Fix full inference E2E model contention

[ajroetker]

Summary

• use ggml-org/gemma-4-e2b-it-gguf as the single default for ordinary, tool-calling, and multimodal generation
• bootstrap explicitly configured large models and pin the Gemma GGUF variant
• keep ReleaseSafe assertions enabled while extending request and job budgets for CPU inference
• accept both current and legacy ClipClap manifest shapes, and replace the brittle synthetic-color semantic assertion with embedding-space invariants

Root cause

The failed full E2E job loaded separate generator defaults under ReleaseSafe. Slow requests timed out after 30 seconds, retained competing model sessions, then cascaded through MODEL_RESOURCE_BUSY into server termination. The job also exposed two independent ClipClap test assumptions: the generated manifest no longer always has a files array, and solid-color images do not provide a stable semantic ranking signal.

Failed job: https://github.com/antflydb/antfly/actions/runs/30694586941/job/91355247573

Validation

• 12 model-helper tests passed
• ClipClap golden-contract test passed
• live ClipClap embedding-space test passed
• all 132 inference E2E tests collect successfully
• workflow YAML parses successfully
• git diff --check passes