Routing advisor parity + reliability
Codex GPT-5.6 tier fix (real correctness bug). The GPT-5.6 family (Sol flagship / Terra workhorse / Luna cheap) was tiered upside-down: gpt-5.6-sol (the flagship) sat in the budget slot and sol was listed as a very-cheap model, so the router handed the most expensive model to the easiest tasks and the floor guarantee treated the flagship as cheap. Verified the lineup against OpenAI's GPT-5.6 preview + Codex CLI and corrected the ladder: budget=luna, mid=terra, capable/frontier=sol.
Config-driven OpenCode ladder. recommend() / baseline() now accept an optional resolved model map (engine stays pure); OpenCode's real ladder is read from opencode.json (model / small_model), falling back to the table when config is absent or malformed. Hermes is a Claude runtime, so its ladder is correct and documented as such.
Pricing. Added gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna pricing rows + normalizer aliases so those models resolve and price.
Copilot. Fixed an invalid bare gpt-5.6 frontier entry to the real, priced gpt-5.6-sol.
README badges. version / last-release / stars badges were showing shields.io's "Unable to select next GitHub token from pool" error. Moved version to a dynamic/json read of the plugin manifest, stars and last-commit to badgen (different backend).
Full test suite: 331 passing (+7 new).