Change proven defaults from el-salvador R&D testing:

CONTEXTUAL_STRATEGY: doc_level -> per_chunk
  Anthropic recipe — section-tailored context per chunk (1 LLM call/chunk
  but better embedding diversity than doc_level).

DATAPREP_CONTEXTUAL_DOC_BUDGET: 6000 -> 100000
  per_chunk needs the full doc visible per call (was tuned for doc_level).

RERANKING_STRATEGY: adaptive -> slice
  Adaptive's multiplicative novelty formula was structurally flawed
  (novelty cliff). Slice top-N by TEI score dominates on both recall
  AND precision. Offline sweep: slice recall 0.91 vs adaptive 0.69.

RERANKER_TOP_N: 1 -> 3
  Balanced compromise: top-1 = precision king, top-5 = recall king,
  top-3 = best of both. Also aligned chatqna default (was 2).

Updated: 4 code files, 2 test files, .gitignore for .tokensave/