Fix critical startup and prompt import regressions

[cursor]

Bug and impact

• Headless server mode with configured GPU workers could emit comfyui:server_ready even when every worker failed or timed out, leaving Docker/server users with a falsely healthy backend and broken generation.
• PNG metadata import split preserved <segment:...> refinement prompts on every comma, corrupting saved prompt metadata and breaking segment round-trips.
• Model-family fallback matched ordinary names containing anima such as Animation/Animagine as Anima, selecting the wrong generation architecture.

Root cause

• server_main.rs did not gate the ready broadcast on any worker reaching Idle, unlike desktop/browser paths.
• stripQualityTags tokenized the entire prompt by comma after preserving supported inline tags.
• Anima detection used broad substring matching.

Fix

• Broadcast server_ready only when at least one configured worker is ready; otherwise emit the existing structured server_error payload.
• Treat parsed segment ranges as atomic during metadata quality-tag stripping.
• Use explicit Anima/Yume marker matching and preserve known animayume detection while avoiding Animation/Animagine false positives.

Validation

• npm run build passed.
• cargo fmt --check passed.
• cargo +stable check --manifest-path src-tauri/Cargo.toml passed after installing required Linux Tauri native packages.