PROJECT-STATUS.md

Perth Pint Prices Project Status

Last updated: 2026-07-18

What this is

Perth Pint Prices (perthpintprices.com) tracks pint prices across 857 Perth pubs (up from 423 in March). Users can discover cheap pints, find happy hours, and report prices. The site is live on Vercel.

Stack, database, routes, components, and lib files are documented in CLAUDE.md (auto-loaded every session). This file covers history, recent work, and the backlog.

What's done recently

World Cup knockout fixture freshness through final (2026-07-18)

• Commit 8ee1ec0: refreshed the FIFA 2026 knockout fixtures from the official FIFA fixture feed. Corrected Mexico v England to 9am AWST, confirmed the remaining quarter-finals (Spain v Belgium, Norway v England, Argentina v Switzerland), both semi-finals, the third-place play-off, and the final (Spain v Argentina). Local naming style remains preserved while TIMES_CHECKED now reflects 18 July 2026.
• Regression coverage: worldCup.test.ts now locks the confirmed quarter-finals, semi-finals, third-place play-off, final, the Mexico v England kickoff correction, and team-colour coverage for every confirmed team in the fixture list.
• Verified: FIFA feed comparison returned zero differences for knockout matches M73-M104; npm test -- --runInBand (346/346), npx tsc --noEmit, and Playwright homepage screenshots at /tmp/perth-pint-prices-worldcup-2026-07-18/home-desktop.png and /tmp/perth-pint-prices-worldcup-2026-07-18/home-mobile.png.

Pub pages linked to their Google listing via schema.org sameAs (2026-07-06)

• Commit 6391d4a: each pub's BarOrPub JSON-LD now emits sameAs + hasMap pointing at the canonical Google Business Profile (built from the stored place_id), plus telephone from the stored phone number. place_id/phone are surfaced through PUB_FULL_COLUMNS and toPub() — deliberately kept out of PUB_LIST_COLUMNS so the big list pulls stay lean. Pubs without a place_id degrade gracefully (lat/lng hasMap, no sameAs). Reconciles each pub page to its real-world Google entity; pub social profiles (FB/IG) are still not collected, so they remain out of sameAs.
• Commit d45277e: restored "target": "ES2017" to tsconfig.json. The config mirrors the Next 14 create-next-app default but had lost its target line, so tsc defaulted to ES3/ES5 and rejected a Set iteration in worldCup.test.ts (TS2802) — making npx tsc --noEmit impossible to pass cleanly. lib is unchanged and next build uses SWC (browserslist-driven), so this only touches tsc's downlevel typecheck; build output is unaffected.
• Verified: npx tsc --noEmit clean (a stale tsconfig.tsbuildinfo from incremental: true masked the fix until cleared), full unit suite 345/345 via tsx --test. New pubJsonLd.test.ts cases lock the sameAs/hasMap place-id path, the lat/lng fallback, and telephone.

World Cup knockout fixture freshness (2026-07-05)

• Commit 7bba585: refreshed the FIFA 2026 knockout fixtures from the official FIFA fixture feed. Confirmed the remaining Round of 16 slots as Mexico v England, Portugal v Spain, USA v Belgium, Argentina v Egypt, and Switzerland v Colombia, and confirmed quarter-final M97 as France v Morocco. Future quarter-finals, semi-finals, third-place play-off, and final remain bracket placeholders where FIFA has not confirmed both teams.
• Regression coverage: worldCup.test.ts now locks the full confirmed Round of 16, the first confirmed quarter-final, knockout kickoff changes, and team-colour coverage for every confirmed team in the fixture list.
• Verified: npm test -- --runInBand (342/342), npx tsc --noEmit, and Playwright homepage screenshots at artifacts/world-cup-fixtures-2026-07-05/home-desktop.png and artifacts/world-cup-fixtures-2026-07-05/home-mobile.png.

World Cup Round of 16 fixture freshness (2026-07-04)

• Commit ebf8761: refreshed the FIFA 2026 knockout fixtures from the official FIFA fixture feed. Confirmed Round of 16 slots for Mexico v England, Portugal v Spain, USA v Belgium, Argentina v Egypt, and Switzerland v Colombia. Later knockout rounds remain bracket placeholders because FIFA still lists them as unresolved winners/losers.
• Regression coverage: worldCup.test.ts now locks the fully confirmed Round of 16 fixtures and continues to require team-colour coverage for every confirmed team in the fixture list.
• Verified: npm test -- --runInBand (341/341), npx tsc --noEmit, and Playwright homepage screenshots at 1280×800 + 375×812.

World Cup Round of 16 fixture freshness (2026-07-03)

• Commit 279c5f9: refreshed the FIFA 2026 knockout fixtures from the official FIFA fixture feed. Confirmed Round of 16 slots for Mexico v England, Portugal v Spain, USA v Belgium, and Switzerland v Winner M87. Kickoff times still match FIFA and unresolved future sides remain bracket placeholders.
• Regression coverage: worldCup.test.ts now locks the newly confirmed Round of 16 teams while the existing team-colour coverage guards every confirmed side in the fixture list.
• Verified: FIFA feed comparison returned DIFFS=0; npm test -- --runInBand (341/341), npx tsc --noEmit, and Playwright homepage screenshots at 1280×800 (/tmp/perth-pint-prices-screenshots/world-cup-home-desktop-2026-07-03.png) + 375×812 (/tmp/perth-pint-prices-screenshots/world-cup-home-mobile-2026-07-03.png).

World Cup Round of 16 fixture freshness (2026-07-02)

• Commit 4c3ac2f: refreshed the FIFA 2026 knockout fixtures from the official FIFA fixture feed. Confirmed Round of 16 slots for Mexico v England and USA v Belgium, updated TIMES_CHECKED to 2 July 2026, and kept unresolved future knockout slots as bracket placeholders.
• Regression coverage: worldCup.test.ts now locks the newly confirmed Round of 16 teams while the existing confirmed-team colour coverage guards England, USA, and Belgium.
• Verified: npm test -- --runInBand (341/341), npx tsc --noEmit, and Playwright homepage screenshots at 1280×800 (/tmp/world-cup-home-desktop-2026-07-02.png) + 375×812 (/tmp/world-cup-home-mobile-2026-07-02.png).

World Cup knockout fixture freshness (2026-07-01)

• Commit 7e30777: refreshed the FIFA 2026 knockout fixtures from the official FIFA fixture feed. Updated the Mexico v Ecuador Round of 32 kickoff to 10am AWST, confirmed Round of 16 slots for Canada v Morocco, Paraguay v France, Brazil v Norway, and Mexico v Winner M80, and moved quarter-final M100 to 9am AWST. Preserved local naming style (Ivory Coast, DR Congo, Cape Verde, Bosnia & Herzegovina) while leaving unresolved knockout sides as bracket placeholders.
• Regression coverage: worldCup.test.ts now locks the confirmed Round of 16 fixtures, FIFA-confirmed knockout kickoff changes, and team-colour coverage for every confirmed team in the fixture list.
• Verified: npm test -- --runInBand (341/341), npx tsc --noEmit, and Playwright homepage screenshots at 1280×800 + 375×812.

Microsoft Clarity analytics added (2026-06-30)

• Commit 3a87b56: added Microsoft Clarity (project xez4co2ysd) for session recordings and heatmaps. The tag loads via next/script with strategy="lazyOnload" in the root layout, mirroring the existing Google Analytics setup so it stays off the critical path.
• Verified: npx tsc --noEmit clean; a headless load check confirmed window.clarity initialises and the tag fetches clarity.ms/tag/xez4co2ysd at runtime.

Removed unowned sameAs profile links (2026-06-29)

• Commit 3ecd439: removed Facebook, Instagram, TikTok, and X links from the Perth Pint Prices Organization.sameAs schema because those arvopints profiles are not owned/controlled by the site owner. sameAs is now omitted until there are confirmed official profiles. Added a regression test so sameAs only returns when verified owned profiles exist.
• Verified: npm test -- --runInBand (338/338) and npx tsc --noEmit.

Site-wide brand schema + sameAs cleanup (2026-06-29)

• Commit 0974a62: moved the canonical Organization + WebSite JSON-LD graph into the root layout so the Perth Pint Prices brand entity is present site-wide, not only on the homepage. Added a small JsonLdScript helper that escapes < per Next.js JSON-LD guidance.
• sameAs follow-up: the initial profile links were removed in 3ecd439 because they were not owned/controlled Perth Pint Prices profiles. Homepage JSON-LD stays focused on the homepage FAQ and links back to the site-level WebSite via isPartOf.
• Article schema now references the same brand entity: author and publisher point at https://perthpintprices.com/#organization rather than creating disconnected anonymous Organization nodes. Verified with npm test -- --runInBand (338/338) and npx tsc --noEmit.

World Cup kickoffs on pub pages + knockout fixtures (2026-06-17)

Added a PubWorldCup card to pub detail pages (/[suburb]/[pub]) — the one World Cup surface that was missing (the /world-cup hub, HomeWorldCup strip and worldCup.ts engine already shipped 2026-06-10). Decided via a quick sub-agent fan-out (data signals / product / SEO / architecture) rather than guessing.

• Extended WC_FIXTURES from 72 → 104: added all 32 knockout fixtures (Round of 32 → final) in AWST, cross-checked 17 June against the official schedule. Teams are bracket slots ("Winner Group A", "Winner M73") until results come in — dates/times/order are fixed. New optional round field on WcFixture; group fixtures leave it unset. Without these, the pub widget would have gone dark after the group stage ends ~28 June (10 days of runway), missing the entire knockout phase.
• PubWorldCup is a server component, passed into PubDetailClient as a worldCup slot prop so the fixture data stays off the client bundle for all 850+ pub pages; only the existing WorldCupCountdown is a client island. Shows the next 3 upcoming kickoffs with Perth time, day heading, knockout round, the tradingStatus permit/early/normal badge, and a "Next · countdown" ticker.
• Gated on goodForWatchingSports === true || hasTab === true. Live Supabase counts: has_tab is near-dead (56 pubs, 6.6% — hand-curated, never seeded), Google's good_for_watching_sports is the real signal (228, 27%); union lights up 250 pubs (29%). Non-sport pubs render nothing (no misleading "watch here" on a wine bar). No SportsEvent/BroadcastEvent schema — would falsely imply the pub hosts the match; honest framing only ("opening for a given kickoff is up to the venue").
• Self-retiring: new shared WC_LAST_DAY = '2026-07-20' (Perth date of the final, which kicks off 3am AWST — the old HomeWorldCup LAST_DAY = '2026-07-19' would have hidden everything during the final; now both import the shared constant). Card returns null once the tournament is past or no upcoming fixtures remain — no manual cleanup. Hub empty-state copy updated (knockouts are listed now, so "lands once the bracket settles" was stale).
• Homepage strip now shows only current games: HomeWorldCup previously rendered the three Socceroos fixtures regardless of whether they'd been played (so a finished game sat at the top of the homepage). Now it shows the next few kickoffs that haven't finished yet — hydration-safe via a mounted clock (mirrors WorldCupFixtures): server/pre-hydration render uses today-onwards (fixtureDay >= today, stable for hydration), then drops matchPhase === 'played' once the client clock is known. Australia's games keep a Socceroos tag; knockout games show their round.
• 2026-06-29 follow-up (5e1920c): tightened the homepage strip again so it shows only future kickoffs, not matches already live. Added upcomingFixtures() plus regression tests proving the strip uses the full fixture list rather than Australia-only.
• 2026-06-29 fixture correction (60d2b86): replaced all 16 Round of 32 bracket placeholders with the confirmed FIFA fixture-feed teams, so the homepage strip no longer shows "Winner C v Runner-up F" style slots or neutral grey fallback stripes. Added tests that lock the Round of 32 matchups and require team colours for every side.
• Verified: tsc clean, 332/332 tests pass (worldCup tests updated to assert 104 = 72 group + 32 knockout, chronological order, final on 2026-07-20), lint clean. Playwright screenshots at 1280×800 + 375×812 against a live sport pub (The Royal on the Waterfront) — card renders with real upcoming fixtures + ticking countdown; confirmed a non-sport pub (Parley Bar) renders nothing. PR #204 (CI green); an independent agent reviewed the live Vercel preview (desktop + mobile, console-error capture) and returned PASS with no blockers.

PageSpeed mobile fixes — perf + a11y, zero SEO change (2026-06-16)

PR #201 (+ follow-ups #202 image, #203 happy-hour contrast). Worked the mobile PageSpeed report. The PSI API was quota-blocked at the project level (defaultPerDayPerProject = 0), so audits were captured by running lighthouse directly against the live URL. Live production after deploy: Performance 72→88, Accessibility 77→100, Best Practices 96→100, SEO 100 (unchanged); FCP 3.6s→1.8s, LCP 5.3s→3.5s, CLS 0.001.

• Perf: AVIF/WebP images.formats in next.config.js (article PNGs −121→−69 KiB); Google Analytics afterInteractive→lazyOnload (160 KB gtag off the critical path — the main FCP/LCP win); display:'swap' on all three next/font families.
• A11y (→100): aria-label on the two FilterSection <select>s + the icon-only Near/$ buttons; min-w-[44px] tap targets (were 23px); tabIndex={-1} on the aria-hidden #ssr-links anchors (hrefs untouched, still crawlable); new amber.deep #A85A00 token for small amber text, ink-on-amber for small badges/buttons (PubCardList cheapest pill + view-all button, HeroSection cheapest-card eyebrow), footer faint-white bumped to AA. text-amber→text-amber-deep in ArticleRail + HomeWorldCup.
• Best Practices: dropped the on-load navigator.geolocation request in HomeClient (the manual "Near" button already triggers it) — fixes geolocation-on-start; productionBrowserSourceMaps: true for valid-source-maps (note: one residual flag from the App-Router inline bootstrap script; chunk maps are emitted + served).
• Reverted mid-work: a modern browserslist — it didn't move the 11 KiB legacy-JS chunk (third-party, Next won't re-transpile), so it wasn't shipped.
• Image follow-up (#202): the 3 homepage article-rail thumbnails were AVIF but served at q=75/sizes=100vw (~82 KB each for a ~360px slot). ArticleImageSlot now uses quality={60} (≈48% smaller, ≈42 KB each, still AVIF, visually clean for below-fold decorative thumbs) + sizes mobile 100vw→calc(100vw - 3rem) so low-DPR devices fetch the w=640 tier. LCP is the hero <h1> text, bounded by the ~780 KB crawlable HTML payload — left alone (SEO).
• Happy-hour contrast follow-up (#203): a live re-audit (run while a pub was in happy hour) caught two color-contrast fails that only render in HH state, absent from the off-peak audit: the HomeClient Live-HH banner price (type-price text-amber, 3.33:1 → text-amber-deep) and the small text-red-on-bg-red-pale "HH"/status badges (4.47:1, a hair under AA) across PubCardList/PuntNPints/SunsetSippers/TonightsMoves/SuburbLeague. Added a red.deep #BA3829 token (mirrors amber.deep) used only on those badges (4.9:1); brand red.DEFAULT untouched. Gotcha for future a11y passes: state/time-dependent elements (happy-hour, crowd) don't render off-peak, so a single audit can miss them.
• No URLs, redirects, canonicals, metadata or crawlable links changed. tsc clean, prod build green, screenshots eyeballed at mobile (375×812) + desktop (1280×800).

Milestone backlog: five issues cleared autonomously (2026-06-12)

Worked the open-milestone backlog (Andrew left untouched by request). Four PRs:

• #193 — Architecture Refactor #57/#58/#59: (a) consolidated price formatting into src/lib/formatters.ts (deleted the src/-root formatPrice + src/lib/priceLabel; both were near-dead — formatPrice only its own test, priceLabel zero importers); (b) dropped the PerthNow.date footgun (a +8h-shifted Date that double-shifts on a non-Perth host; no caller read it) and added a process.env.TZ-flipping test proving host-timezone independence; (c) one price_snapshots read seam (src/lib/priceSnapshots.ts) routing PintIndex / admin-stats / pint-index API / CSV export through shared ordering+limit+coercion.
• #194 — Data Coverage #62: docs/stale-price-refresh-strategy.md (thresholds reusing the freshness.ts tiers, suburb/cheap-outlier prioritisation, weekly+seasonal+crowd-loop cadence) + scripts/stale-venues-queue.sql (read-only ranked queue for the next batch). The widen-pintsweep-to-stale change is flagged as a follow-up, not done (Andrew-adjacent).
• #195 — CI guard #80: duplicate-cluster fingerprint guard. src/lib/pubFingerprint.ts reduces each prerendered pub page to a value-stripped content shell (headings + section count, name/digits tokenised); scripts/check-pub-fingerprints.mjs scans .next/server/app after Build and enforces a distinct-shell floor + largest-cluster ceiling. Current build: 176 pub pages → 39 distinct shells, largest cluster 21%. Wired into CI.
• Held for you (need real-world input, not code): #61 price backfill (real data), #28 GA4 key events (dashboard), #33 press pitch (outreach), #36 GSC alerts, #30/#31/#34/#75 (GSC data + recrawl/backfill prerequisites), #32/#72/#118 (content judgement + per-page visual verification), #66 (major-version upgrade), #79 (AI-extraction + schema), #60 (one useUserLocation hook — implementable but its acceptance needs manual geolocation-prompt verification and HomeClient has a 4-state machine incl. dismissed).

Stale-price flag fixed — it had never worked (2026-06-12)

• Slack notifications confirmed end-to-end: after the #191 redeploy, a marked test report posted to the live API produced a Slack ping with no SLACK_WEBHOOK_URL warnings/errors in the runtime logs. The test report ($14.50 "Test Lager" at The Norfolk Hotel by "Claude (webhook test)") needs rejecting in /admin.
• The same test exposed a day-one bug: the "Report Issue" flow (SubmitPubForm → outdated: true) 500s in production because the API inserts reported_price: 0 as a sentinel, which violates the price_reports_reported_price_check DB constraint. Zero outdated_flag rows have ever existed in prod — every stale flag a user submitted was lost. Diagnosed via Vercel log substring probing (the log search matches full message text even though the display truncates it).
• Fix: outdated flags now insert reported_price: null (check constraints pass on null); new migration supabase/migrations/20260612000000_stale_flag_null_price.sql drops NOT NULL on the column in case it's set — run it in the Supabase SQL editor before relying on the fix. Admin queue shows a "stale flag" badge for these reports instead of a fake $0.00. Admin review already special-cased outdated flags (mark reviewed, never touch the pub price), so approval-side behavior is unchanged.

PR backlog cleared + Slack notifications go-live (2026-06-12)

• All 8 open PRs resolved: merged #188 (egress fix), #190 (consolidated safe dependabot bumps: supabase-js 2.107, openai 6.42, radix-ui 1.5, @types/node 24, upload-artifact v7), #176 (AI article drafting pipeline + dev-only /draft preview), #189 (app-store plan doc). Closed #168 (react-leaflet-cluster 4 needs React 19 — parked with Next 15/Tailwind 4 migration), #163/#170 (superseded), #175 (stale).
• Slack price-report notifications live-ish: the incoming-webhook integration was added to #all-perth-pint-prices (2026-06-12 8:31am) and verified working. SLACK_WEBHOOK_URL believed added to Vercel ~11:17am — after the last production deploy, so this docs commit doubles as the redeploy that activates it. Notification paths: instant ping on each new price report (/api/price-report) + 8am pending-queue reminder (/api/cron/price-check).

Supabase egress fix: shared cached pubs pull + kill browser poll (2026-06-11)

• Why: free-tier egress warnings. Measured through a local byte-counting proxy: one production build = 167.2MB / 1,108 requests (450 full-table-ish select=* pulls — suburb pages pulled the full table up to 8x per render via getSuburbBySlug + getSuburbPubs + getNearbySuburbs + getSiteStats); 7 list pages cost 4.4MB per 5-min revalidation window; and HappyHourClient re-fetched the entire table from every visitor's browser every 60s (~122MB/hour per open tab).
• Fixes: new src/lib/cachedPubs.ts — unstable_cache-shared raw-row pulls (1h TTL, tag pubs), with toPub() mapped fresh per render so live happy-hour status never staleness; ~20 server routes + suburb/pub-page aggregates switched to it. select('*') replaced with explicit column lists; google_opening_hours (~735KB, 35% of a pull) excluded from the shared list pull (keeps the cache entry under Vercel's 2MB data-cache limit) — happy-hour pages get a second small full-column pull filtered to happy-hour pubs; pub detail keeps full columns via getPubBySlug. HappyHourClient poll deleted (live state was already clock-derived). /api/admin/review calls revalidateTag('pubs') on approvals so price changes don't wait out the TTL.
• Measured after: build 1.7MB / 240 requests (−99%); warm cache: 6 different pages = 1 Supabase request; per-pull 377KB → 249KB compressed. Steady state ≈ 6MB/day of hourly refreshes vs multiple GB/day.
• Verification: independent agent run — tsc clean, 309 tests pass, diff review (no correctness bugs), 12 routes content-checked against the built app, exact venue parity with production (/happy-hour 168=168 with identical link sets, /fremantle 63=63), screenshots eyeballed. Its three minor flags (uncached TransportHubPage, dead Pub.source field, stale "auto-refreshes" copy) fixed in 47a69d3.
• Gotcha for future measurement work: an EADDRINUSE'd next start left a stale server on the port and silently invalidated the first after-measurement — caught because the proxy log showed select=* queries that no longer existed in the code. Check the server actually started.
• Commits 2f77fef + 47a69d3.

Pint Signal: stale-answer fix + weekly sweep (2026-06-11)

• Answers vanished after posting — force-dynamic doesn't stop Next's data cache memoising supabase-js fetches, so the 30s poll served a stale answers list and wiped the optimistic row (the answer was always in the DB). fetchCache = 'force-no-store' on /signal/[id] page + GET route, cache: 'no-store' on the client poll.
• Signals now delete themselves — the daily price-check cron sweeps signals whose expires_at is >7 days old (dead links show the burned-out state for a week, then the rows go). noindex + robots Disallow: /signal/ confirmed already live.

Pint Signal Phase 1 (2026-06-10)

• The feature: one mate lights the signal (pub + time), the crew gets the link, everyone answers IN/OUT with one tap; signals burn out 3h after the meet time. No accounts, no push — the group chat is the delivery (share sheet). Plan: docs/pint-signal-plan.md; prototype: docs/prototypes/pint-signal.html.
• New: signals + signal_answers schema (scripts/pint-signal-schema.sql — must be run in the Supabase SQL editor before the feature works), signalId/signals libs (+14 tests), POST /api/signal (5/hr per ip_hash, auto-INs the lighter), GET/POST /api/signal/[id](/answer) (410 after expiry, one answer per IP, 30s poll), /signal/new (price-aware picker, live-HH pubs first, Perth-time chips), /signal/[id] (the dark beacon card + draining-glass timer + crew list, noindex, force-dynamic, graceful burned-out/missing states), robots disallow /signal/, SubPageNav gains an optional Beta badge prop.
• Verification: tsc clean, 321 tests pass, lint clean bar pre-existing warnings, Playwright screenshots of /signal/new (375+1280) and the no-signal state. Built by a subagent against the plan spec, reviewed + verified by the orchestrating session.

Cheapest chip + header breakpoint fixes (2026-06-10)

• Pub list "cheapest" treatment replaced — the #1 row's amber left-bar + star + tint (flagged as looking AI-generated) is gone. The rank number is back, and a solid amber CHEAPEST chip sits next to the pub name. The chip now marks the genuinely cheapest priced pub in view rather than whatever row sorts first (the old star wrongly starred row one under name/distance sort).
• Header no longer wraps at mid widths — at ~700-770px the homepage header crammed brand + 3 nav links + badge + button onto one row, wrapping the brand and "Happy Hours" onto two lines. Desktop nav/button cutover moved sm: → md: in HomeClient, SubPageNav, and MobileNav (hamburger now lives until 768px), brand wordmarks get whitespace-nowrap. Verified zero overflow and single-line header at 375/640/735/768/900/1024/1280.

Homepage live banner softened (2026-06-10)

• The live happy-hour banner was the last solid-black element on the homepage: bg-ink pill with amber-light text (and a bare shadow-hard, which the design system bans). Now white with ink text, the green pulse + LIVE eyebrow kept, amber reserved for the price, shadow-hard-sm. Verified with Playwright; tsc clean.

Pill no-wrap guard + happy-hour grid blowout fix (2026-06-10)

• Global CSS guard in globals.css: anything with rounded-pill gets white-space: nowrap — a pill label breaking onto a second line reads as broken UI (spotted on the world-cup "Tell us about an early open" button, label shortened to "Report an early open").
• While verifying at 320px, found a pre-existing 14px horizontal page overflow on /happy-hour: the best-of pick cards' truncate pub names couldn't clip because grid items refuse to shrink below content width. min-w-0 on the three card links fixes it. All of /, /world-cup, /happy-hour now measure zero horizontal overflow at 320px and 375px.

Amber-on-amber chip restyle, site-wide (2026-06-10)

• Killed every tone-on-tone amber combo (text-amber on bg-amber-pale) — flagged as looking cheap. New rules: active/live chips (happy-hour-active, Golden Hour, HH toggles) go solid bg-amber text-white; passive badges (admin pending/warning, promotion-zone divider) keep the pale tint but switch to ink text; the World Cup permit chip becomes amber-on-white with an amber border.
• Touched: WorldCupFixtures, SunsetSippers, TonightsMoves, DadBar, SubmitPubForm, SuburbLeague, admin dashboard, DiscoverClient, SuburbsClient (hover state). tsc clean; verified on /world-cup at 375x812.

Pint receipt readability pass (2026-06-10)

• Prompted by a user screenshot of The Vale Bar & Brasserie on mobile. Five fixes to PintReceipt.tsx:
• Happy hour row no longer reads "TBC · 7 days 4pm - 5pm" — the schedule is the value when the price is unknown, "7 days" renders as "daily", and a known price shows red with the schedule on a right-aligned sub-line (no more crushed dotted leaders).
• Stale prices flagged — the Checked row renders amber with "· Nd ago" appended when the recency tier is aging/stale (tier now passed into the receipt data).
• Provenance jargon humanized — "from an aggregator lead" → "spotted online" (priceProvenance.ts), moved to its own SOURCE row.
• Duplicate price removed — the standard-pint row only renders when it differs from the hero (it repeated the big number on every non-happy-hour pub).
• Type floor raised 8px → ~10px (banner sub-line, price label, amenity chips, CTA); the vs-avg line now includes the reference ("$1.10 below the $9.10 avg"); CTA flex-wraps.
• Self-branding line removed — the "★ Perth Pint Prices ★" banner row is gone; the pub name now leads the card.
• Verification: tsc clean, 309 tests pass, Playwright screenshots at 375x812 + 1280x800 on both happy-hour variants (The Vale = schedule-only, Ezra Pound = priced). Commit 5edfde2.

Slack notifications for the review queue (2026-06-10)

• Why: a pending price report (Gage Roads, Single Fin $14.50) sat unnoticed for 4 days — nothing told the admin the queue had work in it.
• New src/lib/slackNotify.ts — posts to a Slack incoming webhook via the SLACK_WEBHOOK_URL env var. Best-effort by design: missing env var or a Slack outage logs and moves on, never breaks a submission or cron. Message formatters are pure functions with unit tests.
• Instant ping: /api/price-report POST notifies Slack the moment a report lands — price reports, happy hour reports, and stale-price flags each get their own wording, with pub name, suburb, beer, price, source, and a link to /admin.
• Daily reminder: the existing daily price-check cron (8am Perth) now also counts pending price_reports (with the oldest report's age) and pub_submissions, and sends a queue summary — silent when empty. Piggybacked on the existing cron because Vercel Hobby caps cron jobs at two and both slots are taken.
• Action needed: create a Slack incoming webhook and add SLACK_WEBHOOK_URL to Vercel env vars — until then the code no-ops with a console warning.
• Verification: tsc clean, 308 unit tests pass (+7 new in slackNotify.test.ts). Commit dce54d2.

/world-cup: softer confirmed-opens card + flag chips on fixture rows (2026-06-10)

• The "Confirmed early opens" card was a solid bg-ink block mid-page — restyled to bg-amber-pale with ink text and the standard white pill button (user found the dark block harsh).
• TEAM_COLOURS extended from Group D (4 teams) to all 48 qualified teams with real flag-stripe palettes; fixture rows now show a small stacked-stripe flag chip glued to each team name (no emoji), and the team-colour bands on Socceroos cards pick up the new entries automatically.
• Fixture row layout moved from ragged flex-wrap to a grid: time column, teams with flags, chips right-aligned on desktop and indented under the match on mobile. Trading chips crispened (border-2, stronger tints).
• Verified with Playwright at 375x812; tsc clean, 302 tests pass.

Homepage World Cup strip: mobile polish (2026-06-10)

• The strip's header crammed both labels side-by-side at 375px (each wrapped to two lines); now they stack on mobile, side-by-side from sm: up.
• The fixture rail scrolls full-bleed (-mx-6/px-6 with scroll-pl-6) with snap points and a hidden scrollbar, matching the /discover picks rail — the next card peeks at the screen edge instead of being chopped at the container boundary.
• Verified with Playwright at 375x812 (rest + scrolled states); tsc clean.

Brand assets: stale arvo/PintDex images replaced (2026-06-10)

• The OG share card still said "arvo" with dead stats (420+ venues / 154 suburbs); logo.png, favicon.ico and the PWA/iOS icons were older still ("PintDex") — so every WhatsApp/social preview, browser tab and the Organization JSON-LD logo showed retired branding.
• New scripts/generate-brand-assets.mjs (Playwright render of the real fonts/tokens) regenerates og-image.png (evergreen copy, no counts to go stale), logo.png, icon-512/192, apple-touch-icon; scripts/build-favicon.py assembles favicon.ico. Manifest colours updated to current tokens. Commit eb3abc8. PR #178.
• Note: WhatsApp/Facebook cache link previews — after deploy, re-scrape via developers.facebook.com/tools/debug or just wait out the cache.

Visual audit fixes: SSR guide data, AA contrast, jump nav, empty states (2026-06-10)

• Guide/insight pages + /discover server-fetch pubs — all 8 FeaturePageShell consumers and /discover were client-only (multi-second spinner, blank sections when the browser-to-Supabase fetch failed). Each page.tsx now fetches server-side (revalidate = 300) and passes initialPubs down; new slimPubForFeature in pubPhoto.ts keeps the isDadBar amenity booleans while still dropping the heavy Google enrichment (~850KB HTML per page, under the 2MB crawl limit).
• Empty states — FeaturePageShell and /discover render a "pub list didn't load / try again" card instead of section headers over empty lists.
• Beer Weather page de-thinned — visible H1 + intro via the shell's new intro prop (was a bare widget with an sr-only heading). Duplicate sr-only <h1> removed from all 8 shell pages (every one had two H1s: the crawl block's + the shell's).
• WCAG AA contrast — gray-mid darkened #8A8A85 → #6E6E69 (~3.3:1 → ~4.9:1 on the cream background; body text now passes AA). CLAUDE.md design tokens updated.
• /happy-hour area jump nav — pill anchors (with counts) under the standfirst jump to each region section; the listicle runs ~62k px tall.
• PriceHistory skeleton removed — most pubs have <2 history points, so it flashed a grey block then collapsed (layout shift on the majority of pub pages).
• Homepage desktop nav links Articles — was reachable from every page except the homepage (MobileNav already had it).
• Verification — tsc clean, 302 unit tests (+1 slimPubForFeature), lint clean bar pre-existing SunsetSippers warnings, Playwright screenshots at 1280x800 + 375x812 for home, /discover, beer-weather, /happy-hour, pub detail. Commit af4b69c. PR #178.

Homepage: article rail moved below the pub list (2026-06-10)

• The blog post rail (ArticleRail, "Pub notes with numbers attached") sat between the live happy hour banner and the filter bar, pushing the price list down the page. It now renders after PubCardList (below the "View all venues" button, above How It Works), so prices are the first thing under the hero.
• Verified with Playwright screenshots at 1280x800 and 375x812; tsc clean. Commit a98cc30.

In-repo AI article drafting pipeline + dev-only /draft preview (2026-06-09)

• New scripts/draft-article.mjs — a BYOK content pipeline: pulls real pub data from Supabase, builds a data-grounded brief (optional --keywords for Ahrefs-seeded terms), drafts an article with Claude (claude-opus-4-8, structured-JSON output, cached style exemplars), runs a humanizer pass to strip AI tells, fetches topical Pexels photos (downloaded to public/_drafts/<slug>/, attribution captured), and writes drafts/<slug>.draft.{ts,md,json}. The .ts is a paste-ready Article object for src/lib/articles.ts; the .json feeds the preview route. Live-validated end-to-end (real Claude draft + Pexels) on a Fremantle cheap-pints article — 779 words, grounded in the 12 verified pubs, house voice intact.
• Dev-only preview — extracted the article body into a shared src/components/ArticleView.tsx (live /articles/[slug] rewired to it; SEO/JSON-LD/canonical retained; helpers + body byte-identical, confirmed by an Opus reviewer → no regression). New src/app/draft/[slug] + src/app/draft routes render a draft in the real site styling, 404 in production, noindex, with a "DRAFT PREVIEW · not published" banner. drafts/ + public/_drafts/ gitignored; drafts excluded from tsc.
• Why build vs buy — the deep-research pass on AI content tools concluded an in-repo pipeline beats a paid tool (Cuppa/Koala/Byword) for this code-based stack: BYOK, grounded in proprietary Supabase data, ending in the mandated human-edit/humanize step. Keys (Anthropic/Pexels) live in self-hosted Infisical (infisical.tail54186b.ts.net), not the repo; .env.local holds empty placeholders.
• Verification — tsc clean, 278 unit tests (+4 in pubPhoto/helper coverage), lint clean, Opus code review LGTM (2 hardening fixes applied: query-string guard on the photo resizer, shape-guard on the draft loader). Caveat: preview the /draft route from the main repo (full node_modules) — git worktrees lack deps, so a worktree dev server can't resolve React/Next (reading 'call').
• Voice brief: "polished tells" guard (2026-06-10) — docs/brand-voice-brief.md now guards against precious-quiet copy, not just corny-loud: new "should not feel precious or self-admiring" bullet, a "Polished tells" list (aphorisms, grand openers, virtue-headings, stacked "X, not Y" pivots), a drifted example from the World Cup draft, and voice-test #7 widened to catch both tells. Driven by founder feedback on the first pipeline draft ("too nice").

/world-cup live hub page (2026-06-10)

• New /world-cup route — a FANZO-style fixture hub with the trust layer they fake: all 72 group-stage fixtures in AWST (captured from FANZO's geo-localised rail, cross-checked against two AEST schedule sources; Socceroos times triple-verified against the official announcement), each tagged with its WA trading window — permit hours (midnight–6am any day, plus Sunday before 10am), early doors (6–9am, legal on a standard licence), normal trading. Knockouts deliberately omitted until the bracket settles — nothing TBC ships.
• src/lib/worldCup.ts — fixtures data + tradingStatus() classifier built on perthClock (standard hours 6am–midnight Mon–Sat / 10am Sunday per DLGSC, checked 10 June 2026; World Cup extended-trading permit announced 5 June 2026), formatKickoff (midnight/midday/3am/8.30am style), CONFIRMED_OPENINGS ledger (renders an honest empty state until venues confirm door times — no Andrew for now, so confirmations are manual + report-form), FORM_GUIDE_SLUGS.
• src/components/WorldCupFixtures.tsx — client fixture list grouped by Perth day: live clock (60s tick), ON NOW / NEXT chips, played matches auto-collapse, All / Socceroos & Group D filter, Socceroos rows highlighted.
• Page (src/app/world-cup/page.tsx) — hero stat cards, Socceroos cards, dark confirmed-opens card with report CTA, screens form guide pulling live pint prices + checked dates from our own pubs data (the moat: FANZO lists venues, we list venues with the price and the date we checked it), Northbridge Piazza free-screen note, FAQPage JSON-LD + visible Q&A, breadcrumb schema, sitemap + footer link.
• Verification — tsc clean, lint clean (one pre-existing SunsetSippers warning), 294 unit tests pass (+18 new for fixtures data sanity, the licensing classifier incl. Sunday edge cases, kickoff formatting, match phases), Playwright screenshots at 1280x800 + 375x812 (mobile chip-wrap fixed).
• Live countdowns (same day, direct to main) — new WorldCupCountdown client component (1s tick, mount-gated to dodge hydration mismatch, shows "On now" during the ~2h match window then disappears) on the three Socceroos cards ("Kicks off in 9d 16h 42m") and inside the fixture list's NEXT chip ("Next · 1d 16h 42m"). formatCountdown() in worldCup.ts drops ticking seconds beyond a day; +5 unit tests. GSC indexing requested for /world-cup the same day (priority crawl queue confirmed).
• Homepage World Cup strip + team colours (same day, direct to main) — HomeWorldCup renders the three Socceroos fixture cards (kickoff in Perth time, live countdown, flag-stripe band) between the homepage header and hero; horizontal scroll on mobile, 3-up grid on desktop; self-removes after the final (gated on perthToday() > '2026-07-19'). TeamStripes paints each card with both teams' flag colours via TEAM_COLOURS in worldCup.ts (Group D only, neutral fallback for knockout opponents); same stripes applied to the /world-cup Socceroos cards.

Fix Ahrefs size regressions: slim list payloads + resize venue photos (2026-06-09)

• Diagnosis from the 2026-06-08 Ahrefs crawl: the photo-recovery work (#173) introduced every new Error this week. Verified live with curl: the homepage shipped 2.22MB of HTML — all 857 pubs' full objects serialised into the RSC payload — tripping Googlebot's 2MB crawl limit; pub photos were hotlinked at up to 4800px (~208KB each), flagging "Image file size too large" on 15 pages. The real bulk wasn't the photos (~240KB) but the full Pub enrichment the lists never render: periods (441KB), editorial summaries (352KB), 15 Google attributes (282KB), opening-hours text (217KB).
• Fix — src/lib/pubPhoto.ts (new): resizeGooglePhoto() swaps the lh3 size token (=s4800-w1024 → =w640, ~208KB → ~90KB measured, under the 100KB threshold) while keeping the photo hotlinked to Google — their terms forbid re-hosting, so next/image is deliberately out (honours the existing "plain img on purpose" note at the render sites). slimPubForList() drops the list-irrelevant Places enrichment before serialising; every dropped field is optional on Pub, so SlimPub stays assignable to Pub — zero component/prop-type changes.
• Applied: homepage + suburb list payloads slimmed (suburb getSuburbStory + JSON-LD keep the full data, slimming only the client payload); photos resized at both render sites (pub detail + happy-hour listicle). /happy-hour HTML stays ~835KB on purpose — it genuinely renders photos + write-ups + hours — but its image transfer drops ~57%.
• Result: homepage 2.22MB → ~0.77MB (est.), well under the 2MB limit; each venue photo 208KB → 90KB. Clears Ahrefs "Page size exceeds 2MB" + "Image file size too large" + "HTML file size too large". The other ~38 audit issues (meta/title hygiene, orphans, the intentional 622 noindex tier pages) were left out of scope.
• Verification: tsc clean, 278 unit tests (+4 new in pubPhoto.test.ts), lint clean. (OneDrive-path space still breaks npm lifecycle scripts → ran tsc/tsx/next directly via node.) Full next build + e2e run in CI. Commit 5e61cfe.

Structured-data audit + Organization entity (2026-06-08)

• Full schema/breadcrumb audit of all 34 JSON-LD surfaces. Verdict: already in good shape — every breadcrumb is a proper BreadcrumbList with absolute item URLs and a sensible hierarchy (guides/insights crumb under "Discover", not the URL path); the pub page emits a correctly interlinked @graph (WebPage + BarOrPub + BreadcrumbList) with Offers/opening-hours/amenities/geo; article schema carries author + publisher + dates. Confirmed live: homepage + pub-page @graph parse and the pub breadcrumb renders Home → Suburb → Pub with absolute URLs.
• Did NOT "fix" the ItemList url props — checked Google's carousel spec: summary-page ListItems correctly use url (only BreadcrumbList uses item). The existing code was right; leaving it avoided a churning regression.
• Real gap closed — Organization brand entity: the homepage had a bare WebSite node and no Organization (no logo, no sameAs), so nothing tied the domain to the brand for Google's knowledge graph. New src/lib/siteJsonLd.ts (buildOrganizationJsonLd / buildWebSiteJsonLd, both @context-less for @graph embedding, with siteJsonLd.test.ts). Homepage now emits one @graph: Organization (@id #organization, square /logo.png 1024² as ImageObject, sameAs → the footer's facebook/instagram arvopints profiles) + WebSite (@id #website, publisher → Organization, inLanguage en-AU) + the existing FAQPage (@id #faq). articleJsonLd publisher logo switched from the landscape OG image to the square /logo.png and given the matching #organization @id.
• Validated against Google's official docs: carousel ItemList summary pages require url on each ListItem (only BreadcrumbList uses item) — confirmed, so the existing ItemLists were left alone; BreadcrumbList needs position + name + absolute item (last item's item optional) — all pages conform; Organization logo accepts an ImageObject/square image ≥112px — our 1024² /logo.png qualifies.
• Verification: tsc clean, 274 tests (+2 site-schema cases), live homepage @graph validated (Organization/WebSite/FAQPage all parse, logo + sameAs + publisher link present), live pub breadcrumb validated.

GSC indexing fixes: section-redirect bug + retired-URL hygiene (2026-06-08)

• Diagnosis from the live GSC Page-indexing report: 421 "not indexed" vs 1.15k indexed (indexed trend still climbing). The headline is mostly benign — 273 of the 421 are Page with redirect (the intended legacy /pub/, /suburb/, http→https, etc. — working as designed). The genuinely actionable buckets: 67 Discovered + 66 Crawled – currently not indexed, which drill-down confirms are almost entirely individual /[suburb]/[pub] pages — the long tail of price-less / near-template venues Google judges too thin. That's a data problem (663 pubs still missing a regular price), already handled architecturally (Tier-C pages are noindex + sitemap-excluded), so no code change ships for it — it needs prices, not markup. Deferred + flagged to the user.
• Real bug found & fixed — /guides + /insights "Redirect error": both were page-component permanentRedirect('/discover'). On Vercel's CDN that serialised to a 308 with NO Location header + a 9 KB HTML body (confirmed with curl against prod) — a redirect to nowhere, which is exactly what Googlebot reports as a redirect error. Moved both to vercel.json edge redirects (statusCode 301, exact-source so /guides/* + /insights/* children are untouched), matching the existing /suburb/:slug + www pattern, and deleted the two redirect-only page.tsx files (the layout.tsx files stay for the child pages). Edge redirects emit a clean 301 + Location — proven by the analogous /suburb/north-perth → /north-perth already doing so on prod.
• Retired-URL hygiene: legacy /pub/{slug} misses (removed/re-slugged venues: badlands-bar, jack-rabbit-slims, w-churchill, wolf-lane, ruin-bar, the-flying-scotsman) now return 410 Gone instead of 404 — the precise status for a permanently-retired URL scheme. (Per Google's HTTP-status docs, 404 and 410 are treated identically for indexing, so this is a correctness choice, not a crawl-speed win — the earlier "drops faster" rationale was wrong.) The 301 redirect still fires when the slug resolves. The current-structure 404s (/northbridge/ruin-bar, /mount-lawley/five-bar, /perth-cbd/halford-bar) are genuinely-removed pubs — 404 is correct, they'll drop naturally. The 4 Duplicate canonical (incl. /suburb/north-perth, which 301s cleanly) and the /guides+/insights redirect-error validations are stale classifications that clear on recrawl.
• redirect-seo.test.mjs extended to assert the new /guides + /insights edge redirects and the /pub/{slug} 410 (tested contract, not just config).
• Verification: tsc clean, 272 tests + redirect-SEO test pass, vercel.json valid; the 308 → /discover (with Location) + child-page-200 + 410 behaviour confirmed live on a local dev server before moving to the edge-redirect form. (npm lifecycle scripts still break on the OneDrive-path space → ran tsc/tsx/next directly via node.)

Recover hidden pub photos + manual-photo fallback (2026-06-08)

• Re-curated all 191 previously-hidden pubs. The earlier long-tail pass hid ~191 pubs because the auto-pick landed on food/crowds — but most had a real venue shot deeper in the (stable-order) photos array. Re-pulled up to 10 candidates each, then ran a 4-stage QC: 15 picker agents → 12 adversarial verifier agents (flagged 54) → 6 re-pick agents → every one of the 159 final photos eyeballed by me via @napi-rs/canvas montage sheets, hand-picking the ~20 hard ones from their full folders. Net: 159 pubs recovered a genuine venue shot (exterior/bar/interior/beer-garden/clubhouse); 32 stay blank (beer brands with only cans, golf courses with only fairways, venues with only food/people). Overrides still 248; applied index-based and verified what's actually stored matches every pick (no silent-fail).
• Lesson reinforced: the vision agents hallucinate at the index level — they confidently mis-described specific files (a rainbow-over-fairway called a "clubhouse interior", a chicken box called a "storefront"). So agent picks are a first pass only; the human montage review is mandatory. New throwaway QC tooling: scripts/recurated-picks.json (decision record), apply-recurated-picks.mjs, build-verify-montage.mjs / build-folder-montage.mjs / build-stored-montage.mjs.
• Manual-photo fallback (image_url): pub detail page, happy-hour listicle, and BarOrPub JSON-LD now fall back to a manually-supplied image_url (credited "courtesy of {pub}") when Google has no photo — so the ~85 pubs Google has zero photos for (e.g. The Deen) can get a venue-supplied shot dropped straight in. image_url was empty across all pubs, so zero risk of surfacing stale data. (Lifting photos from venue sites / Eventfinda / socials stays off the table — attribution ≠ licence.)

Photo curation fix: index-based overrides + strict listicle sweep (2026-06-06)

• Critical fix — curated photos never applied: Google's Place Photo refs rotate every request (verified: two back-to-back fetches return different ref tokens), so the ref-based pickBestPhoto override always missed and silently fell back to the heuristic. The 51 curated listicle re-picks (incl. The Bird) never landed — prod was still serving the original food/people shots (The Bird = a head-in-the-foreground courtyard). The photos array order is stable, so overrides now pin by index (scripts/photo-overrides.json stores integer indices, not refs). Re-applied + verified what's actually stored now: The Bird = the back bar, NBC = the brick exterior, Palace Arcade = the arcade-bar interior.
• Strict re-sweep of the listicle: agents re-reviewed all 132 happy-hour photos at a tougher bar (any prominent person/head/food/drink in the foreground = reject). Re-picked the 10 non-curated stragglers the lenient first pass missed — 7 real venue shots + 3 hidden (Quarie / Sporting Globe Belmont / Oceans 6019 had only food/crowds). Overrides now 248 (index picks + NONE hides).
• Lesson: the pick agents mis-call food-with-branding as a venue shot ~20–30% of the time, so every pick is now human-verified before apply (this caught 3/9 on the last batch). Added scripts/repick-photos.mjs (--slugs per-pub re-pick) + scripts/verify-applied.mjs (download what's actually stored) for ongoing fixes; backfill-place-attributes.mjs gained --slugs for targeted single-pub applies.

#164: happy-hour area listicle + attributed venue photos + opening hours (2026-06-06)

• PR #164 (054340e): /happy-hour rebuilt as an editorial area-grouped listicle (CBD/Northbridge, North, Coast, Freo/South, Hills — cheapest first), with a live "on right now" section (countdowns), best-of picks, per-venue write-ups, and a collapsible opening-hours + address line. Every pub page gains an attributed Google Places hero photo, structured opening hours, and a visible street address; BarOrPub JSON-LD now carries the photo as an ImageObject (creditText + author) for richer local results. New src/lib/perthRegions.ts (suburb→region map).
• Google Places photo pipeline (scripts/backfill-place-attributes.mjs + migration 20260605000000_pub_google_photos.sql): one attributed photo per pub at $0 (rides the existing Details call), stored in separate columns (google_photo_url/_attribution/_attribution_uri) so curation never clobbers image_url. Smart-picks the best candidate (business + landscape over food/portrait). Photos are hotlinked + monthly-refreshed (not re-hosted) and always shown with the contributor + Google Maps attribution, per Places API policy. ~772 pubs backfilled.
• Photo quality — vision curation: Google's pub photos run ~30–38% food/drink close-ups, which no metadata heuristic can catch. Agents viewed every photo: the 51 weak listicle ones were re-picked from candidates (49 real venue shots + 2 hidden), and 186 weak long-tail ones hidden (hide-bad-photos.mjs) — all recorded as slug→ref/"NONE" in scripts/photo-overrides.json (237 entries) so the monthly refresh keeps the hand-vetted choices. Net: every photo on prod is a real venue shot or none — no burgers/selfies/menus.
• Verification: tsc clean, Playwright CI green, prod live (perthpintprices.com/happy-hour rendering photos), SSR + DOM confirmed on listicle + pub pages.
• Flagged follow-up: Supabase egress is over the 5 GB free tier — the app's getPubs pulls the full ~2.4 MB pubs table uncached on every render. Structural fix: trim select('*') / lean on getPubsLite / cache the list.

SEO #4: money-page keyword pass + /happy-hour rebuild (2026-06-05)

• Keyword research (Ahrefs, AU) → docs/seo/keywords.md. The on-brand terms are wide open: happy hour perth (250/mo, KD 0), best pubs perth (400, KD 0), beer garden perth (100, KD 0), dog friendly pubs perth (150, KD 0); cheap pints perth triggers an AI Overview (citation prize). Flagged rooftop bars perth (1,500, KD 30) + bars perth (800) as high-volume but off the pint-price USP. Filled the Ahrefs project id (9843078) + keyword-store blanks in .claude/seo-content.config.md.
• /happy-hour optimised for "happy hour perth" (the #1 opportunity). The SERP rewards a comprehensive list (Perth is OK / sitchu listicles rank), but our page filtered the visible list to isHappyHourNow — so outside ~4–6pm it showed "No happy hours right now" + an empty list, unrankable for the query. Now it lists every timed happy hour, cheapest first, with active ones flagged live: keyword H1 ("Happy hours in Perth"), an answer-first intro, a sharper title ("Happy Hour Perth — Every Deal, Live & Priced"), and a FAQ + FAQPage schema (what time / where cheapest / which pubs). Keeps the live board + countdowns — the freshness the static listicles can't match.
• Verification: tsc clean, 272 tests, rendered HTML confirmed (full list, keyword H1, FAQPage, no empty state).
• Next: the KD-0 guide gaps — /guides/dog-friendly-pubs, /guides/beer-gardens (the allows_dogs / outdoor_seating data already exists).

Pub page SEO: price schema + answer-first + happy-hour titles (2026-06-05)

• feat/pub-page-seo: an SEO audit (via the seo-content skill) of the post-receipt pub template found the pages under-optimised for a wide-open SERP — searching "how much is a pint at [pub]" returns TripAdvisor / booking sites, none of which answer the price. Three fixes, hitting every priced/HH pub at once:
  • Exact price in structured data (pubJsonLd.ts): added makesOffer → Offer/MenuItem with the real pint price (+ happy-hour price) in AUD — previously the price was only a coarse priceRange: "$$". Now machine-readable for rich results + AI answers.
  • Answer-first line + FAQPage (PubDetailClient.tsx): a plain, un-boxed one-line answer under the H1 — "A pint at {pub} is ${price}, last checked {date}." — restoring the extractable answer the removed Quick-read box used to carry; plus FAQPage JSON-LD emitted from the price/HH/nearby Q&A (visible-FAQ gate lowered 3→2 so it renders + matches the schema). FAQ now uses the real suburb average (was the city average, mislabelled).
  • Happy-hour titles (page.tsx + voiceCopy.ts): HH-only pubs (no standard price) now title "$6.00 happy-hour pints" / meta "pours a $6 pint during happy hour…" instead of the flaky "Price TBC" (old title keyed off the time-dependent effective price, so it flipped in/out of HH).
• Already good, left alone: SSR/crawlable render, tiered indexability (dataless pubs noindex), self-canonicals, breadcrumb + LocalBusiness schema, internal linking; aggregateRating correctly omitted (can't claim Google's stars as first-party).
• Deferred (#4, next): thin per-page content depth — the bigger lever is the suburb/discover/insight "money" pages feeding link equity down; an Ahrefs keyword pass goes there. Added .claude/seo-content.config.md.
• Verification: tsc clean, 272 tests, signals confirmed in rendered HTML (Offers, answer line, FAQPage, HH titles) on priced + HH-only pubs.

Pub page receipt card — site-wide template (2026-06-05)

• feat/pub-receipt-card: the pint price card on /[suburb]/[pub] is now the "receipt" for every pub — an amber-headed docket (★ Perth Pint Prices ★ banner + venue name in DM Serif), the hero price, an itemised dotted-leader sheet (Standard pint · Happy hour · Cheaper nearby · Checked · Google rating), amenity stamps, and an in-card report CTA. New self-contained src/components/PintReceipt.tsx; the throwaway _receipt-prototype/ route (A thermal docket / B raffle ticket / C banner card, toggled via a dev-gated ?variant switcher) was deleted after C won.
• Unpriced pubs (most of the 857) handled: TBC renders cleanly — no false "vs-avg" line, the CTA reads "Know the price?" not "a better price", and the Tier-C "nearest checked price" stub was lifted out of the old card into its own bordered block (verified on The Flaming Galah, Freo → "12 nearby pubs have a checked price — start with Whisper Wine Bar at $9.00").
• Page de-dup (the receipt owns this data now): removed the dark "Quick read" answer-first box (its numbers live in the receipt + the FAQPage schema), the duplicate full-width report button, and the repeated rating/amenity chips below the card (GoodToKnow gained a summaryOnly prop → editorial blurb only). About + that blurb now share one bordered card instead of floating. Dropped the redundant "Family-friendly in {suburb}" subtitle (the vibe pill already says it).
• Verification: tsc clean, 272 tests, priced + unpriced pubs confirmed in-browser via DOM reads (Chrome screenshot capture wedges on the dev sticky Leaflet map).

The Deen suburb correction (2026-06-05)

• Data fix (scripts/fix-deen-suburb.mjs): The Deen (pub #7) was tagged suburb = "North Perth" but sits at 82 Aberdeen St, Northbridge — corrected to Northbridge. Surfaced by the content-review address/suburb audit (the venue is its own Pint-of-the-Day pick, so the wrong tag was visible, and it skewed the Northbridge cluster stats). The pub route looks pubs up by slug and 301-redirects on a suburb mismatch, so /north-perth/the-deen now 308-redirects to /northbridge/the-deen (verified) — no code redirect needed. Script is idempotent + guarded (only touches the row while it is still "North Perth" with a Northbridge address). The wider ~127 address/suburb mismatches stay deferred — most are false positives (CBD naming, border streets).

Pint-price stat consolidation + content cleanup (2026-06-05)

• PR #157: one source of truth for every headline number, driven by a full site content review. New src/lib/suburbStats.ts (getSuburbStats/getSuburbExtremes/getPricedSuburbCount, on verified regularPrice) and a cached /api/pint-index route; HomeClient, VenueIntel, SuburbLeague, getSiteStats, the Pint Index widget + header chip all route through getPintPriceStats + getSuburbStats. Result: avg $9.10 · median $9.00 · cheapest Hillarys · priciest Kalamunda on every surface (was $9.04/$9.09/$9.10 and three different "cheapest suburb" answers). Snapshots now feed only the trend line + % change; the weekly-snapshot cron recomputes canonically and /api/weekly-snapshot is an alias.
• Content fixes: Dad Bar 11 → ~67 venues via a shared isDadBar() (curated kidFriendly + Google goodForChildren with a beer garden/sports, Perth CBD excluded; curated playground venues pinned on top; Discover hub card now matches). Suburb Rankings capped to 18 (one per AFL club) with verified-priced counts (kills the "Perth (84) $8/$8/$8" anomaly). Cosy Corners renders its list instead of a blank page (RainyDay weather is now an enhancement, not a gate). Venue Breakdown "% cheaper than median" computed over priced venues (was the wrong 8%, dividing by all 857). Happy-hours article board limited to genuine confirmed pint drops (excludes the small-pour/unconfirmed entries the prose disclaims); formatHappyHourDays now collapses full/short day names to "Mon–Fri". "Article brief" block removed; latent Rules-of-Hooks crash in the Pint Index sparkline fixed.
• Tailwind: off-design-system utilities → design tokens site-wide (CrowdReporter, MobileNav, PintIndex, FilterSection, SubmitPubForm, crowd-level colours) — zero off-system utilities remain outside vendored shadcn.
• Held for a separate data pass: ~127 address/suburb mismatches surfaced (e.g. The Deen tagged North Perth but addressed in Northbridge) — kept out of this PR because correcting a suburb changes the pub's public URL (needs a redirect), and most flagged rows are false positives (CBD naming, border streets).
• Verification: tsc clean, 272 tests (incl. new suburbStats cases), production next build clean, cross-page numbers confirmed live in-browser. (npm lifecycle scripts still break on the OneDrive path → ran build/tests/tsc directly via node.)

Pub page Variant B "receipt" redesign (2026-06-04)

• PR #154 / merge commit d666038: folded the chosen prototype direction into /[suburb]/[pub] — the pint price card is now a betting-slip "ticket" (dark PINT PRICE header bar, centred hero price, vs-avg / freshness / HH-now badges, dashed-divider happy-hour receipt row). Built on the new .type-* system; the page also adopts type-hero/type-eyebrow/type-card-header/type-card/type-price.
• All real machinery preserved: watchlist/share, the answer-first Quick read card, live HH-now countdown, price provenance line, the Tier-C "no verified price" verification stub (now TBC + stub inside the receipt), price history, the permanently-closed banner, the Leaflet map, nearby-pubs with analytics, and the report-price flow. Chosen via a throwaway /prototype route (3 toggleable directions, since deleted).
• Verification: live via Chrome MCP (rich pub, no console errors) + Playwright (mobile reflow + Tier-C state); tsc clean, 267 tests, lint clean. CI caught the e2e PR-proof asserting the old "Pint Price" label → updated to the new "Pint price" copy.

Semantic typography system + site-wide codemod (2026-06-04)

• PR #153 / merge commit bef5b08: added a .type-* component layer in globals.css (one definition per role: type-hero, type-hero-editorial, type-section, type-card, type-card-header, type-eyebrow, type-price) and migrated ~30 components/pages onto it — fixing the typography-audit root causes.
• Root-cause fixes: the base h2,h3 → DM Serif + bold rule (faux-bold + accidental serif; DM Serif Display ships weight 400 only) replaced with h1–h4 → mono, deliberate serif now opt-in via font-display; editorial heroes unified to one size (was 4.55/4.7/4.8/5rem drift); removed the misleading font-heading alias; fixed undefined text-coral/bg-coral → red and SVG monospace → the brand mono var.
• Peer-reviewed by 3 parallel agents (regressions / spec-adherence / coverage). Findings folded in — notably the type-card-header role (uppercase panel headers were being shrunk into type-eyebrow) and two card titles that had shrunk. Verified tsc, 267 tests, lint, and desktop/mobile QA across 6 routes.

Closed-venue noindex + place_id data-quality pass (2026-06-04)

• PR #152 / merge commit e2b2594: business_status = CLOSED_PERMANENTLY (from the Places sweep) now forces a venue to isIndexable: false (robots noindex + sitemap exclusion) while keeping its data-driven tier for rendering, plus a "permanently closed" banner. Auto-handles future closures on each monthly sweep.
• Address audit + fixes (data-only): ran a cheap (Pro-tier) Places address audit across 815 pubs; auto-corrected 23 genuine address errors (Running With Thieves + 22 others) to the matched place_id's Google address, skipping unit/formatting noise; held 3 ambiguous ones for manual review (they turned out to be correct place_ids on re-check).
• Phantom-venue cleanup (data-only): the 4 dead-place_id pubs (Bar 2wo, Martinis Oysters & Jazz, Long Point Brewing, Parsons Bar & Grill) — all price-less, never-verified, no current Google match — were delisted (CLOSED_PERMANENTLY, place_id nulled). Re-match search confirmed the 3 "held" place_ids (Steve's, Lyric's Bar, Samuels on Mill) were already correct.

Monthly Places refresh automation (2026-06-04)

• .github/workflows/places-refresh.yml (in PR #151): a GitHub Actions cron (0 17 1 * *) re-runs scripts/backfill-place-attributes.mjs monthly to keep amenities/hours/ratings and the visible "checked" date current — $0 under the 1,000-call/month Enterprise+Atmosphere free quota (~815 pubs/sweep). workflow_dispatch supports a limit input for cheap smoke tests; secrets GOOGLE_PLACES_API_KEY / SUPABASE_SERVICE_ROLE_KEY / NEXT_PUBLIC_SUPABASE_URL set on the repo. Smoke-tested green.

Google Places attribute backfill + Good-to-know pub module (2026-06-04)

• Related to issue #78 / branch feat/places-attribute-backfill: built the Google Places (New) attribute spine over the place_ids already stored on pubs. A single Place Details call per pub is already billed at the Enterprise+Atmosphere SKU because of the amenity booleans, so the sweep pulls everything useful in that one call at zero marginal cost — and ~857 pubs fit inside the 1,000-call/month free quota, so one sweep is $0.
• Migration 20260604000000_google_places_attributes.sql: added typed columns to pubs — amenity booleans (serves_beer, serves_food, outdoor_seating, good_for_children, good_for_groups, good_for_watching_sports, allows_dogs, live_music, restroom, reservable), plus google_rating, google_rating_count, google_price_level, business_status, google_editorial_summary, google_opening_hours (jsonb), and a google_attrs_updated_at provenance stamp. Stored separate from the hand-curated kid_friendly/cozy_pub/has_tab/sunset_spot columns so curation always wins; the sweep only fills a blank website as a freebie.
• Backfill script scripts/backfill-place-attributes.mjs: mirrors backfill-phones.mjs — one FieldMask, --dry-run + --limit flags, 80ms throttle, prints a per-attribute coverage table, and guards the business_status CHECK constraint against unexpected enum values.
• Render — GoodToKnow component: a compact, sourced "Good to know" chip row on the pub page (Beer garden · Dog-friendly · Shows the footy · Kids welcome · Kitchen · Live music · Good for a group · Takes bookings), plus the Google rating and editorial summary, attributed inline with the checked date. Truthful-absence: renders nothing when Google has no signal, so it is safe to deploy ahead of the sweep.
• Schema: extended buildPubJsonLd with amenityFeature (only Google-affirmed true attributes) and a real openingHoursSpecification array built from Google's trading-hours periods. Google's star rating is shown visibly but deliberately kept out of aggregateRating to respect the locked no-self-serving-rating guardrail.
• Shipped to production: migration applied via the Supabase SQL editor (all 17 columns + 2 constraints live); full sweep wrote 811 of 815 pubs (4 skipped on stale 404 place_ids), filling 79 blank websites as a bonus. Catalogue-wide coverage: google_rating 96%, opening hours 89%, serves_beer 85%, good_for_groups 79%, reservable 68%, outdoor_seating 56%, live_music 39%, good_for_children 32%, serves_food 30%, editorial_summary 30%, good_for_watching_sports 28%, allows_dogs 21%.
• Verification: npx tsc --noEmit clean, npm test (266/266 tests, including 3 new pubJsonLd cases for amenity emission/omission + regular-hours mapping), npm run lint clean on all changed files, and desktop + mobile Playwright screenshots of a populated pub page (/burswood/the-camfield) confirming the sourced chip row, Google rating, editorial summary, and dated attribution render on-brand. npm run build not run locally (the OneDrive path breaks npm lifecycle scripts here); Vercel will clean-build on merge.

Answer-first pint-cost page shipped (2026-06-03)

• Issue #27 / PR #147 / merge commit c561fd2: added /how-much-is-a-pint-in-perth for the zero-click "how much is a pint in Perth" cluster. The page answers in the first screen with the live checked average, median, range, cheapest verified pint, glass-size notes, visible Q&A, and internal links into Cheapest Pints, Happy Hours, Pint Index, and the proper-pint explainer.
• Pint Index answer block: added a "Perth beer prices" answer-first card to /insights/pint-index, backed by the same verified regular-price stats. The mobile Pint Index methodology wrapper was tightened to remove an 8px horizontal overflow found during screenshot QA.
• Shared stats helper: added src/lib/pintPriceStats.ts with tests so average, median, range, under-$10 count, and verified-row filtering stay consistent across answer pages.
• SEO wiring: added canonical metadata, OG/Twitter image, Breadcrumb + ItemList JSON-LD, and sitemap/footer//llms.txt entries.
• Verification: verified npm test (263/263 tests), npx tsc --noEmit, git diff --check, npm run lint (existing warnings only), npm run build, npm run test:e2e (4/4 Playwright tests), reviewer sub-agent LGTM, and clean desktop/mobile screenshots for /how-much-is-a-pint-in-perth + /insights/pint-index under /tmp/perth-pint-prices-answer-first-27/. The repeated local build flake (/_document not found) only occurred while a dev/e2e server was touching .next; the solo build passed.

Student-pints landing page ready for preview (2026-06-03)

• Issue #32 third slice / branch codex/student-pints-landing-page-32 / commit 8ec3168: added /student-pints-perth, a verified under-$10 regular-pint page for UWA, Curtin, and Murdoch.
• Campus ranking: added shared student-pint ranking helpers and tests. Campus sections use verified regular prices only, direct-distance radii, and campus-specific notes; Murdoch deliberately uses a wider 8km radius because the current checked data has no useful sub-$10 cluster inside 4-6km.
• SEO wiring: added canonical metadata, OG/Twitter image, Breadcrumb + ItemList JSON-LD, useful internal links, and sitemap/footer//llms.txt entries.
• Verification: verified npm test (260/260 tests), npx tsc --noEmit, git diff --check, npm run lint (existing warnings only), npm run build, npm run test:e2e (4/4 Playwright tests), reviewer sub-agent LGTM, and desktop/mobile screenshots under /tmp/perth-pint-prices-student-pints-32/.

Transport-hub landing pages ready for preview (2026-06-03)

• Issue #32 second slice / branch codex/transport-hub-landing-pages-32 / commit f6c8adf: added four direct-proximity pub landing pages: /pubs-near-perth-station, /pubs-near-optus-stadium, /pubs-near-rac-arena, and /pubs-near-elizabeth-quay.
• Reusable hub system: added shared transport-hub config, direct-distance ranking, ItemList JSON-LD, canonical metadata with OG/Twitter images, answer-first copy, stats, ranked nearby rows, and cross-links between transport guides. Sitemap, footer, and /llms.txt now expose the hub pages.
• Data guardrails: unverified regular prices render as TBC and do not affect equal-distance price tie-breaks. Page copy uses direct-proximity/direct-distance wording rather than claiming real walking-route distances.
• Verification: verified npm test (257/257 tests), npx tsc --noEmit, git diff --check, npm run lint (existing warnings only), npm run build, npm run test:e2e (4/4 Playwright tests), reviewer sub-agent LGTM, and desktop/mobile screenshots for Perth Station + Optus Stadium under /tmp/perth-pint-prices-transport-hubs-32/.

High-intent landing pages ready for preview (2026-06-03)

• Issue #32 first slice / branch codex/high-intent-landing-pages-32 / commit 68754c5: added /cheapest-pints as a verified regular-price landing page and /happy-hour/[day] static pages for all seven weekdays, with canonical metadata, OG/Twitter tags, Breadcrumb + ItemList JSON-LD, answer-first copy, ranked rows, Q&A blocks, and useful internal links.
• Happy-hour day linking: added a planning-board rail to /happy-hour, linked Cheap Pints from the footer, added the new pages to sitemap.xml, and exposed /cheapest-pints in /llms.txt.
• Data guardrails: added a shared happy-hour day parser with tests for weekday ranges, wrapped ranges, en-dash ranges, daily labels, weekday/weekend labels, and Postgres array-ish strings. Day pages sort by happyHourPrice, show TBC when a deal has no confirmed price, and avoid mixing regular pint prices into happy-hour claims.
• Verification: verified npm test (254/254 tests), npx tsc --noEmit, git diff --check, npm run lint (existing warnings only), npm run build, npm run test:e2e (4/4 Playwright tests), reviewer sub-agent LGTM, and desktop/mobile in-app browser screenshots under /tmp/perth-pint-prices-high-intent-32/.

Editorial/page-depth measurement ready (2026-06-03)

• Issue #140 / branch codex/measure-editorial-impact-140 / commit c587238: added one Vercel+GA4 event helper and tracked link wrapper, then wired article hub/rail/detail clicks, article-to-pub clicks, report-price CTAs, pub-page nearby/internal clicks, and website/directions clicks so the page-depth/editorial push can be measured instead of guessed.
• Snapshot process: expanded docs/seo-snapshots/TEMPLATE.md with editorial URL baselines, article engagement metrics, pub-page engagement events, and thin-content watch notes for the first weekly baseline.
• Verification: verified npx tsc --noEmit, git diff --check, npm run lint (existing warnings only), npm test (249/249 tests), npm run test:e2e (4/4 Playwright tests), npm run build, reviewer sub-agent LGTM, and desktop/mobile screenshots for /articles, /articles/pints-under-10-perth, and /perth/18-knots-rooftop-bar under /tmp/perth-pint-prices-measure-140/.

Pub-page depth work in progress (2026-06-03)

• Issue #138 / branch codex/pub-page-depth-138 / commits 6c737d6, c690a0f: added fact-earned pub-page quick reads, cleaner price-card contrast, best-time/nearby summary copy, gated seeded mini FAQs for pubs with enough real answers, and de-duplicated metadata via the pub voice builders. Reviewer follow-up normalized happy-hour metadata labels, removed duplicate Tier-C stub copy, and updated PR proof for the new nearby heading states. Visual proof covers priced, Tier-C, and FAQ-rich pub states under /tmp/perth-pint-prices-pub-depth-138/.
• Verification: verified npx tsc --noEmit, git diff --check, npm run lint (existing warnings only), npm test (249/249 tests), npm run test:e2e (4/4 Playwright tests), npm run build, and desktop/mobile browser screenshots.

Article rails + page-depth hubs ready for preview (2026-06-03)

• Branch codex/hub-article-rails / commits ae5b705, 063f5e8, fff9bc5, 84b9402: added the article rail system to Home/Discover, fact-checked the three new pub articles, wired the generated article image set under public/articles/, and added reusable image slots for the article hub/rails.
• Inline article imagery (85c2d96): article detail pages now render the primary article image plus four section-matched supporting figures per article, with captions, alt text, and explicit Next Image dimensions.
• Page-depth modules: beefed out /happy-hour with a live happy-hour board, /suburbs with data-led suburb rails, and dynamic suburb pages with verified-price story cards, local FAQs, and FAQPage JSON-LD.
• Data-source corrections: suburb summaries now use verified regular pint prices for verifiedCount, cheapest, most expensive, and averages, so metadata, crawler text, nearby chips, and visible suburb copy do not mix live happy-hour prices into regular-price claims. Article dates are formatted in Australia/Perth to avoid SSR/client date drift.
• Verification: verified npx tsc --noEmit, npm run lint (existing warnings only), npm test (249/249 tests), npm run build, reviewer sub-agent LGTM, and desktop/mobile browser evidence under /tmp/perth-pint-prices-depth-final-aggregate/.

Page-depth audit + articles system shipped locally (2026-06-03)

• Branch codex/articles-system / commits a6ef748, 7443705: added the page-depth editorial audit (docs/page-depth-editorial-plan.md) and the first pub editorial article system. New routes are /articles plus three SSG article pages: /articles/pints-under-10-perth, /articles/perth-happy-hours-by-day, and /articles/proper-pint-schooner-middy-perth.
• Editorial plumbing: articles are driven from src/lib/articles.ts, wired into footer/mobile/sub-page nav, sitemap.xml, and /llms.txt. Article detail pages emit canonical metadata and BlogPosting JSON-LD, and can render live pub-data modules for verified sub-$10 pints, happy hours by day, and glass-size explainers.
• UI polish: removed the repeated OG-logo image treatment, rebuilt the article hub as a text-led editorial surface, cleaned the featured/latest card hierarchy, and switched detail pages to compact editorial briefs plus useful live-data panels.
• Verification: verified npx tsc --noEmit, npm run lint (existing warnings only), npm test (249/249 tests), npm run build, and desktop/mobile visual evidence for the article hub and detail pages under /tmp/perth-pint-prices-articles-system/.

Official menu extractor hardening + second dry run (2026-06-03)

• Branch codex/official-menu-extractor-hardening / commit f69c55e: tightened official-menu source and row filters after the larger dry run surfaced false positives. The extractor now skips Non-Alc shorthand, apple-cider-vinaigrette food rows, and beer-in-food-description rows such as macaroni/panko/deep-fried items; source discovery and seed building now block breakfast/kids menu URLs even when filenames use underscores.
• Second dry run: rebuilt seeds from the large discovery artifact with the hardened filters, producing 202 crawl sources. Ran a no-write crawl with linked-source follow-up and OCR disabled for native-tooling stability: 195 fetched, 131 linked fetched, 11 linked failed, 39 candidates, 7 failed sources, 0 inserted.
• Review pack: generated scripts/official-menu-review-next.json and .csv locally with suggested decisions: 9 approve_suggested, 30 needs_manual_review. No production import was run; rows still need manual review before any pending reports are inserted.
• Verification: verified focused official-menu tests, npm test (248/248 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build (existing warnings only), git diff --check, and seed/crawl/review smoke checks.

Official menu crawler scaled to review + pending reports (2026-06-02)

• Local uncommitted work: expanded the official-menu workflow from MVP into a dry-run-first discovery, seed, crawl, review, suggested-decision, and guarded import pipeline. The approval pack lives at docs/official-menu-crawler-approval.html; generated crawl/review/import artifacts are gitignored.
• Crawler improvements: added rendered HTML fallback, PDF text/table extraction, image OCR, scanned-PDF OCR fallback, JSON-LD MenuItem extraction, adjacent price-line handling, source/provider discovery, and stricter filtering for ecommerce/product pages, generic image assets, unsupported AVIF OCR, cocktails/wine/spirits, food combos, non-alcoholic beers, stale CMS templates, and other false positives found during live dry runs.
• Review workflow: added scripts/build-official-menu-seeds.mjs, scripts/build-official-menu-review.mjs, and scripts/import-official-menu-review.mjs. Review exports include source URL, evidence text, extraction mode, confidence, decision, and notes; suggested decisions separate explicit pint/draught/tap evidence from rows needing manual review.
• Admin review: the admin stats payload now returns pending price reports as the review queue instead of only the latest 10 reports, and the Reports tab surfaces official-menu source/evidence fields so the 26 imported rows can be actioned through the normal approval path.
• Follow-up provenance fix: linked-source crawl candidates now carry the actual linked menu/PDF/image URL and extraction mode into review exports instead of falling back to the parent page URL.
• Latest batch: a 140-source dry run fetched 129 sources, found 26 review candidates, and inserted 0 rows during crawl. After owner approval, the importer inserted 26 pending price_reports with submission_source = "official_menu". Admin review is complete: 5 clear official-menu rows were approved into canonical pub prices, and 21 ambiguous/duplicate rows were rejected.
• Verification: verified node --check for the new workflow scripts, a dry-run import plan, npm test, npx tsc --noEmit, npm run lint (existing warnings only), npm run build (existing warnings only), and desktop/mobile admin screenshots.
• Commits: d8e976d, 99b8200

Official menu source discovery ready (2026-06-01)

• PR #130 / merge commit 92f2c8d: added a read-only discovery CLI that scans missing-price pubs with websites, ranks likely menu/drinks/PDF links, and writes scripts/official-menu-source-candidates.json for review. The generated artifact is gitignored.
• Source scoring: added pure link extraction/scoring helpers with tests for relative URL canonicalisation, menu/drinks/PDF ranking, low-intent filtering, and deduping.
• Verification: verified npm test (232/232 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, npm run discover:official-menu-sources -- --limit 5 (5 fetched, 2 candidates), and git diff --check.

Official menu crawler MVP ready (2026-06-01)

• PR #129 / merge commit 3ded19a: added a manual, dry-run-first official menu crawler for curated source URLs. It fetches reviewed venue/menu pages, extracts conservative pint-price candidates, and can insert pending price_reports with submission_source = "official_menu" only when run with --write.
• Guardrails: extraction skips happy-hour/special/deal lines in v1, keeps reviewer evidence text, stores raw extraction metadata, and never updates canonical pubs.price. The committed seed file is an example only; real crawl lists stay explicit and reviewed.
• Verification: verified npm test (229/229 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, git diff --check, and a dry-run against scripts/official-menu-seeds.example.json with zero writes.

Price intake plumbing implemented (2026-06-01)

• Branch codex/price-intake-plumbing-spec / commit e2de1c5: added structured price_reports provenance/evidence columns (submission_source, source_url, evidence_text, observed_at, raw_extraction, extractor_version) with a constrained source vocabulary and legacy backfill for menu scans, Tier-C report heroes, and stale flags.
• Intake + moderation safety: public reports now store structured sources, menu-scan submissions no longer rely on note parsing, and admin approval maps provenance/confidence through one helper. Outdated flags are reviewed without touching live prices, happy-hour reports no longer write regular price_history.price, aggregator leads are blocked from direct approval, and admin stats count all pending reports.
• Verification: verified npm test (225/225 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and git diff --check. No Playwright screenshots were needed because the UI change was payload-only.

Price provenance persisted (2026-06-01)

• Issue #76 / branch codex/price-provenance-76 / commit 2c96ffc: added structured current-price provenance fields (price_source, price_verified_at, price_confidence) on pubs, with matching verified_at and confidence audit fields on price_history.
• Write paths + display: Andrew mid-call writes, Andrew post-call fallback writes, admin-approved price reports, and approved pub submissions now persist source, verified time, and confidence. Pub pages can render compact copy such as Checked by Andrew on 31 May 2026, with low-confidence rows labelled.
• Verification: verified npm test (214/214 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, npm run test:e2e (4/4 Playwright tests), git diff --check, and saved desktop/mobile before/after Playwright evidence plus mockups under artifacts/playwright/provenance-*.

Price recency chip humanised (2026-06-01)

• Branch codex/humanize-recency-chip / commit b41ceaf: changed pub-page recency copy from warning-style labels plus a visible date to one plain chip such as Checked 103d ago. The exact date remains available as hover text, but the card now speaks in the dry, normal-person voice.
• Verification: verified focused recency/indexability tests (210/210 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and desktop/mobile Playwright proof for /northbridge/the-court-hotel.

Stale-price warning trimmed (2026-06-01)

• Branch codex/remove-stale-warning-blob / commit d39e819: removed the duplicated stale-price warning paragraph from pub price cards. The compact May be out of date pill and last-verified date remain, keeping the freshness signal without the bulky text block.
• Verification: verified focused recency/indexability tests, npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and desktop/mobile Playwright proof for /northbridge/the-court-hotel.

Playwright PR proof gate ready (2026-06-01)

• Branch codex/playwright-pr-proof / commit 5da3158: installed Playwright Test and added desktop/mobile PR proof smoke coverage for the homepage and a pub detail page, with screenshots attached to the HTML report.
• CI artifacts: the main CI workflow now installs Chromium, runs npm run test:e2e after the production build, and uploads a playwright-proof artifact containing the report and test output. Videos are retained for failures, and npm run test:e2e:video records intentional proof videos when motion or multi-step UI needs it.
• Verification: verified npm test (209/209 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, npm run test:e2e (4/4 Playwright tests), a video-mode pub-page proof run, and the PR #124 GitHub artifact upload.

Price-recency tiers ready (2026-06-01)

• Issue #77 / branch codex/price-recency-tiers-77 / commit a5183ea: added one getPriceRecency helper for fresh (<30 days), aging (30-90 days), stale (91+ days), and unknown price states derived from last_verified.
• Render + prerender: pub pages now receive the recency tier from the server and stale prices show a visible May be out of date badge plus the checked-days count. The same tier feeds dataScore and Tier A/B/C selection, so stale verified prices fall to Tier B and only Tier A pub pages are pre-rendered.
• Verification: added recency-boundary tests and indexability scoring tests. Verified npm test (209/209 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and desktop/mobile screenshots for stale-price pub page /northbridge/the-court-hotel.

Geo-aware Cheaper Nearby module ready (2026-06-01)

• Issue #73 / branch codex/cheaper-nearby-73 / commit cae6329: replaced same-suburb-only pub recommendations with a geo-aware nearby helper. Pub pages now rank verified priced pubs inside a 2km radius, fall back to same-suburb links when sparse, and keep TBC pages useful with nearby verified prices.
• Cross-suburb internal links: the module now renders correct cross-suburb pub URLs, distance text, suburb labels, and price-delta copy such as $4.30 cheaper, with crawler-visible links using the same canonical URL helper.
• Verification: added pure ranking tests for cross-suburb cheaper pubs, sparse-radius fallback, and TBC/unknown-price pages. Verified npm test (203/203 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and desktop/mobile screenshots for priced and TBC pub pages.

Community + Reddit discovery loop documented (2026-06-01)

• Issue #115 / branch codex/community-reddit-loop / commit 3f4d135: added docs/community-discovery-loop.md with the first 10 Reddit/community targets, self-promo norms, and a transparent Pint Index community-post draft.
• Content + tracking: added Reddit-sourced #27/#32 backlog questions and a referral plan that tracks reddit.com, old.reddit.com, reddit, and old-reddit alongside the existing SEO snapshot template.
• Guardrails: explicitly bans inauthentic mention seeding, aged-account boosting, sockpuppets, vote manipulation, unsolicited DMs, link-only replies, and undisclosed promotion.
• Verification: docs-only change; verified with git diff --check.

AI/community SEO snapshot tracking documented (2026-06-01)

• Issue #116 / branch codex/ai-community-seo-snapshot: added docs/seo-snapshots/TEMPLATE.md as the #36 weekly SEO snapshot template, with required AI-referral and community-referral callouts for any source sending >0 sessions.
• GA4 + UTM process: documented the GA4 exploration definitions for AI/search-assistant, community/social, DuckDuckGo, landing-page, and future key-event reporting. Added a lowercase UTM convention for community posts and PR outreach.
• Verification: docs-only change; verified with git diff --check.

Pint Index methodology + CSV citation asset shipped (2026-06-01)

• Issue #117 / branch codex/pint-index-methodology: Pint Index now includes a visible methodology block covering what counts as a pint, which prices are included, how prices are checked, why TBC venues stay out of averages, and how weekly snapshots power the index.
• Citation export: added /insights/pint-index/data.csv as a public CSV download for the Pint Index snapshot series, with text/csv, Content-Disposition, stable headers, and escaped row formatting. /llms.txt now links both the Pint Index methodology page and the CSV export.
• Verification: checked the CSV route locally for 200, text/csv, attachment headers, and live snapshot rows. Verified npm test (200/200 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and desktop/mobile screenshots for /insights/pint-index.

Tier-C report-price hero shipped (2026-05-31)

• Issue #74 / branch codex/tier-c-report-hero: price-less Tier-C pub pages now render the truthful absence stub with last Andrew call timing, verified nearby-price count, and a link to the nearest checked pub instead of leaving the page as a uniform husk.
• Report-price conversion path: Tier-C pages flip the pub-page CTA to a primary Report the price hero button and tag resulting submissions with source:tier-c-report-hero in price_reports.notes for M8 measurement.
• Verification: checked /fremantle/federal-hotel locally as a real Tier-C sample: noindex still present, stub renders, nearest checked pub link renders, and desktop/mobile screenshots show the hero CTA and copy fit. Verified npm test (199/199 tests), npx tsc --noEmit, npm run lint (existing warnings only), and npm run build.

AI citation signals shipped (2026-05-31)

• Issue #35 / branch codex/llms-article-schema: added /llms.txt as a markdown citation map for the live high-signal pages (Pint Index, live insight pages, happy hours, discover, and key suburb pages). The deleted weekly-report route stays deleted; Pint Index is the Article/schema asset per docs/seo-action-plan.md.
• Structured citation + crawler access: added Article JSON-LD to /insights/pint-index with author, publisher, dateModified, and mainEntityOfPage.lastReviewed from real pub freshness data; added explicit GPTBot / PerplexityBot / Google-Extended allow rules while keeping /admin and /api blocked.
• Visible freshness: pub pages now surface Last verified <date> in a semantic <time> element where last_verified exists. Verified npm test (199/199 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, local /llms.txt / robots / Article schema smoke checks, and desktop/mobile screenshots for the pub-page freshness row.

Pub-page SEO Phase 1 schema shipped (2026-05-31)

• Issue #29 / commit e6d861d: replaced disconnected pub-page JSON-LD with one connected @graph: WebPage#webpage points to BarOrPub#pub through mainEntity, links BreadcrumbList#breadcrumb, and exposes dateModified from last_verified where present.
• Schema correctness: removed invalid "$X.XX per pint" priceRange strings and now emits schema-valid $ / $$ / $$$ bands from regular price vs suburb average, with site average only as fallback. Kept the owner decision intact: no Menu, MenuItem, Offer, telephone, or guessed venue hours.
• Review + verification: peer review first blocked merge on site-average banding; fixed with getSuburbAveragePrice, tightened partial-coordinate and invalid-time guards, and re-reviewed cleanly ("Ready to merge? Yes"). Verified npm test (195/195 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and live-row schema checks for Tier-A and Tier-C pubs.

Site voice workstream kicked off (2026-05-31)

• Voice standard + copy specs landed (PR #84, ab4704a): adopted docs/brand-voice-brief.md — the dry/deadpan perthisok register (no exclamation marks, AU spelling, freshness as the trust signal, the 7-question Voice Test as the per-surface gate) — and docs/content-pack-v1.md, whole-site surface copy written as data-fed string builders. Captured as PRD #83 under new milestone #8 "Site Voice & Content".
• PRD #83 sliced into 8 tracer-bullet issues (#86–#93): seededVariant (#86), corny-drift cleanup (#87), hub standfirsts (#88), /[suburb] lead + suburbObservation (#89), homepage + global chrome (#90), 5 guide standfirsts (#91), 5 insight standfirsts (#92), pub-page voice strings voiceCopy (#93, blocked by #86). Copy slices gated AFK + screenshots-on-PR against the Voice Test; all tracked on board #6.
• #86 seededVariant shipped: deterministic anti-sameness phrasing picker (rendezvous/HRW hashing over a cyrb53 string hash) so the ~850 templated pages don't converge — same id always renders the same wording, picks spread across the pool, and appending a variant only ever moves a bounded share onto the new entry (minimal-disruption growth, asserted). Pure src/lib module + co-located tests; no UI, so the screenshot gate doesn't apply. Verified npx tsc --noEmit and 189/189 tests (tsx --test).

Corny-drift cleanup shipped (2026-05-31, #87)

• Swept the weather/guide components (BeerWeather, RainyDay, SunsetSippers) for the corny/American drift the voice brief flags. Found ~4× the content-pack §10 table: replaced exclamation-mark boosterism and stacked slang (Scorcher!, grab a mate/cold one, sesh, hunker down, get amongst it, Hold onto your hat!, Quick!) with dry equivalents; fixed American cozy→cosy in visible prose (the cozy-corners URL slug + the cozyPub data field stay); dropped the brief-flagged That's the Perth way tag. PuntNPints was already clean.
• Words-only — no layout/logic change. Verified npx tsc --noEmit clean + lint (pre-existing react-hooks warnings only). Gated through the brief's 7-question Voice Test. (Humanizer skill not installed this session; the brief's anti-AI-tell standard applied by hand.)

Pub-page SEO Phase 0 shipped (2026-05-31, PR #82)

• Issues #69/#70/#71 / merge commit da2ec80: shipped the indexability/freshness foundation from docs/pub-page-content-plan.md (merged in PR #81). Pub pages now use a tested Tier A/B/C dataScore helper: Tier-C price-less husks stay live for users/internal links but emit noindex,follow; Tier A/B pages remain indexable.
• Sitemap honesty: sitemap.xml now includes only Tier A/B pub URLs, uses real last_verified / updated_at / last_updated freshness dates instead of build-time Date.now(), and derives suburb/static lastModified from all child pub freshness rather than only the indexable subset. Live-data check: 857 routable pubs, 232 indexable pub rows, 397 sitemap URLs.
• ISR + write-path freshness: pub pages now revalidate hourly (3600) and pub reads are cached with pub:${slug} tags; Andrew record-price and public price-report writes trigger revalidateTag + revalidatePath.
• Review + verification: independent sub-agent review initially caught the suburb freshness bug, the fix was reviewed cleanly ("Happy to merge: yes"). Verified npm test (181/181 tests), npx tsc --noEmit, npm run lint (existing warnings only), npm run build, and runtime spot checks for Tier-C robots + sitemap exclusion.

Andrew guardrails shipped (2026-05-31, PR #68)

• Issue #63 / merge commit 97b3d63: revived the unmerged Andrew call-safety work on a fresh codex/andrew-dnc-cooldown branch. Adds phone_call_log-based do-not-call exclusion, 72h-minimum call cooldown in /api/pintsweep/kickoff, queued-call reservations before ElevenLabs batch submit, unique reservation IDs, call_initiation_failure logging, post-call fallback persistence from ElevenLabs data collection, and scripts/mark-dnc.mjs for manual DNC markers.
• Review + closeout: first peer review blocked merge on repeated-call gaps; e41ecba addressed those gaps and f150f6a fixed the second-review reservation ID collision risk. Final peer review returned no findings and "Ready to merge? Yes." #63 is closed and the roadmap card is Done.
• Verification before merge: npx tsc --noEmit, npm test (177/177 tests), npm run lint (passes with the existing SubmitPubForm/SunsetSippers warnings), and npm run build.

HappyHourEngine refactor shipped (2026-05-31)

• Issue #56 / commit 2279802: collapsed the dual happy-hour engines into the structured happyHourLive path and deleted the lossy src/lib/happyHour.ts parser. happyHourLive now owns active/starting-soon status, effective price, minutes remaining, starts-in-minutes, countdown text, and structured labels from raw happy_hour_days/start/end fields.
• Fixed the two deferred HH consumers: DiscoverClient and TonightsMoves no longer re-parse pub.happyHour; both read structured schedule fields through the unified engine, so weekend-only, single-day, half-hour, and am-to-pm windows behave correctly.
• Regression coverage + review: added injected-clock tests for weekend half-hour active windows, am-to-pm starting-soon windows, and the independent-review catch that already-ended same-day deals must not show as "Starting Soon". Verified tsc, 170/170 tests, next build, and desktop/mobile screenshots for /discover + /insights/tonights-best-bets (local-only Vercel Analytics 404 expected off Vercel).

Backlog organised into milestones + dependency triage (2026-05-31)

• Milestones: closed the catch-all "SEO Push — Q2 2026" and split the board into 6 themed milestones — SEO: Indexing & Technical (#2), SEO: Content & Schema (AEO) (#3), SEO: Measurement & Growth (#4), Architecture Refactor (#5), Data Coverage (#6), Andrew (voice agent) (#7). Every open issue is bucketed; the 10 new issues (#56–#64, #66) were added to project board #6.
• Ticketed the untracked work: refactor as #56–#60 (HappyHourEngine, Formatters, PerthNow.date, price_snapshots accessor, useUserLocation); data as #61–#62; Andrew as #63 (revive DNC/cooldown/post-call hardening) + #64 (voice-quality decision). #25 (www→apex) closed as a no-op.
• Dependency triage (0 open PRs): merged #18 (setup-node 4→6), #19 (checkout 4→6), #38 (minor/patch group), and #65 (consolidated lucide 1.11 + twilio 6 + @vercel/analytics 2.0.1 — one clean lockfile; tsc + 167 tests + build green). Next 14→16 (#21) deferred — peer-reviewed as a coordinated Next 16 + React 19 + ESLint 9 upgrade, closed + dependabot-ignored, tracked as #66. Prod smoke-tested green on deploy 755d156 (routes 200, lucide icons render).
• Branch hygiene: pruned all merged branches + 6 dead ones; kept claude/audit-e2e-improvements-WQb6N (it is #63's source). Remote branches now: main + that one.

Documentation accuracy sweep (2026-05-31, PR #55)

• Audited every doc, the repo, and the open issues against what's actually live, then corrected or deprecated the drift. Docs + issue metadata only — no code changes.
• CLAUDE.md / README: lib count 11→13 (added supabaseGateway, urls); legacy /pub+/suburb stubs corrected 308→301; /guides+/insights noted as 308-redirects to /discover; homepage nav fixed (no "Pint Report" — it's Discover + Happy Hours + a "Submit a Price" CTA); dropped stale "pub crawl / 23 pages / leaderboard / pub golf / weekly report" references.
• SEO-MASTER.md: added a dated reconciliation banner; removed deleted features from the sitemap/ISR/keyword tables; repointed /pub/[slug]+/suburb/[slug] to the live /[suburb]/[pub]+/[suburb] routes; FID→INP; GBP marked deprioritised.
• architecture-refactor-plan.html: banner marking it superseded (CrawlPlanner/GolfScoring/Phase-6 moot post-deletion; CEO decisions resolved).
• seo-action-plan.md: #1 (www→apex) downgraded — proven live that www already 308-redirects and 308 ≡ 301, so it's not the "biggest win"; #2 (legacy 301) marked done; #4 GA4 event list pruned of the deleted pint_crawl_start/pub_golf_start.
• Issues: #25 corrected + downgraded; #28/#29/#31/#35 refreshed (deleted-feature refs removed, #31 unblocked).
• Live verification today: 857 pubs (API), sitemap 1022 URLs and clean, all 4 deleted-feature routes 404, /suburb/*→301, llms.txt still 404 (#35 open), pub pages still missing FAQPage/MenuItem schema (#27/#29 open).

Refactor Phase 2 (Pub mapper): one row→Pub mapper (2026-05-30)

• PR #52 (39ef38d). getPubs / getPubBySlug / getNearbyPubs / getSimilarPricePubs each carried a byte-identical ~52-line block that computed live happy-hour status + effective price and built the 30-field Pub object. Collapsed into a single toPub(row) in src/lib/supabase.ts; each accessor now delegates. Net −133 lines, behaviour unchanged.
• getPubsLite (smaller PubLite type) and getPriceHistory (different shape) left as-is.
• Verified: tsc clean; 167/167 tests; next build green (service key unset) — all 857 pub pages prerender through toPub on real data; live pub-detail page renders correct name/suburb/price. Independently reviewed (LGTM).

Refactor Phase 1 (PubUrls): one suburb-slug source of truth (2026-05-30)

• PR #50 (957c886). Collapsed 14 copy-pasted suburb-slug functions into one. toSuburbSlug moved into the pure src/lib/urls.ts seam (it had lived in supabase.ts — a backwards dependency for a URL helper); supabase.ts re-exports it for back-compat. Every pub/suburb link now builds via pubUrl / suburbUrl / absolutePubUrl. Deleted 13 byte-identical inline toSuburbSlug copies across components + pages.
• Fixed a live 404: SuburbLeague used a divergent toSlug that omitted the apostrophe-strip, so it linked the suburb O'Connor to /o-connor (404) instead of the canonical /oconnor (200) — the only one of 150 live suburbs with an apostrophe. Now routed through suburbUrl.
• SEO tripwire: src/lib/urls.test.ts snapshots all 150 live production suburb slugs (pulled from Supabase) so any future slug-logic change that would move a Google-indexed URL fails CI first.
• Verified: tsc clean; 167/167 tests (snapshot stable ⇒ zero URL drift); next build green with the service key unset; sitemap.xml emits 1022 canonical URLs with zero /o-connor; routes resolve (suburb/pub 200, legacy /pub/* 301). Independently reviewed (LGTM) before merge.

DB security lockdown + architecture keystones (2026-05-30)

• Refactor Phase 0 + SupabaseGateway (PRs #45–#48): added src/lib/perthClock.ts (the time keystone) and src/lib/supabaseGateway.ts (anonClient() / serviceClient()), centralising 16 inline createClient(...) blocks and cutting the duplicated anon-JWT literal from ~14 files to 1. Fixed a real security bug — admin/review silently degraded service-role→anon (#47). Deleted 4 dead files. A CI deploy gate (#43) now runs node:test via tsx and blocks merges.
• RLS audit + lockdown: a probe-driven scare ("anyone can write prices") was a false alarm — pubs / price_history / price_snapshots were always service-role-only. The audit did find 3 over-permissive tables: after migrating their writer routes to the service-role key (#48), dropped the public WRITE + read-leak policies on pub_price_cache, push_subscriptions, and agent_activity. Also fixed the stale Pint Index (the weekly-snapshot crons had been failing silently on anon writes). The DB change is recorded in supabase/migrations/20260530120000_rls_lockdown.sql.

Removed non-core features (2026-05-30)

• Product call: killed everything that isn't the core pint-price experience. Deleted pub-golf, pint-crawl, leaderboard, and weekly-report (the "Pint Report" stats page) — all had 0 organic clicks/impressions over 90 days and are rebuildable from git history.
• Removed the 4 page routes + the orphaned /api/weekly-report, and cleaned every reference: sitemap entries, the /discover "Activities" carousel, and nav/footer links (MobileNav incl. now-unused icon imports, Footer, SubPageNav, HomeClient, homepage SSR crawl-links). Fixed the weekly push deep-link (was broken at /weekly) → /insights/pint-index. 18 files, −2,635 lines. tsc + next build green; /discover visually verified (no empty section, footer nav trimmed).
• Twilio left untouched (separate AI phone-call price-collection system). The refactor plan's CrawlPlanner/GolfScoring modules + Phase-6 game-teardown are now moot.
• Commit: 0f93002

Whole-app refactor plan + deploy gate (Phase -1) (2026-05-30)

• A 17-agent workflow (survey → 4 independent architects → reconcile → red-team → finalize) produced a full module architecture for the app, rendered as docs/architecture-refactor-plan.html. Headline: the app has ~8 problems copy-pasted 50+ times; the fix is 24 deep modules across a 4-layer onion (substrate seams → pure kernels → data accessors → external-I/O → UI) + a giant teardown, in an 8-phase migration (Phase -1..6). Builds on the earlier 7-candidate review (folded in).
• Shipped Phase -1 (the deploy gate): main auto-deploys to production but CI ran only tsc/lint/build — tests were not gated, and the two node:test unit tests couldn't even run (node couldn't resolve their extensionless TS imports). Added tsx + a test script and wired Test + redirect-test steps into the existing Typecheck, lint, build CI job (job name kept verbatim to preserve the branch-protection required check). Bumped CI Node 20 → 22.
• Surfaced (not yet fixed) two live security holes for later phases: an ElevenLabs HMAC shared-secret bypass (post-call:66) and a service-role→anon silent degrade (admin/review:6).
• Two CEO decisions still open: keep/kill the game features pint-crawl + pub-golf (both have 0 organic clicks/impressions over 90 days per GSC; in-app usage is unmeasurable — no working GA4/first-party analytics), and whether Twilio (in package.json but unused in code) is still in the stack.
• Commits: 76f11aa (plan doc), 5a716ca (Phase -1 deploy gate)

Live-bug fixes from architecture review (2026-05-30)

• A multi-agent architecture-review workflow surfaced three "live bugs"; two were real (fixed here), one was a false alarm.
• Happy-hour badge (5247efb): the pub list/card components re-parsed the lossy pub.happyHour string through happyHour.ts's regex, which can't match Weekends, single days, or am-pm spans (and the generated string drops :30 minutes), so the "HH" badge silently showed off during weekend/half-hour/midday happy hours. PubCardList, SunsetSippers, and PuntNPints now read the structured pub.isHappyHourNow (computed by happyHourLive from raw fields). TonightsMoves + DiscoverClient deferred — their isToday/countdown use needs the happyHourLive extension (arch-review rank 1).
• Pint Index badge (f609b48): PintIndexBadge fetched price_snapshots ascending + limit(30) = the 30 oldest snapshots, so the homepage badge price and "% this week" came from the start of history. Now fetches the newest 30 (descending) and reverses to chronological order.
• Not a bug: record-price not writing happy_hour_price — Andrew's tool never sends a happy-hour price, so nothing is dropped. Latent gap (arch-review rank 5), not a live bug; left untouched.
• Next body of work is the 7-candidate architecture deepening backlog from the same review (rank 1 = consolidate the happy-hour engine into one deep happyHourLive module and delete happyHour.ts).
• Commits: 5247efb, f609b48

SEO redirect consolidation (2026-05-03)

• Fixed the legacy redirect half of the highest-priority canonical redirect work:
  • Legacy /suburb/[slug] redirects now return 301 to the current /[suburb] URL in production.
  • Legacy /pub/[slug] redirects now return 301 to /[suburb]/[pub] in production.
• Updated the app-level www.perthpintprices.com redirect rule to request 301, but production verification showed Vercel's project-domain redirect still returns 308 before app config. Issue #25 is reopened; remaining action is to update the Vercel domain configuration for www.perthpintprices.com with redirectStatusCode: 301.
• Added npm run test:redirects to guard the redirect status codes against regression.
• Commits: 8818575, 6eea19a

Project governance + CI infrastructure (2026-04-27)

• README, CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, CHANGELOG.md
• .github/CODEOWNERS, dependabot.yml, PR template, 5 issue forms (bug, feature, SEO action, content, data quality)
• CI workflow on every PR + push to main (typecheck, lint, build) — required status check on main
• ESLint wired up (was previously absent — next lint ran interactively); 9 pre-existing warnings tolerated, 3 pre-existing errors fixed
• Custom labels: triage, blocked, in-progress, p0/p1/p2, feature, chore, seo, andrew, content, data, security, performance, schema, indexing, ci, dependencies
• Repo description + 15 topics set, branch protection on main (no force-push, no deletion, required CI, required conversation resolution)
• Milestone "SEO Push — Q2 2026" with 12 issues filed against it (#25-#36)
• Commits: c98bdf3, cf00ce5, c24cc72, 7ef63b2, e30491d

SEO research + action plan (2026-04-27)

• Pulled GSC + GA4 data: 131 clicks / 5k impressions / 2.6% CTR / position 17.1 over 90 days, 838 unique queries, 28% of pages not indexed (439 of 1,573)
• docs/seo-research-2026.md — what's new in late-2026 SEO since SEO-MASTER was written. AI Overview / AEO / GEO playbook, Information Gain post March 2026 core update, MenuItem schema, AU local quirks, INP replaces FID. 27 sources cited.
• docs/seo-action-plan.md — prioritised 12-action punch list driven by the GSC + GA4 data. Each issue in milestone #1 maps to a section here.
• Headline gaps: www vs apex canonical splitting (~38 clicks at risk), legacy /pub///suburb/ 308s still indexed, 373 "Discovered, currently not indexed", 0 GA4 key events, 12 zero-click queries on page 1
• Commit: cf00ce5

Andrew voice agent: full pipeline + voice tuning (2026-04-26 + 2026-04-27)

• Production pipeline already shipped from a parallel session before this work began:
  • src/app/api/agents/record-price/[slug]/route.ts — mid-call webhook (partial-data tolerant, unit conversion)
  • src/app/api/agents/post-call/route.ts — HMAC-signed post-call webhook with Claude transcript fallback
  • src/app/api/pintsweep/kickoff/route.ts — batch trigger using ElevenLabs Batch Calling, 24h dedupe, openNow filter
  • 70-term Scribe v2 keyterm list, 6 few-shot transcripts, DTMF/IVR navigation
• First successful real-pub capture (2026-04-26): $12.80 Great Northern Super Crisp at Kalamunda Hotel from Cassie
• ElevenLabs MCP wired into ~/.claude.json user-level
• Nexelle agent + phone number deleted from ElevenLabs; voice-proxy Vercel project deleted; local source removed
• Voice config tuned per docs/andrew-voice-research.md: model eleven_flash_v2 → eleven_v3_conversational (Expressive Mode), stability 0.30 → 0.70, similarity_boost 0.85 → 0.75, turn eagerness eager → patient. flash_v2_5 was rejected by API ("English Agents must use turbo or flash v2"), v3 conversational got through via expressive_mode toggle.
• docs/andrew-voice-research.md — voice model + TTS tuning research
• Commits: d03b883, ae4b44e (reverted in 6d4cd96 — duplicate of pre-existing route)

Arvo → Perth Pint Prices rebrand finished (2026-04-27)

• Companion to the earlier code/SEO-title rebrand (5b3df48). Replaced remaining "Arvo" references in CLAUDE.md, PROJECT-STATUS.md, SEO-MASTER.md, the price verification kit, the service worker, the PWA manifest, and every guide/insight/suburb page.
• Homepage title now uses Next's title.absolute pattern; "Arvo" preserved as alternateName in WebSite JSON-LD for SEO continuity
• SubPageNav brand mark switched from SVG icon to text-only (<amber>Perth</amber> Pint Prices)
• Commit: 5a11568 (39 files, 157 insertions, 204 deletions)

Pub location audit and fix (2026-03-10)

• Audited all 423 pub lat/lng coordinates against Google Places Text Search API
• Fixed 146 incorrect locations: 126 auto-applied, 11 user-confirmed, 9 geocoded from addresses
• Accuracy improved from 64.5% to 97.6% within 100m of Google's coordinates
• Scripts: audit-locations.js, apply-confident-fixes.js, apply-manual-fixes.js (gitignored)

Homepage UX overhaul + dead code cleanup (2026-03-10)

• Filter section consolidated, distance dropdown swaps in for suburb when sorting Nearest, copy aligned with community-powered messaging
• 16 unused components removed (1,676 lines), src/lib/emoji.ts deleted, 163 testing screenshots deleted
• Commits: 48f139f, 9fe1345

SEO + Humanizer audit (2026-03-09)

• BreadcrumbJsonLd schema bug fixed, Twitter cards added, sr-only H1s added, titles/descriptions trimmed
• Em dashes replaced with periods in 5 components
• Commits: 2a6e082, f3e93b5

Research agent merge (2026-03-09)

• 190 research agent JSON files merged: 49 updates + 3 new pubs into Supabase via scripts/merge-research.js

What's still to do

Architecture refactor (in progress — full plan in docs/architecture-refactor-plan.html)

From the 2026-05-30 architecture review. Done: Phase -1 (deploy-gate CI), Phase 0 (PerthClock + SupabaseGateway keystones), DB security lockdown, Phase 1 (PubUrls — one slug source of truth + 150-suburb SEO snapshot test, fixed the O'Connor 404), Phase 2 (toPub row→Pub mapper, −133 lines). Remaining, roughly by size:

• HappyHourEngine (biggest / highest-value) — collapse happyHour.ts + happyHourLive.ts into one "effective-price-now" module and delete the lossy string parser. Also fixes two deferred happy-hour badge bugs: TonightsMoves + DiscoverClient re-parse the lossy pub.happyHour string instead of the structured pub.isHappyHourNow, so their badge/countdown is wrong for weekend / half-hour / am-pm windows (their isActive is entangled with isToday/countdown, which needs this engine first).
• Formatters — relocate formatPrice.ts / priceLabel.ts into one seam.
• PerthNow.date footgun (Phase-0 review nit) — its UTC fields hold Perth wall-clock, so .getHours()/.getDate() on it double-shifts. Rename or drop.
• price_snapshots accessor (5 fetch/coerce sites) and useUserLocation() hook (6 getCurrentPosition copies).
• Tiny: getPubsLite's prelude still duplicates toPub's happy-hour computation.

Now ticketed under milestone Architecture Refactor (#5): #56 HappyHourEngine (p1), #57 Formatters, #58 PerthNow.date footgun, #59 price_snapshots accessor, #60 useUserLocation. Plus #66 — the coordinated Next 16 + React 19 + ESLint 9 upgrade (deferred from Dependabot #21).

Operating loop: each increment ships behind an independent code review + full verification (tsc, node:test, next build with SUPABASE_SERVICE_ROLE_KEY unset); merge to main = production deploy. Continue autonomously; only surface 100%-real blockers.

SEO push — Q2 2026 (3 milestones: SEO: Indexing & Technical #2 · SEO: Content & Schema (AEO) #3 · SEO: Measurement & Growth #4)

See docs/seo-action-plan.md for the prioritised punch list. Top of the list:

• #25 www → apex — closed (no-op: www already 308-redirects and Google treats 308 ≡ 301; vercel.json is set to 301). The only residual lever was a GSC reindex nudge; not a ranking risk.
• #27 Reclaim zero-click page-1 queries with answer-first blocks + FAQPage schema (4h, p1)
• #28 Configure GA4 Key Events (30m, p1)
• #29 Add answer-first block + MenuItem schema to all 857 pub pages (1-2d, p1)
• #30 Get 373 "Discovered not indexed" crawled (1d, p1)
• #31 Resolve 39 "Duplicate, Google chose different canonical" (unblocked — legacy 301 shipped + #25 closed; re-pull the current count from GSC)
• #32 Build missing high-intent landing pages (3-5d, p1)
• #33 Press-pitch the Pint Index to WA media (1w, p2)
• #34 Migrate to next/image + INP audit (4-6h, p1)
• #35 Add llms.txt + Article schema for AI citation (1h, p2)
• #36 Set up weekly SEO snapshot + GSC alerts (15m/wk, p2)

Data

• 663 of 857 pubs missing regular prices (was 222 in March; new pubs are being added faster than prices are being collected)
• Andrew (the voice agent) is the strategy — see agents/andrew.json and the /api/pintsweep/kickoff batch trigger
• Consider price refresh strategy for stale prices (some pubs haven't been updated in months)
• Ticketed under milestone Data Coverage (#6): #61 backfill the 663 missing prices (p1), #62 stale-price refresh strategy.

Andrew voice agent (milestone Andrew (voice agent) #7)

• #63 revive do-not-call + cooldown + post-call-fallback hardening — draft PR #68 is open from codex/andrew-dnc-cooldown (not merged/deployed yet). It supersedes the old unmerged claude/audit-e2e-improvements-WQb6N source branch; keep the old branch until #63 ships.
• #64 voice-quality decision. Place test calls with phone OFF silent to evaluate cadence/volume after the v3 conversational + stability 0.70 change
• If sound quality is good: kick off Professional Voice Clone (PVC) per docs/andrew-voice-research.md §3 — book a 30 min recording session via Voicebooking / Voices.com or record a willing AU male mate. Cleanest signal for production phone bots.
• If still inconsistent: pilot Cartesia Sonic-3 + Line per research §6

Dependabot — triaged 2026-05-31 (0 open PRs)

• Merged: #18 setup-node 4→6, #19 checkout 4→6, #38 minor/patch group, #65 (lucide 1.11 + twilio 6 + @vercel/analytics 2.0.1, consolidated — one clean lockfile; tsc + 167 tests + build green).
• Deferred: next 14→16 (#21) — closed + @dependabot ignore'd; tracked as #66 (coordinated Next 16 + React 19 + ESLint 9 upgrade) under milestone Architecture Refactor.

Stale PRs to clean up (#1-#12)

Done — all of those old Feb–Mar 2026 PRs are closed (0 open PRs ≤ #12), and as of 2026-05-31 the Dependabot backlog is cleared too (see above) — 0 open PRs total.

Features (ideas, not committed)

• Price alerts / watchlist notifications
• Pub comparison tool
• Historical price charts on pub detail pages

Manual setup the user owns

• GA4 Key Events: per issue #28, configure 6 key events in GA4 Admin → Events
• Twilio cleanup: the orphaned Nexelle phone number +61851226384 is still allocated and billing on the Twilio side — release it in the Twilio console if you want it gone.

Project board

Perth Pint Prices — Roadmap (v2 Project, Team Planning template).

• Linked to iamjohnnymac/perthpintprices; holds the roadmap issues across the 6 milestones (22 items). The 10 issues created 2026-05-31 (#56–#64, #66) were added to the board.
• Built-in views: Backlog (table grouped by Status), Board (Kanban: Todo / In progress / Done), Current iteration, Roadmap, My items
• Fields: Status, Priority (P0/P1/P2), Size (XS-XL — used as effort proxy: XS=Quick win, S=Half day, M=Full day, L=Multi-day, XL=Ongoing), Estimate (number), Iteration (sprint), Start date, Target date, plus custom Area (SEO / Andrew / Content / Performance / Schema / Indexing / Chore)
• Each item already tagged with Priority + Size + Area; Status defaults to Todo

Note: project was created via the GitHub web UI (Team Planning template + bulk import) after gh project create produced one where API-added items were silently invisible in views (forward query returned 0 even when reverse lookup confirmed items attached). UI-created project works. Don't recreate via CLI.

Utility scripts (local only, gitignored)

• scripts/merge-research.js, analyze-json.js, compare-prices.js, fix-seo.js, audit-locations.js, apply-confident-fixes.js, apply-manual-fixes.js, backfill-phones.mjs, backfill-place-ids.mjs, discover-venues.mjs, finalize-backfill.mjs, insert-discovered-venues.mjs, sample-voices.mjs, sample-voices-female.mjs, test-responsive.mjs