fix(eval): remove erroneous await on synchronous answer_question

test(rag): probe determinístico de presencia de cartas (guard del fix entity-aware)

fix(rag): auto-keywords no desplazan chunks semánticos en el contexto

refactor(rag): extrae _assemble_context y agrega keyword-eviction probe

Merge pull request #17 from GonzaloAsencio/perf/authority-boost-sim102

fix(eval): endurece el harness del juez (5 findings del code-review #17)

feat(eval): persiste respuesta completa y agrega modo --rejudge

feat(eval): timeout de judge configurable y override JUDGE_PROVIDER=gemini

perf(retrieval): probe Lever 2 lineage offline — medido NO gana @5 (59->53), solo rescata 1 miss a rank9 @10

perf(retrieval): probe Lever 1 FTS-keyword — medido PERDEDOR (regresa 59->47@5, no rescata semantic-gap misses)

perf(retrieval): miss_diagnosis script clasifica chunking-misses en granularidad vs gap semantico (Phase 0, LLM-free)

docs(scripts): aclara que la probe modela solo el brazo raw, no fuse_eq completo

feat(eval): --ids para correr batches disjuntos y armar un eval limpio por partes

perf(retrieval): baja authority boost a sim_102 (errata 1.02/patch 1.01) recupera +6pp recall@5

feat(eval): --limit para correr subset estratificado y no agotar la cuota LLM

chore(scripts): backup de corpus_chunks a CSV restaurable previo a re-ingest

feat(ingest): budget de chunk chico para el rulebook (estrategia A)

feat(llm): backoff+retry en 429 para HyDE, generacion y judge

feat(retrieval): RRF fusion raw+HyDE lifts recall@5 41%->59%

feat(eval): sonda de retrieval recall por rank y fuente

fix(eval): corrige falsos negativos en retrieval recall via rule_codes

feat(eval): agrega eval harness LLM-as-judge para RAG

chore(db): consolida fix_constraint en migracion 005 y borra scripts sueltos

test(scripts): smoke v2.0.0 y medicion de recall del indice

feat(ingest): metadata por set y chunking fino del rulebook

feat(corpus): parse errata y build_corpus por expansion

fix(cards): collapse cosmetic card variants by shared artwork

feat(cards): emit frontend lookup index from parse_cards.py

feat(rag): ingest cards as a new corpus_chunks source_type

feat(cards): add Riftcodex card parser and corpus generator

chore(corpus): add parser and migration scripts for v1.3.0

feat(corpus): add Unleashed Rules FAQ as corpus source v1.3.0

feat(rag): add @tag system and per-keyword chunks for precise retrieval

feat(rag): re-chunk rulebook by rule number for precise retrieval

feat: add validate_query script for retrieval validation

fix: use uuid5 for chunk IDs to match Supabase uuid column type

feat: add ingest pipeline script (chunking + embeddings + pgvector upsert)

feat: add errata parser (parse_faq.py)

feat: repo structure, corpus pipeline scripts, and PDF parser