#70-Debug: Diar Weak-Fold (7->4 Sprecher), LLM-Robustheit + Whole-Pass-Modellgrenze, Thermal-Cap fuer analyzeRecording

- diar.js: Stufe-3 Weak-Fold (DIAR_MIN_TOTAL_SEC 20s) gegen Phantom-Sprecher
- llm.js: End-Instruktion, JSON-Repair, Sanity+Fallback, Chunk-Retry, Quote-Sanitizer, gelabelter/gewichteter Reduce; WHOLE_MAX 20000 (Gemma-4B-Grenze)
- rkllm_server.cpp: max_context_len 8192->12288
- server.js: analyzeRecording unter A76-Takt-Cap (95-97 Grad ungedeckelt)
- Doku Kap. 03/04/05 + Changelog

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0113vDKdSNtD3Dwcw4QMu9yV

app/rkllm_server.cpp

@@ -49,7 +49,7 @@ int main(int argc, char** argv) {
     if (argc < 2) { fprintf(stderr, "usage: %s model.rkllm\n", argv[0]); return 1; }
     RKLLMParam p = rkllm_createDefaultParam();
     p.model_path        = argv[1];
-    p.max_context_len   = 8192;
+    p.max_context_len   = 12288;
     p.max_new_tokens    = 1536;
     // Gemma-3-4B (w8a8) degenerates into repetition loops ("X, X, X…") and echoes
     // the prompt under pure-greedy decode → needs real sampling + strong repeat/