fix(cli): preserve existing agent settings with --model

[devanshrj]

Summary

• treat an existing agent's persisted generation configuration as the baseline when startup combines --agent and --model
• apply the selected model identity, provider-specific settings, and reasoning tier while preserving context window, max tokens, and temperature
• clamp preserved caps to backend-reported model limits, with preset limits and backend validation as the fallback
• rebuild destination provider settings instead of carrying incompatible provider-specific fields across providers
• stop silently refreshing existing agents from model presets when no model override was requested

Fresh --new-agent --model creation and established interactive /model behavior continue to apply complete preset defaults. Same-handle reasoning tiers remain model-bearing updates, and --agent --model composes with both --conv default and --new conversation selection.

Verification

• bun run check — all 12 checks passed
• focused model/startup matrix — 113 passed, 0 failed
• affected API/local update suites after final cleanup — 19 passed, 0 failed
• isolated local non-inference CLI smoke — Sonnet to GPT-5.6 changed model/provider/reasoning while preserving 65,536 context, 4,096 max tokens, temperature 0.2, and the custom system prompt
• Letta Cloud non-inference CLI smoke — verified the same persisted before/after state through the real API; disposable agent deleted and deletion verified

Compatibility

• existing-agent CLI model overrides now preserve effective persisted caps/sampling values rather than replacing them with preset defaults
• values above the destination catalog limit are clamped with a warning
• missing catalog limits fall back to preset limits and backend validation
• fresh-agent creation and interactive /model semantics are unchanged