feat: add Fireworks Serverless Training backend

[xiaoyifan]

Summary

• add TRAINING_BACKEND=fireworks as a Tinker-compatible LoRA training backend
• keep the existing RSIBench algorithm unchanged: offline LoRA SFT followed by sampler-checkpoint evaluation
• separate the Hugging Face target/tokenizer name from the Fireworks accounts/.../models/... training resource
• persist only the Fireworks metadata needed for checkpoint evaluation: training session ID, sampler checkpoint, tokenizer, provider model, and endpoint
• reconnect the E2B OpenAI-compatible proxy to the Fireworks session for Harbor attempt and official evaluation sampling
• keep existing Tinker artifact names and budget-state schema unchanged while using provider-neutral artifacts for Fireworks runs
• validate credentials, account-specific rates, and the Fireworks model-resource format before paid training or evaluation work can start
• normalize Fireworks checkpoint names to the service DNS-label contract
• pin compatible Fireworks/Tinker versions in the reusable E2B proxy template

Scope

This PR keeps Fireworks behavior parallel to the existing Tinker integration:
LoRA SFT, sampler checkpoint creation, checkpoint evaluation, and provider
budget accounting. It does not add optimizer-state persistence, an online RL
algorithm, or a new training orchestration layer.

Flow

flowchart LR
  A[Data agent] --> B[train_messages.jsonl]
  B --> C{Training backend}
  C -->|Tinker| D[Tinker LoRA SFT]
  C -->|Fireworks| E[Serverless LoRA SFT]
  D --> F[Sampler checkpoint]
  E --> F
  E --> G[Session ID and tokenizer]
  F --> H[E2B checkpoint proxy]
  G --> H
  H --> I[Harbor evaluation]
  I --> J[Score and provider cost state]

Loading

Review cleanup

• removed the Fireworks-only save_state call and resumable-checkpoint metadata because no RSIBench runner consumes them and Tinker has no parallel behavior
• restored the existing Tinker metadata, sampling-cost artifact, preflight artifact, and budget-state field names for new Tinker runs
• added fail-fast validation that rejects a Hugging Face ID in Fireworks TRAINING_MODEL before constructing the remote service client
• moved Fireworks sampling-rate validation ahead of E2B sandbox creation and provider sampling

Validation

• python3 -m unittest discover -s tests -v — 75 tests passed
• Python compile checks for all changed runtime modules
• bash -n for runner, pipeline, official-eval, environment, setup, and Docker wrapper scripts
• credentialed Fireworks multi-step SFT:
  • accounts/fireworks/models/qwen3p5-9b, LoRA rank 8
  • 5 optimizer steps, 2 gradient accumulations per step, batch size 2
  • 20 real forward/backward calls and 1,180 processed training tokens
  • reported loss sum decreased every step from 162.01 to 149.37
  • sampler checkpoint saved with a 24-hour TTL
• credentialed checkpoint sampling through the Fireworks E2B proxy and Harbor:
  • one SWE-bench Verified task with step_limit=5
  • 5 model/tool turns completed against the saved sampler checkpoint
  • 14,216 prompt tokens and 607 completion tokens
  • zero Harbor infrastructure errors, zero retries, and automatic E2B sandbox cleanup

The credentialed runs exposed and verified fixes for provider model naming,
explicit tokenizer loading, checkpoint DNS-label length, dependency
compatibility, and pricing fail-fast behavior.