Lazy vs eager model loading strategy at app startup

[brendanddev]

Problem

The model currently loads on the first user message instead of at app startup.

This introduces a noticeable delay on first interaction, based on recent local runs:

• ~12.9s from loading model to generating
• ~16.5s from loading model to generating
• ~12.2s from loading model to generating

This impacts first-use UX and makes the app feel unresponsive at the start of a session.

Current behavior

• app starts quickly
• first user request triggers model load
• first response has high latency due to model load + subsequent prefill/generation

Desired behavior

Evaluate whether model initialization should be:

• eagerly loaded at app start
• optionally preloaded
• lazily loaded with clearer user feedback

Constraints

• must not interfere with runtime ownership or tool flow
• must not complicate CLI responsiveness unnecessarily
• must remain compatible with local models (llama.cpp)

Why this is deferred

This is a performance / UX improvement, not a correctness bug.
It is intentionally out of scope for Phase 9 investigation quality work and should be addressed separately.

Evidence

Recent logs show:

• loading model at +876ms → generating at +13784ms
• loading model at +24576ms → generating at +41069ms
• loading model at +1968ms → generating at +14161ms