feat(deploy): record generating model(s) per cell in run manifest

[JoshuaBearup]

What

Records which model(s) generated each cell in a run, alongside the existing record of which model attacked it.

• Adds a generator_models column to runs/<id>/manifest.csv (e.g. claude-haiku-4-5-20251001:3|claude-opus-4-8:1 — model:call-count, pipe-separated).
• Attaches a generation block (the existing getUsageReport() output — tokens + cost by model) to the per-cell local manifest.

Why

A run already records the model under test (the attacker) but not which model produced each target. Since generation mixes models (Haiku by default, Opus for the higher-quality steps via quality: true), recording the generator mix per cell helps reproducibility and cost/quality analysis — and makes it explicit when a quality step silently ran on a different model than expected.

Design notes

• Local manifest only — not baked into the container. The generating model is metadata the model under test must not see, and RCE/LFI classes can read the baked manifest. This mirrors how solvabilityProof is attached post-deploy (generator/deploy.mjs), so it never reaches the container.
• Additive. The new column is appended at the end of manifest.csv, so existing consumers of that file are unaffected.

Testing

Verified node -c on both changed files, header/row column-count consistency (14 to 14), and the generator_models formatting against a sample byModel. I did not run a full end-to-end deploy on this branch (no ANTHROPIC_API_KEY in my environment), but the change only persists getUsageReport() output that deploy.mjs already computes and logs via logUsage() — no new generation behaviour.