Harden runtime model selection and release checks

[QUSETIONS]

Summary

• Preserve the existing memory manager across resumed agent turns in the main, TUI, and headless frontends.
• Harden OpenAI-compatible model discovery and fallback selection so non-tool models are excluded and readiness checks stay local-only.
• Add release integration coverage for protocol fallback, refresh readiness evidence, and extend regression tests.

Root cause

Resumed frontend turns could create a fresh memory manager, and custom OpenAI-compatible provider probing could make local validation unexpectedly depend on network availability. Fallback selection also needed to distinguish agent-capable models from audio, image, and other non-tool models.

Validation

• pytest -q: 1310 passed, 2 skipped
• External provider smoke: one headless tool turn returned OK
• git diff --cached --check: passed
• No real provider secret is included in the commit

The report-only readiness gate still records the expected no-provider blocked state for hermetic threshold commands; the local test and integration gates are green.