PERF: Cache TitaNet/secondary ONNX weights through durable OPFS

[MikeSchirtzinger]

Problem

The app has a durable OPFS model cache for Nemotron, but secondary ONNX models can still rely on browser HTTP cache or engine-specific caches. That makes repeat-load behavior less predictable under browser eviction.

Evidence

• apps/web/js/model-cache.js is generic enough to store model bytes by logical key.
• nemotron-engine.js:343-362 uses OPFS read/write-through for Nemotron model files.
• TitaNet and other secondary model paths should be checked and wired through the same durable cache where they fetch raw ONNX bytes.

Acceptance criteria

• Audit TitaNet/diarization and other secondary ONNX fetch paths.
• Route applicable raw model-weight fetches through the OPFS model cache.
• Preserve privacy boundaries and avoid caching meeting data.
• Add cache-hit/miss logging consistent with Nemotron.
• Verify repeat load behavior after browser HTTP cache eviction where feasible.