[0xShug0]

Arena can run the same LFM2 package twice while marking two different precision entries complete. Its resident-model check reused Studio's matcher, which applies restored Studio session settings to the expected package options. A restored Q8 selection can therefore make another precision incorrectly match the resident Q8 model.

Arena now compares the resident model against the exact session options sent for each queue item, both before reusing it and after loading. Studio matching is unchanged. Includes the rebuilt WebUI bundle; no server, framework, or model changes.

Validation:

• Reproduced through browser settings import and a two-package Arena queue: one model load before, two distinct package loads after; both generations complete.
• Exercised real LFM2 Q8/Q4 weights against a Debug CUDA server. A browser inventory fixture mapped the second precision choice to Q4 because F16 weights were unavailable; this validates package switching, not F16 inference.
• Verified source UI and rebuilt bundle. Frontend type checking reports zero errors/warnings; frontend build passes.