[piotrlaczkowski]

Add optional externally served Laya/OpenAI-compatible predictions for bounded knowledge reranking and agent guidance, while keeping acceptance gates authoritative. Harden real-model delivery against mismatched role finalization, lost team contracts, blocked recovery reads and empty Node test suites. The release smoke now rejects failed or unfinished delivery instead of accepting team structure alone.

Improve Studio's live progress, connection feedback, mobile task navigation, motion controls and camera behavior, preserving the newer command-center and floor animation work from main. Update feature, configuration, operational and release documentation, including a dated qualification record and v0.28.0 migration notes.

Validation:

• Local make check passed: 72.2% coverage, zero Go lint findings; merged Studio passes 365 tests and its build/typechecks.
• Live oMLX/Qwen3-Coder delivery passed: 3/3 tasks, zero failed/unexecuted; native Laya and OpenAI-compatible decision protocol smokes passed.
• Independently ran the generated application's Go race tests and current acceptance guard, then exercised HTTP/DOM rendering, refresh, 24 concurrent increments and HTTP/network failure states in Chromium.
• Studio desktop/mobile/3D browser checks and packaged-binary authentication/navigation checks passed.
• Six platform/architecture cross-builds, strict docs, version/reference checks and dependency vulnerability audits passed.

Laya hosting, weights and training remain external. Predictions are opt-in; protocol checks do not establish a measured speedup or model accuracy. Final merged-tree CI and release artifact checks remain required before publication.