MaxCode: Add MaxText conversion path with dedicated agent and demo scripts

[gvanica]

Summary

Adds a second conversion target — instead of only converting PyTorch to standalone JAX/Flax, MaxCode can now also convert to MaxText artifacts: a YAML config overlay, optional custom layer files, and a best-effort checkpoint converter. This enables direct integration with the MaxText training framework.

MaxTextConversionAgent (agents/migration/maxtext_conversion_agent.py)

A multi-stage LLM-powered agent that produces MaxText-compatible output:

1. Classification — Analyzes the PyTorch source to determine which MaxText base model is closest (e.g., Llama, Qwen, DeepSeek) and what custom components are needed
2. YAML config overlay — Emits a MaxText YAML configuration that maps the source architecture to MaxText parameters (hidden size, num heads, num layers, etc.)
3. Custom layers (conditional) — If the source has components not covered by MaxText built-ins (e.g., GatedDeltaNet), emits a standalone Flax module
4. Checkpoint converter (best-effort) — Generates a weight-mapping script for converting PyTorch checkpoints to MaxText format

Demo scripts

Script	Purpose
step4_convert_maxtext.py	Runs MaxText conversion pipeline, writes YAML + layers to timestamped output directory
step5_verify_maxtext.py	Runs verification scorecard against MaxText output, accounting for components delegated to MaxText built-ins
step4_convert_api.py	Alternative conversion path using the REST API endpoint

Design decisions

• Multi-stage prompting — Each LLM call is focused on one artifact type, keeping prompts short and outputs well-structured
• Delegation awareness — The verification scorecard (in step5_verify_maxtext.py) understands that some PyTorch components are handled by MaxText built-ins and excludes them from completeness scoring
• Timestamped output — MaxText artifacts are written to output/maxtext_YYYYMMDD_HHMMSS/ to preserve multiple conversion attempts

Test plan

• Run step4_convert_maxtext.py on a cloned Qwen3 repo — verify YAML config and custom layers are produced
• Run step5_verify_maxtext.py — verify scorecard shows delegated components and adjusted completeness
• Verify step4_convert_api.py works with a running API server

---

Split from #17 — PR 7 of 8

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.