feat(discovery): classify gateway models by capability

[dgokeeffe]

Summary

• classify Responses, Gemini and OSS models from advertised capabilities instead of brittle names
• filter explicitly NOT_READY endpoints from generated agent catalogues
• persist validated context/reasoning specifications alongside discovered model IDs
• retain gateway-only fallbacks when Unity Catalog has not registered a model yet
• consume validated OSS capabilities in OpenCode and disable its unsupported prompt_cache_key

Why

AI Gateway's model inventory changes continuously. Name-based assumptions can hide new models, expose non-chat models, advertise endpoints that cannot serve traffic, or generate requests the gateway rejects. Capability-driven discovery keeps ucode aligned with the live workspace contract.

Stack

PR C of five. Depends on PR B (pr/claude-discovery-hardening), which depends on A. It is opened as a draft until those land; the isolated review delta is c8e7114..763fa3e.

Planned dependency order: A → B → C → D and E.

• A: #446 — Pi routing foundation
• B: #447 — Claude discovery hardening
• C: #448 — this PR
• D: #449 — Pi OSS streaming
• E: #450 — OpenCode native Responses routing

Validation

• 792 focused discovery/CLI/Claude/OpenCode tests passed
• explicit NOT_READY, malformed metadata and gateway-only fallback cases are covered
• Ruff lint and changed-file formatting passed
• ty check src and git diff --check passed
• live discovery returned ready native Responses and validated OSS models
• the complete integration stack passed 2,358 tests with 37 skips; its three failures reproduce unchanged on main

Non-goals

This PR does not add Pi's MLflow repair proxy or OpenCode's native Responses provider; those remain separate dependent PRs.