April 25: ProcessMemoryEnforcer, OCROptimizer, PagedBlockPool deadlock fix, test suite and UI cleanup

[cnshsliu]

Two April 25 commits plus working-tree refinements.

ProcessMemoryEnforcer

Actor-based memory pressure manager, 1-second polling loop. Configured via ServerConfig.maxProcessMemory — four modes:

• auto — reserves max(4GB, min(8GB, physMem/5)) for the system, rest for NovaMLX
• percent — e.g. 80% of physical RAM
• fixed — absolute: 24GB, 4096MB
• disabled — no enforcement

Worker subprocess reports memory stats to the supervisor every 5 seconds. Settings page has a mode picker GUI with conditional percent/value inputs.

OCROptimizer

Auto-detects OCR models (deepseekocr, dots_ocr, glm_ocr) and injects optimal defaults: prompt prefix, sampling parameters (temperature=0, top_p=1, etc.), and extra stop sequences. Detects whether the user already provided text content to avoid redundant prompt injection.

PagedBlockPool deadlock fix

Nested locking in PagedBlockPool was causing crashes (70+ occurrences). Fixed by extracting freeBlockUnlocked and incrementRefUnlocked private methods that operate without acquiring the lock — callers hold it once.

Test suite consolidated

swift test now covers all 125 tests across all targets. VLM integration tests (base64 image, HTTP URL image) are in-tree and gracefully skip when no server is running. Fixed a misleading filename: E2ETests.swift → RateLimiterTests.swift — it was rate limiter unit tests, not E2E. Real end-to-end HTTP tests live in the separate NovaMLXE2ETests target.

UI fixes

• Status page: removed the "Active Models" section (Models tab already shows this)
• "Inference Speed" → "Real-time Inference Speed" with all 9 language translations updated
• Agents page: fixed the arrow.clockwise icon animation that kept flying in from screen center. Root cause: SwiftUI .animation(nil, value:) doesn't cancel — it triggers an implicit transition. Replaced with independent rotationAngle state controlled via withAnimation blocks.

Documentation

• 8 new sections in features.md / features.zh-CN.md: Worker Subprocess Isolation, Control Token & Thinking Filtering, Speculative Decoding, Agent Integration, Memory Management, Cloud Model Proxy, Internationalization, Homebrew Distribution
• Added VLM image recognition curl examples (base64 data URI and HTTP URL)
• Architecture diagram updated