fix: add safe analytics floor repair endpoint

fix: stabilize analytics totals across warm containers

fix: preserve exact catalog collections

fix: initialize new tenant during catalog ingest

fix: allow stable widget route in client iframes

production: harden catalog tenants analytics and widget install

fix: stabilize tenant routing and active db persistence

tests: tenant-scoped analytics + CSAT-scale charts, eval sqlite-first rows

reliability(§3): shared cross-container key-health store (modal.Dict)

eval panel: fix custom-tests 500, judge-key UX, product-question generation

security: production hardening — tenant isolation, email XSS, write limits, CSP

security(§5): key _ROW_CACHE by tenant+collection identity, not bare collection name

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

fix(catalog): bidirectional name-match in price/list path (typo robustness)

fix: make modal crawls durable under large writes

fix(retrieval): tighten quote-first and eval compatibility

feat(crawl): harden extraction cleanup and retrieval gating

fix(eval): correct product scoring and retrieval ranking

eval: derive tenant identity from runtime evidence

eval: trust runtime-grounded tenant rows

eval: fall back to runtime chunk seeds

eval: derive product questions from chunks

eval: tighten tenant scope matching

eval: fix chat retrieval scoring and tenant scope fallback

eval: harden live run output and grounding filter

eval: tighten tenant scoping and fallback diagnosis

eval: respect requested generate and run counts

eval: tighten tenant filter and judge guidance

eval: confirm trace-backed judge verdicts locally

eval: deepen judge trace certainty

eval: trust tenant-local kb rows and tighten analytics filter

eval: allow request judge key and set row statuses

eval: fallback to kb when saved eval set is empty

eval: add workflow-trace evidence to llm judge

eval: filter stale eval sets and diagnose idk rows

eval: filter cross-tenant analytics seeds

eval: add prompt-aware judge diagnostics

eval: add separate Groq judge layer

evals: add context recall and likely-cause hints

evals: improve live scoring and admin diagnostics

evals: add ranked retrieval and grounded answer metrics

evals: harden kb question quality and grading consistency

evals: remove cwd assumptions and expand coverage

fix: patch _owner_eval_blocker in evals smoke test

evals: tighten grading and harden standalone suite

eval: harden DB-grounded suite and rotation

evals: add owner RAG scoreboard UI + API

refactor: extract services/config.py (session 2)

fix: sync-github owner-only + correct test failures found by running pytest