experiment: validate EOS-aware diverse prompts across four model backends #227

Problem

The four-model matrix mostly forces a fixed number of tokens from repeated prompts. Matching those tokens does not establish normal EOS-aware behavior, prompt diversity or cross-backend answer quality.

Evidence

#201 and #225 demonstrate transport and capacity, including Gemma's corrected native attention. Fixed outputs continue after EOS and previously concealed an invalid causal reference.

Scope

Add shared, checkpoint-tokenized conversational/long-context fixtures and explicit EOS-aware generation for Qwen3-0.6B, Qwen3-4B, KONI-8B and Gemma3-1B. Compare GPU and ATOM native outputs first, then use the same input IDs for RNGD PD. Record per-case hashes, stop reasons, numerical/output mismatches and factual checks separately from performance.

Done when

All four model fixtures run on available GPU/ATOM and RNGD paths with reported EOS/length outcomes; divergence is retained and diagnosed, rather than hidden by exact-match-only summaries. Keep the broader production PD/overlap goal active.

Owner

Codex, maintainer review required. Related #201, #225.