Ground Oghma retrieval in catalog entities

[RANGROO]

Summary

• extract Oghma topics locally from canonical names and aliases before lore retrieval
• tolerate guarded STT spelling/phonetic errors while rejecting ambiguous neighbors, speaker labels, and wrong-sense homonyms
• keep repeated central topics and multi-topic mention order, including ordinary creature and object mentions
• fetch the selected canonical article exactly; use MiniMe/custom extraction only as a constrained fallback when no local topic is grounded
• add related-concept tags as a last-resort, exact-only contextual fallback after canonical/alias extraction abstains
• allow a unique multiword tag to select its sole article; require two shared tags plus explicit knowledge intent when a tag belongs to two or three articles
• reject one-word tags, phrases owned by more than three articles, fuzzy/corrupted tags, and tag-only speaker labels
• prefer longer overlapping tag phrases so common nested phrases can block unsafe matches without becoming candidates
• expose a test-only fallback toggle through the existing extractor function's optional argument; production calls keep tag fallback enabled

Validation

• unchanged frozen conversational/STT suite: 500/500
  • 100% required recall, relevant precision, and primary ordering
  • zero forbidden retrievals and zero over-retrieval
  • p50 13.97 ms, p95 26.14 ms, p99 33.46 ms end-to-end
  • retrieval-only p50 0.25 ms, p95 0.40 ms, p99 0.53 ms
• catalog-derived tag-policy A/B on identical 500 fixtures:
  • previous product library: 260/500; all 240 safe tag-positive cases absent
  • tag-aware product library: 500/500
  • p50 10.26 ms, p95 16.95 ms, p99 19.07 ms (previous p95 16.91 ms)
  • 160 unique multiword positives, 80 corroborated shared-tag positives, and 260 ambiguity/corruption/speaker-label/canonical-precedence guards
• 7/7 standalone tag-policy probes
• temporary candidate catalog focused retrieval: 41/41; p95 13.96 ms
• php -l lib/oghma_retrieval.php
• git diff --check
• final deployed runtime library SHA-256 matched the committed worktree and passed the tag-policy suite 500/500 (p95 16.88 ms)

Limits

• transcript-level validation only; microphone/upstream STT was intentionally not tested
• no live MiniMe calls were needed by these deterministic suites
• not tested in Skyrim
• the external harness and generated tag-review tooling remain outside this repository

[RANGROO]

Superseded by consolidated draft #701, which contains this work plus the completed Oghma catalog, retrieval, knowledge-class, biography, parity, and UI changes. Please review the combined draft there.