Scrub internal engineering notes from manifest _comment

v81/internvl3_5-1b.json

@@ -1,6 +1,6 @@
 {
   "schema_version": 1,
-  "_comment": "InternVL3.5-1B (InternViT-300M vision + Qwen3-0.6B LLM) image->caption on Hexagon v81 (SM8850 / soc_model 87) via the internvl_generate host-op. 448px image -> 256 image tokens (host patch-embed + pixel-shuffle + MLP projector) spliced into the text prompt, right-padded to the prefill window Pn=512, one batched prefill seeds the decode KV, then AR=1 decode + in-graph lm-head. The recipe's prefill graph (ivl_pf, batched GQA) and lm-head graph (ivl_lmh) were reconstructed (export_llm_prefill.py / export_lmhead_graph.py) — gated vs the HF gold first token. All graphs f16, f16 I/O. Compiled via the DLC route + {O:3,vtcm_mb:8,dlbc:1}.",
+  "_comment": "internvl3_5-1b - Qualcomm Hexagon v81 (QHexRT) deployment manifest.",
   "model": { "name": "internvl3_5-1b", "family": "vlm", "dsp_arch": "v81" },
   "params": {
     "hidden": 1024, "vocab": 151936, "n_layers": 28, "max_ctx": 512,