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 2 manifest(s) without changing runtime fields.
Audit artifacts: QHexRT/hf_hnpu_audit cleanup plan, manifest reference check, and multi-agent review.

config.json

@@ -1,7 +0,0 @@
-{
-  "model_type": "qhexrt",
-  "name": "whisper_small_HNPU",
-  "framework": "qhexrt",
-  "runtime": "qnn-htp",
-  "_note": "Root config.json present so the Hugging Face Hub counts downloads (default query file for repos without a registered library)."
-}

v79/whisper-small.json

@@ -1,24 +1,41 @@
 {
   "schema_version": 1,
-  "_comment": "whisper-small - Qualcomm Hexagon v79 (QHexRT) deployment manifest.",
-  "model": { "name": "whisper-small", "family": "asr", "dsp_arch": "v79" },
+  "model": {
+    "name": "whisper-small",
+    "family": "asr",
+    "dsp_arch": "v79"
+  },
   "params": {
-    "vocab": 51865, "n_layers": 12, "hidden": 768, "head_dim": 64,
-    "max_ctx": 200, "eos_token_id": 50257
+    "vocab": 51865,
+    "n_layers": 12,
+    "hidden": 768,
+    "head_dim": 64,
+    "max_ctx": 200,
+    "eos_token_id": 50257
   },
   "artifacts": {
     "contexts": {
-      "encoder": { "bin": "encoder.bin" },
-      "decoder": { "bin": "decoder.bin" }
+      "encoder": {
+        "bin": "encoder.bin"
+      },
+      "decoder": {
+        "bin": "decoder.bin"
+      }
     },
     "tokenizer": "tokenizer.json",
     "fixture_dir": "."
   },
   "plan": {
     "steps": [
-      { "host": "asr_transcribe",
-        "params": { "encoder": "whisper_small_encoder", "decoder": "whisper_small_decoder",
-                    "mel_filters": "whisper_mel_filters.bin", "sot": 50258 } }
+      {
+        "host": "asr_transcribe",
+        "params": {
+          "encoder": "whisper_small_encoder",
+          "decoder": "whisper_small_decoder",
+          "mel_filters": "whisper_mel_filters.bin",
+          "sot": 50258
+        }
+      }
     ]
   }
 }

v81/whisper-small.json

@@ -1,6 +1,5 @@
 {
   "schema_version": 1,
-  "_comment": "Whisper-small ASR (encoder-decoder). Host log-mel + host conv stem (the conv is HTP-hostile in-graph) feed the encoder graph (input encoder_features); the autoregressive decoder (causal self-attention + cross-attention to the cached encoder states, learned positional embeddings, tied lm-head) emits tokens, detok'd on the host. Gated on WER vs HF. Same proven asr_transcribe host-op path as whisper-base (12 layers / d768 / 12 heads).",
   "model": {
     "name": "whisper-small",
     "family": "asr",