production: harden catalog tenants analytics and widget install

coref: LLM rewrite is primary for subjectless follow-ups (ends the regex treadmill)

coref/minmax: fix embedded-number name truncation + Urdu anchor pollution

coref: make the set-reference/projection class language-agnostic (Roman-Urdu)

coref: structurally resolve the prior-result-set follow-up CLASS

fresh2 100Q fixes: count_split phrasing + ordering pole-strip + coref category re-scope + compare name-strip

coref: fix follow-up cascade exposed by Codex independent 100Q (6 fails)

coref: both-stock follow-up uses 'X or Y' not 'X and Y' (Q65 determinism)

coref: fix Roman-Urdu/English follow-up cascade (100Q Q62/63/65/67)

fix: 3 remaining diecast live fails — Roman-Urdu compare verdict, stock-status phrasing, superlative follow-up

fix: 3 more universal class fixes (Q10/Q29/Q44 from live diecast eval)

fix: 8 universal answering-class fixes from fresh diecast 50Q eval

catalog: LLM-primary understanding for code-mixed (Roman-Urdu) questions

catalog_query/coref: universal fixes for remaining DIECAST round-3 fails

coref: fix follow-up resolution — compare-continuation, verbose antecedents, plural set-reference

fix: respect category in stock list queries

fix(coref): don't bind a pronoun to policy/logistics prose

fix(catalog+coref): strip follow-up adverbs so "do you also stock X" resolves X

feat(chat): request-scoped follow-up coreference + superlative/greeting parse fixes