest -q tests/test_packaging.py — 20 passed.

Detailed configurations, machine profile, limits, metrics, and retained artifacts are linked from docs/benchmarking/metric_database.md and docs/benchmarking/results.md.

BREAKING CHANGE: Existing environmental-sound indexes use the prior representation and must be rebuilt for PE-A-Frame search.