Sprint 0+1: Package restructure, source tiers, verdicts, multi-variant editions

- New oracle/ package (11 modules) with unified CLI (python -m oracle)
- Source tiers: Tier 1 (arxiv/github/hf), Tier 2 (rss/hn), Tier 3 (reddit)
- Composite verdicts: PUBLISH/WATCH/ARCHIVE/DROP based on signal score + age
- Content-hash dedup: SHA-256[:16] normalized, atomic at insert time
- Multi-variant editions: 4 YAML configs (default/research/devops/brief)
- Variant engine: filter → rank → render (HTML + JSON, themed)
- Per-adapter timeout (10s) + threading fallback
- Consolidated 12 root scripts → thin wrappers + oracle/ package
- Archived stale scripts (_engagement, _live_compare, reddit_proof)
- Updated .gitignore, README.md, schema.sql

adapters/arxiv.py

@@ -355,6 +355,18 @@ class ArxivAdapter(SourceAdapter):
         if paper.get("_applied_domain"):
             tags.append(paper["_applied_domain"])
 
+        # Local-serving / efficient-inference signal (suggested source: arXiv cs.LG
+        # MoE/quantization papers → "local-serving" tag filter)
+        serving_kw = [
+            "quantiz", "quantization", "moe", "mixture of experts",
+            "serving", "efficient inference", "pruning", "distill",
+            "knowledge distillation", "low-rank", "lora", "parameter-efficient",
+            "vram", "memory efficient", "edge inference", "on-device",
+        ]
+        combined = f"{title_lower} {summary_lower}"
+        if any(kw in combined for kw in serving_kw):
+            tags.append("local-serving")
+
         return tags
 
     def fetch(self, query: str = "", limit: int = 20) -> list[dict]: