[Research]: Benchmark which :free OpenRouter models reliably support tools

[shauryagangrade]

Problem

README warns some free models return "No endpoints found that support tool
use", and free-tier rate limits vary wildly — but there's no data on which
models actually work with GCode's tool schema and stream reliably.

Proposed approach

• Run a scripted benchmark: for each :free model on OpenRouter, send a
  fixed prompt that requires one tool call (e.g. read_file), record
  success/failure, tool-support, latency, and 429 rate-limit behavior.
• Publish results as a table in the README/models docs and propose a curated
  default list plus a warning badge in /models for models known to lack tools.

Where to look

• gcode/models.py (supports_tools derivation) — cross-check vs reality.
• README.md:78-85 (Models section).

Acceptance criteria

• A reproducible script + published results table.
• README recommends specific models; /models can flag untested/failing ones.

Difficulty

Medium — great non-Python-heavy contribution.