Hound v11.2.0

Six-signal search ranking + source type detection + 41% latency reduction

Added

• Six-signal ranking: title relevance (+0.10) + URL relevance (+0.08) on top of neural + consensus + domain reputation + answer signals. All zero-latency.
• Source type detection: every result classified as docs/paper/repo/blog/forum/reference/news/other. Surfaced in fetch_hint.
• Snippet merging: same URL from multiple engines = merged snippets (capped 600 chars). Richer info per result, zero extra fetches.
• Smarter fetch_relevance tiers: consensus + domain + answer signals promote quality results to high tier.
• Heading-aware BM25: blocks under matching headings get 1.5x boost. Tables/code with query terms always preserved.
• Neural rerank pre-filter: scores top 12 instead of all 30+ results. 41% latency reduction (5.8s -> 3.4s).
• Faster search deadline: 8s -> 5s per engine. Soft deadline 2s with early-return-on-quorum.

Removed

• include_content: removed entirely. Failed experiment: added 6-10s latency + 33k token damage per session with no quality improvement.

Tests

• 62 new adversarial tests. Total: 466 tests.