v0.5.14 — Neural Engine auto-recovery

Changed

• The Neural Engine now recovers itself — no button required. The crash-fence that protects against a native Core ML compile crash was a permanent latch; it's now a bounded crash budget. Every launch auto-retries the compile, a clean compile resets the budget, and only after 2 failed compiles does it give up and settle on the CPU engine. A transient failure (memory pressure on a 711 MB fp16 compile) self-heals with zero taps; a genuinely uncompilable model can't thrash — at most two bad launches, then it stops.
• Catchable compile failures charge the budget too, so a deterministic non-crashing failure can't retry every launch forever.
• The legacy permanent-poison flag is ignored, so installing this build grants a fresh budget — a device stuck on the CPU from a past crash auto-attempts the recompile on first launch.
• The manual Retry on Neural Engine button stays as an override for once the budget is spent (still no re-download).

Telemetry

• ai_lens_ready now carries coreml_crashes (how many attempts the ANE took).

Carries v0.5.13 (in-app version in Settings) and everything prior.