v1.0.0-beta.12

What was fixed

• Structured evidence on every query_memory response. Each retrieved source now carries explicit source_file, lineno, name, entity_type, and a relevance score — not just buried inside a single flattened paragraph.

• Inline citations with scores. Claims in the synthesized answer now cite their source directly in the text — #file:line (score, L2 or rerank) — so you can see exactly which line backs which statement as you read, not just at the end.

• Labeled scoring (L2 vs. rerank). Every citation is tagged with which scoring method produced it, so a keyword-boosted rerank match and a pure semantic match are never confused for the same thing.

• Sources footer, consistent format. A clean bulleted reference list appears after every answer, listing all cited sources with file:line and score — easy to scan, easy to click through.

• Fixed empty-answer bug on reasoning-model queries. Queries that routed to the reasoning-enabled DeepSeek model (e.g. anything triggering deeper "architecture" analysis) previously could return a blank answer; now correctly falls back to the model's reasoning content when needed.

• Divide-by-zero guard on rerank scoring. Eliminated a bug where an exact/zero-distance match could produce an absurd, unusable rerank score.

• Empty-query handling. Submitting an empty query now returns a clean, helpful prompt instead of a malformed or garbage result.