Feat: capture inbox (Phase B) — phone captures merge into the library

Feat: /capture endpoint core (Phase A) — single-post extraction

Feat: living-library merge polish (Phase 3)

Feat: living-library cost-skip on re-extract (Phase 2)

Feat: living-library merge on import (Phase 1)

Fix: fast-fail on hosted URL fetch with clear error message

Fix: URL fetch works on hosted via Instagram's embed/captioned endpoint

Fix: Sonnet default, thinking-block skip, multi-venue URL extraction

Feat: single Instagram URL extraction via CAPTION_PROXY_URL

Revert: remove single Instagram post URL feature

Fix: API key gate for URL mode; faster timeouts; better network error message

Feature: chatbot (free) mode works for single URL extraction

Feature: extract a single place from an Instagram post URL

Fix KML round-trip: embed City/State in description CDATA; read them back on import

Fix: roulette location search matches city OR country; strip UNKNOWN from display

Web UI: session restore for returning users + food-only filter

Web UI: tab-bar SPA, KML/CSV import, refactor index.html

Fix: index 500 on HF — use new Starlette TemplateResponse signature

#19: public-URL hardening

#18: extend guardrail test for client-geocode + set-coords

#12: bound multi-user concurrency on the shared container

Fix: client-side geocoding for the chatbot path too

HF deploy: container hosted-readiness + client-side geocoding (#8–#11)

Demarcate the two deployment modes + add guardrails, CLAUDE.md

Hardening: multi-user safety, rename to places, geocoder cache, robust chatbot

Feature: Tier 2 Docker + env-gated Ollama for hosted deployments

Feature: Tier 1 hosting — BYOK + free chatbot handoff path

Feature: generalize beyond restaurants — category field + place extraction

Feature: batched + cached classification — 10× fewer API calls + prompt caching

Feature: roulette status filter, new-since-last-visit banner, README positioning

Feature: city trip roulette, multi-format exports, visited tracking

Feature: file upload validation + keyword pre-filter before LLM calls

Feature: in-app results browser (R2–R4 backend + R6–R11 frontend)

Results browser #R5 + #R1: geocode_one helper + GET /results endpoint

Feature: quality benchmark suite + CI regression test

Feature: Ollama local LLM provider + model quality/cost comparison in UI

Feature: richer extraction — cuisine, highlight, price, occasion, creator, saved_at

Reorganise into pipeline/, web/, data/ folders