Staging - #387

fix(retrieval): enable production node filter

…-cache

perf(retrieval): optimize map-nav snapshot and scoring

Replace character-level CJK tokenization with the shared jieba path and
bump map-unit index format to v2 so query hashes match stored tokens.
Add --tokens-only/--skip-current-format for lightweight reindex after
tokenizer changes without rebuilding manifests or namespace snapshots.

Co-authored-by: Cursor <cursoragent@cursor.com>

…nit-tokens

feat(retrieval): word-level map-unit BM25 tokens

…ew evidence text rendering

Deleted the legacy evidence rendering functionality and tests. Introduced a new method for rendering evidence text that groups documents and sections more effectively. Updated related components to utilize the new rendering approach.

…sset-inline

feat(retrieval): unify evidence text and placeholder asset inline

…g-readiness

perf(retrieval): optimize serving-index reads without changing quality

…ll-runtime-path

fix: make statistics backfill runnable in API image

…enames

fix(worker): normalize encoded document filenames

…ain-conflicts

perf(retrieval): add content trigram index