fix(loop): clamp runaway expansion to model output limit

[zhanghanduo]

Summary

• Clamp the expanded runaway retry budget to the model's max_output_tokens as well as available context.
• Skip expansion when the active cap already reaches the model output limit, allowing reduced recovery to proceed.
• Resolve the limit through LLMFallbackChain as the smallest declared limit among its legs: the chain binds one max_tokens for whichever leg serves the request and exposes no capabilities of its own. Legs with no declared limit leave expansion bounded by context only, as before.
• Cover Claude Opus 5.5 at 128000 tokens, partial expansion to the model limit, and chains at and below their limit.

Verification

• Rebased on 0.16.0 main: full suite 2312 passed, 2 skipped
• Ruff, Pyright and scripts/check_version_bump.py pass; adds changes/74.fix.md

🤖 Generated with Claude Code