Prune QHexRT HNPU bundle to runtime-minimum artifacts

Remove 1 non-runtime/sensitive files from the QHexRT HNPU bundle.
Strip non-runtime _comment metadata from 1 manifest(s) without changing runtime fields.
Audit artifacts: QHexRT/hf_hnpu_audit cleanup plan, manifest reference check, and multi-agent review.

v81/README.md

@@ -1,10 +0,0 @@
----
-license: other
-base_model: nvidia/Llama-3.1-NemoGuard-8B-Topic-Control
-tags: [qualcomm, hexagon, npu, qnn, qhexrt, v81, llama, guardrail, topic-control]
----
-# Llama-3.1-NemoGuard-8B-Topic-Control — Hexagon NPU (v81)
-QHexRT-runnable, arch-pinned QNN bundle of NVIDIA's NemoGuard topic-control guardrail for **Hexagon v81**
-(SM8850, soc_model 87). LoRA on Llama-3.1-8B-Instruct, merged onto an ungated base, converted via QHexRT's
-**dense decode-split** (8 chained `decode_p*` parts + `llama_generate_sharded`). **Device-validated W8
-greedy-EXACT vs HF gold @ ~3.6 tok/s.** See `runanywhere/nemotron_nano_8b_HNPU` for the run recipe.

v81/nemoguard-topic-8b.json

@@ -1,6 +1,5 @@
 {
   "schema_version": 1,
-  "_comment": "Llama-3.1-Nemotron-Nano-8B-v1 (NVIDIA reasoning fine-tune of Llama-3.1-8B) on Hexagon v81 (SM8850 / soc_model 87). Standard dense Llama-3.1-8B (32 q heads -> materialized MHA + MAXCTX 512; llama3-scaled RoPE theta 500000 factor 8; untied lm-head). The ~7GB W8 decode exceeds the v81 cDSP per-context ceiling, so the 32 layers are split into 8 CHAINED decode parts (decode_parts=8, ~0.85GB W8 each) driven by the llama_generate_sharded host-op (decode-over-prompt seed, no batched prefill). Export gate: decode 8/8 match gold (forge). W8 device gate is greedy_tol.",
   "model": {
     "name": "nemoguard-topic-8b",
     "family": "llm",