feat: unify training rendering and harden FrozenLake VL

[benjibc]

Summary

• harden the FrozenLake multimodal rollout path for Kimi K2.5 VL and Qwen3 VL, including prompt-faithful token/mask accounting, verifier-side validation, and validation screenshots for ep logs
• unify cookbook training onto shared token-level rendering/masking so SFT, DPO, ORPO, and FrozenLake RL all use the same target_tokens + weights representation
• keep FrozenLake training aligned with eval-protocol by deriving the training datum from the same per-token mask that the UI renders
• add a qwen3-4b smoke-test CI workflow for SFT, DPO, and GRPO, plus local unit/import coverage for the shared rendering path
• require eval-protocol>=0.3.23 so the cookbook picks up the released ep logs transcript/token-debug UI changes and the relaxed fireworks-ai dependency range from eval-protocol

Eval Protocol Releases

• UI PR: feat: align ep logs chat view with tokenized rollout prompts eval-protocol/python-sdk#431
• Hotfix PR: Relax fireworks-ai dependency pin eval-protocol/python-sdk#432
• Release: https://github.com/eval-protocol/python-sdk/releases/tag/v0.3.23
• PyPI: eval-protocol==0.3.23

Validation

• ./training/.venv/bin/pytest -q training/tests/unit training/tests/test_smoke_imports.py training/examples/frozen_lake/test_masking.py
• env -u FIREWORKS_API_KEY -u FIREWORKS_ACCOUNT_ID -u FIREWORKS_BASE_URL -u FIREWORKS_INFERENCE_URL -u FIREWORKS_HOTLOAD_API_URL -u FIREWORKS_GATEWAY_SECRET ./training/.venv/bin/pytest -q training/tests/smoke_test/test_sft_smoke.py training/tests/smoke_test/test_dpo_smoke.py training/tests/smoke_test/test_grpo_smoke.py
• real visual verifier runs against:
  • accounts/fireworks/models/kimi-k2p5
  • accounts/fireworks/models/qwen3-vl-30b-a3b-instruct
• Chromium verification against the live ep logs UI

Smoke CI

The new workflow expects these GitHub secrets/vars for remote smoke runs:

• secrets: FIREWORKS_API_KEY, FIREWORKS_ACCOUNT_ID
• optional vars: FIREWORKS_SMOKE_TRAINING_SHAPE, FIREWORKS_SMOKE_BASE_MODEL, FIREWORKS_SMOKE_TOKENIZER_MODEL, FIREWORKS_BASE_URL, FIREWORKS_INFERENCE_URL, FIREWORKS_HOTLOAD_API_URL, FIREWORKS_CUSTOM_IMAGE_TAG

Defaults are wired to:

• model: accounts/fireworks/models/qwen3-4b
• tokenizer: Qwen/Qwen3-4B
• shape: ts-qwen3-4b-smoke-v1

Screenshots

Kimi K2.5 VL

Qwen3 VL