fix(agents): reuse deterministic interactive Prefill KV

[FluffyAIcode]

Summary

• remove random run nonce from interactive Generator and Critic model prompts
• centralize deterministic prompt construction for exact KV identity across repeated tasks
• compact system instructions while retaining open-problem honesty and evidence-window constraints
• reduce default Critic evidence budget from 128 to 64 content tokens
• retain random IDs only for benchmark telemetry, outside model tokenization

Measured prompt reduction

For 证明黎曼猜想 with a long Generator response:

• Generator Prefill: 93 → 40 tokens (57% reduction)
• Critic Prefill: 311 → 176 tokens (43% reduction)
• repeated identical tasks become eligible for exact allens/Primary KV hits because tokenized prompts no longer contain random entropy

The first unseen prompt remains bounded by the 16GB allens Gemma Prefill rate; this change removes unnecessary repeated cold work without violating Primary decode-only mode.

Tests

• deterministic prompt identity and retained safety instructions
• Agent GAN targeted suite — 13 passed
• full platform-neutral suite — 859 passed, 1 skipped
• Python compile and diff checks pass

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.