本番のトレンド版を main に統合し、Cloudflare への自動デプロイを追加

.env.example

@@ -5,3 +5,8 @@ TYPESAFE_DEFAULT_MODEL=jev-latest
 
 # Optional original live feed
 NEXT_PUBLIC_WEBSOCKET_URL=
+
+# Server-side live Japanese X trends (15-minute cache)
+TREG_TOKEN=
+# Required for an identity token; use your own active team slug.
+TREG_ORG=

.github/workflows/deploy.yml

@@ -0,0 +1,66 @@
+name: Test and deploy to Cloudflare
+
+# Pull requests: tests, type check and a Cloudflare build with a Workers dry run.
+# Pushes to main (and manual runs): the same checks, then `wrangler deploy` to meta-earth-wave.com.
+# Runtime secrets (TREG_TOKEN, TREG_ORG, TYPESAFE_API_KEY, NEXTAUTH_SECRET) live on the Worker and survive deploys.
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.event_name == 'pull_request' && format('pr-{0}', github.event.number) || 'deploy-production' }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    env:
+      WRANGLER_VERSION: 4.129.1
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version-file: .nvmrc
+          cache: yarn
+
+      - name: Install
+        run: yarn install --frozen-lockfile
+
+      - name: Test
+        run: node --test tests/*.cjs
+
+      - name: Type check
+        run: yarn -s tsc --noEmit
+
+      - name: Build static assets for Workers
+        run: yarn build:cloudflare
+
+      - name: Workers dry run
+        run: npx --yes wrangler@${WRANGLER_VERSION} deploy --dry-run
+
+      - name: Deploy to Cloudflare
+        if: github.ref == 'refs/heads/main' && (github.event_name == 'push' || github.event_name == 'workflow_dispatch')
+        uses: cloudflare/wrangler-action@v3
+        with:
+          apiToken: ${{ secrets.CLOUDFLARE_API_TOKEN }}
+          accountId: ${{ vars.CLOUDFLARE_ACCOUNT_ID }}
+          wranglerVersion: ${{ env.WRANGLER_VERSION }}
+          command: deploy --message "${{ github.sha }}"
+
+      - name: Smoke test production
+        if: github.ref == 'refs/heads/main' && (github.event_name == 'push' || github.event_name == 'workflow_dispatch')
+        run: |
+          build_id=$(cat .next-cloudflare/BUILD_ID)
+          for i in $(seq 1 10); do
+            if curl -fsS https://meta-earth-wave.com/ | grep -q "/_next/static/${build_id}/"; then echo "meta-earth-wave.com serves build ${build_id}"; break; fi
+            [ "$i" = 10 ] && { echo "::error::meta-earth-wave.com is not serving build ${build_id}"; exit 1; }
+            sleep 6
+          done
+          curl -fsS -o /dev/null -w "GET /api/themes -> %{http_code}\n" https://meta-earth-wave.com/api/themes

.gitignore

@@ -34,3 +34,8 @@ yarn-error.log*
 
 # typescript
 *.tsbuildinfo
+
+# Cloudflare
+.wrangler/
+.next-cloudflare/
+.dev.vars*

CLOUDFLARE.md

@@ -0,0 +1,71 @@
+# Cloudflare deployment
+
+URL: https://meta-earth-wave.com
+
+Fallback: https://metaearthwave.watanabe-koto.workers.dev
+
+The apex custom domain is managed by the `routes` entry in `wrangler.jsonc`.
+
+The existing Next.js 12 frontend is statically exported; `cloudflare/worker.ts`
+adapts the existing Pages API handlers to Workers requests and responses.
+No Next.js server or SSR endpoint is exposed. API handlers, including NextAuth,
+run with Node compatibility. `next/headers` is stubbed only for NextAuth's unused
+App Router branch. Pages Router requests always pass explicit request/response objects.
+
+## CI/CD
+
+`.github/workflows/deploy.yml` runs on every pull request and on pushes to `main`:
+tests, type check, `yarn build:cloudflare` and `wrangler deploy --dry-run`.
+On `main` (push or manual "Run workflow") it then runs `wrangler deploy` and checks that
+meta-earth-wave.com serves the new build ID and that `/api/themes` answers.
+
+Repository settings it needs:
+
+- Secret `CLOUDFLARE_API_TOKEN`: a Cloudflare API token from the "Edit Cloudflare Workers" template
+- Variable `CLOUDFLARE_ACCOUNT_ID`: the Cloudflare account ID (not secret)
+
+Worker runtime secrets are managed with `wrangler secret put` as below; CI never sees them.
+To roll back, use `wrangler rollback` or pick an earlier version in the Cloudflare dashboard.
+
+## Build and publish (manual)
+
+Use Node 22 (see package.json engines) and Wrangler 4.129.1 or later:
+
+```sh
+npm run build:cloudflare
+wrangler deploy --dry-run
+wrangler deploy
+```
+
+The build uses `.next-cloudflare` so the ordinary local `.next` build remains separate.
+Images are served directly. The unused archival `public/earth.jpg` (28 MB)
+is excluded via `out/.assetsignore`; the actual earth texture stays in the deployment.
+
+Set server-only secrets using `wrangler secret put NAME`:
+
+- TYPESAFE_API_KEY: Jev analysis
+- TREG_TOKEN, TREG_ORG: trend/post retrieval
+- NEXTAUTH_SECRET: session signing
+- TWITTER_CLIENT_ID, TWITTER_CLIENT_SECRET: optional X login
+- NEXTAUTH_URL: set to the public origin when configuring X login
+
+Never use NEXT_PUBLIC_* for authentication secrets.
+The initial deployment has Jev/treg credentials and NEXTAUTH_SECRET.
+X credentials and a posting WebSocket server were not present locally, so X login/posting
+is not configured. Geolocation requires the visitor's browser permission.
+Individual waves remain local to the visitor's browser, as before deployment.
+
+## Verification on 2026-09-25
+
+Production build and export, Workers dry run, and 24 existing tests passed.
+The deployed site renders the globe. Live Jev input and trend availability were
+checked separately; a deployment does not replenish the external treg balance.
+Theme caches are per Worker isolate, not a durable shared cache.
+
+## Refresh interval
+
+Trend refresh, server cache TTL and treg accepted cache age are 6 hours.
+Existing trend waves remain visible for up to two refresh periods (12 hours)
+so they do not disappear before the next refresh. Refresh remains visitor-driven,
+not a centralized scheduled job; the interval is not an account-wide spend cap.
+Personal input preview remains real-time (minimum request start spacing 800 ms).

WAVES.md

@@ -12,12 +12,16 @@ is pinned in package.json and .nvmrc and verified by the production build.
 Copy `.env.example` to `.env.local` and set `TYPESAFE_API_KEY` on the server.
 Without a key, emotion analysis is disabled and POST returns 503. No keyword/demo
 fallback is included. Upstream errors return 502, with a 12-second request timeout.
-Real Jev accuracy and latency have not been verified with a live credential.
+Live Jev calls were verified locally for SNS samples and a user-authored wave.
+Model accuracy has not been independently evaluated.
 
 ## Usage and storage
 
-START → 波を生み出す → text and location → 感情の波を生み出す.
-The location defaults to Tokyo; users can choose a city or enter latitude/longitude.
+START → 波を生み出す → text → browser location permission → 感情の波を生み出す.
+User-authored waves use browser geolocation at submission time. No city or coordinate
+selector is shown. A denied, unavailable, or timed-out location blocks submission
+with a retryable message; no default location is silently substituted. Geolocation
+is requested before paid analysis, with a 10-second timeout and a 60-second cache.
 The message panel displays the five estimated strengths without test sliders.
 
 Jev posts live in this browser session only (20 maximum); reload clears them.
@@ -33,7 +37,7 @@ and empathy mixes gently. The five strengths blend with normalized weights.
 Colored waves remain within radius 0.16 on the unit sphere; weaker displacement
 travels outward at 0.065 units/second and fades by radius 0.36. A post does not
 change the entire globe or background. Posts expire after 60 seconds, fading
-between seconds 42 and 60. The 20 available visual layers prioritize newer posts
+between seconds 42 and 60. The 40 available visual layers reserve one blended layer per mapped region and prioritize newer personal posts
 and never split a blend when capacity is exhausted.
 
 ## Wave captions
@@ -57,6 +61,182 @@ Planner: `src/lib/waveCaption.ts`, renderer: `src/lib/waveCaptionDraw.ts`, overl
 
 ## Validation
 
-`node --test tests/wave-emotion.cjs tests/wave-caption.cjs`
+`node --test tests/*.cjs`
 `yarn tsc --noEmit`
 `yarn build`
+
+## Live themes
+
+Set `TREG_TOKEN`, optional identity-token `TREG_ORG`, and `TYPESAFE_API_KEY`
+in server-side environment variables. Do not prefix these with NEXT_PUBLIC.
+The original globe page now offers five Japanese X trend themes, automatically
+refreshing while the page is open every 15 minutes. No background scheduler runs
+when no one is viewing the site.
+
+GET `/api/themes` retrieves Japanese trends via TikHub through treg. Category
+coverage is preferred over a pure rank list (sports, culture, society, everyday).
+Exact hashtag/width/case variants are merged; semantic event alias merging and
+inference of which current match is being discussed are not implemented.
+GET `/api/themes/:id` accepts only an ID in the current list. It follows up to
+12 AnyAPI search pages (requested limit 50, actual provider page size may be lower),
+aiming for 120 recent posts in the last 24h, one per author, deduplicating text and
+IDs. Fewer than 100 posts is explicitly marked as a limited sample. Search is no
+longer restricted to Japanese, but Japanese trend queries still bias the sample;
+these results do not represent entire countries or the world.
+
+Batches of 10 posts (at most 3 in flight) score five emotions and infer a broad
+posting region using text, language and public profile location when returned by
+the search provider. Missing profile fields are left empty; no extra profile
+lookup runs. The current atlas supports 24 countries and eight Japanese regions.
+It does not infer nationality or exact personal coordinates. News locations,
+travel destinations and team names are not sufficient posting-location evidence.
+Chosen-region probability must reach .70; language-only assignments require .85
+and are restricted to coarse Japanese, Korean and Thai locations. These model
+probabilities are not calibrated geolocation accuracy. Other ambiguous languages,
+unsupported regions and insufficient evidence remain unknown and are not mapped.
+
+Each mapped region averages its own five emotions, blending color and movement
+at its display centroid with a country/region sized footprint. Centroids are
+visual anchors, not measured origins. No unknown posts fall back to Tokyo. Group
+buttons rotate the globe to that region. Public responses contain aggregate scores,
+counts, inference notes and source links, not author profiles or post text. Details
+show the actual sample size and unknown count. Browser geolocation remains the
+source for the user's own submitted wave.
+
+The aggregate mood gently changes ambient flow speed over four seconds, only
+when at least three mapped samples exist. A single user-authored post never
+changes global flow. Local region counts can still be small even with 120 total
+posts, and should not be interpreted as population opinion.
+
+Aggregates animate as a dated visualization until their age reaches 30 minutes;
+this is not a live firehose or a count of newly arriving posts. User-authored
+waves last 60 seconds and retain their chosen topic and browser-provided coordinates. Unrelated
+local topic posts are hidden when another theme is selected.
+
+Requests are coalesced with a 15-minute bounded in-process cache, a 1-minute
+failure backoff, and stale results explicitly labelled for at most 30 minutes.
+Treg uses per-endpoint/query/time-window idempotency keys. Caches are per hosting
+instance; multi-instance Jev spend is not globally capped. Configure provider
+spending limits for production, and use a shared cache for large deployments.
+Observed catalog prices: trends $0.001/call, search ~$0.00075/call; Jev is billed
+separately. A theme performs up to 12 batch inferences, never one per frame. Search costs
+at most approximately $0.009 per theme, plus separately billed inference.
+The first analysis is a blocking request with a four-minute client timeout.
+Production hosts with short request limits need a background job and shared
+result storage before this larger analysis can be deployed reliably.
+
+## Globe-first X discovery (current UI)
+
+The permanent theme selector is replaced with clickable waves on the existing globe.
+`/api/themes` assigns a broad topic-related location through a bounded inference
+of the title/category. This is not posting-origin inference. If confidence is below
+.85 or inference fails, it uses the explicitly labelled Japan trend observation
+anchor. This does not claim a real event origin. The source feed remains Japan;
+no multi-country collector was added after the interaction design changed.
+
+Waves at the same anchor are grouped; clicking opens a compact list of X search
+links. A single-topic wave opens its X search directly. Every link encodes the
+original title (including hashtags), opens a new tab with noopener/noreferrer,
+and leaves the globe available. Waves stop after 30 minutes without a fresh list
+and disappear on the next list update when no longer selected. Selection covers
+five category-diverse topics, not every trend.
+
+Emotion analysis runs for the five visible topics with two client jobs at a time;
+search remains bounded at 12 calls per topic (about $0.045 total maximum for five),
+plus inference. The neutral discovery wave is available immediately, then receives
+the overall topic's emotion blend after analysis. Co-located topics share a mixed
+wave and have individual emotion-colored dots in their link list. Errors leave a
+neutral wave and never manufacture scores. Trend rank and sample size are not
+presented as measured discussion volume. Provider fees and caches above still apply.
+
+## Global floating keywords (supersedes globe-first regional anchors)
+
+The current discovery view collects the top five trends from eight markets:
+United States, United Kingdom, Brazil, India, France, South Africa, Australia and
+Japan. Failed markets are disclosed, and successful coverage is shown in the UI.
+This is an eight-market sample, not complete worldwide coverage or a post count.
+Exact normalized aliases merge first. A bounded choice evaluation merges other
+same-event aliases only at probability >= .90; related-but-different events must
+stay separate. If that service fails only exact aliases merge.
+
+For each event, the best rank per country contributes 1/sqrt(rank). Contributions
+sum across countries; the top twelve events become individual floating keywords.
+Search includes collected aliases joined by OR without a Japanese-language filter.
+Keyword placement is a visual layout, never a claimed origin. Font size uses this
+cross-market prominence plus emotion intensity. Text gradients preserve all five
+emotion weights; a neutral white word indicates analysis is not yet available.
+Clicking each keyword opens X directly, without grouping topics into a region list.
+Emotion analysis still targets up to 120 posts per event, with two events in flight.
+At twelve events the search ceiling is approximately $0.108 per refresh, plus
+$0.008 for trend feeds and separately billed analysis/alias matching. This sample
+can still be biased by source coverage, aliases and languages returned by search.
+
+If alias inference returns HTTP 401/402/403, the list explicitly reports analysis
+unavailable and no post searches are launched. The visible words remain neutral;
+exact-name aggregation and rank-based sizing continue. A live verification on
+2026-09-21 retrieved all eight countries but the inference provider returned 402,
+so live emotion color/size verification is blocked until that account is usable.
+
+## Full-sphere coverage update
+
+Discovery now samples sixteen markets, adding Mexico, Argentina, Nigeria, Kenya,
+Indonesia, South Korea, Germany and Turkey to the previous eight. All sixteen
+feeds succeeded in local verification. Feed cost ceiling is about $0.016/update;
+twelve-topic post search limits remain unchanged. This is still sampled coverage,
+not every country in the world.
+
+Keywords use an equal-area golden-angle distribution across the entire globe,
+covering both hemispheres and all longitude quadrants. They are no longer confined
+to the front/Asian hemisphere. Orbit controls slowly auto-rotate while the keyword
+view is active; users can drag to see the other side. Layout remains non-geographic.
+
+## Trend details before external navigation
+
+Wave and keyword clicks now open an accessible modal on the existing page. Only
+its explicit `Xで見る` link opens the external search in a new tab. Escape, the
+close button and `地球に戻る` dismiss the modal; auto-rotation pauses while open.
+The modal follows refreshed analysis results, showing five independent 0–100
+strengths, actual sample count, participating markets, rank, aliases and timestamp.
+Provider-supplied descriptions appear as the overview when available; missing
+summaries and missing scores are labelled rather than fabricated. Pending,
+unavailable and failed analysis states are distinguished.
+Live verification on 2026-09-22 displayed five emotion values from 120 actual
+posts for #Brownlow; the previous billing failure was not present in this check.
+
+## Unified emotion analysis (2026-09-22)
+
+Personal waves and each SNS post now use `server/analyzeEmotions.ts`: one trimmed
+`state.utterance`, the same five questions, decoder, model setting and 12-second
+upstream timeout. Empathy means the author's expressed understanding of and care
+for another person's feelings, not the reader's reaction to the author's distress.
+Region questions remain in separate batches and never share the emotion request.
+SNS analysis retains three workers, the 120-post target and the existing TTL cache.
+At 120 posts this uses 120 emotion requests plus 12 region requests instead of 12
+combined requests; latency and request volume increase. Synthetic regression
+checks do not establish accuracy on real SNS data. Quote attribution and ambiguous
+short text still need independent evaluation.
+
+## Live personal-wave preview
+
+The existing text modal previews feelings while typing, including IME text changes.
+Request starts are throttled to at least 800 ms apart, with only one request in flight.
+Completed snapshots update the wave while newer text is queued. The last wave stays
+visible during updates; clearing text invalidates outstanding snapshots. Cached
+newer results cannot be overwritten by an older in-flight result. Closing the modal cancels the request
+and clears the 20-entry in-memory cache. The UI explains that draft text is sent
+for analysis before submission; previews neither request location nor post to X.
+The preview blends five coloured moving wave layers with smooth intensity changes
+and honours reduced-motion preferences. Submission reuses a completed result only
+when its trimmed text exactly matches, otherwise it analyses the submitted text.
+Location permission is still required to place the final wave on the globe.
+
+## Topic overview and observed-country anchors
+
+When a feed has no description, `summarizeTheme` asks Jev Choice to select explanatory
+sentences from up to 120 fetched posts. This is an extractive overview, not generated
+prose or fact verification. The modal labels excerpts and links to their source posts.
+Invalid or low-confidence selections produce no overview; summary failures do not
+fail emotion analysis. Reports carry the enriched theme to the globe and its modal.
+Waves now anchor near the observed country with the best local trend rank (ties use
+country-name order), with a small visual offset. Multiple countries still form one
+global topic. The location is an observation point, not a claim about an event's origin.

cloudflare/next-headers.ts

@@ -0,0 +1,4 @@
+// NextAuth's unused App Router branch imports this on Next.js 12.
+// Workers call its Pages API handler with explicit request/response objects.
+export function headers(): never {throw new Error('App Router headers are not used by this Pages Router application');}
+export function cookies(): never {throw new Error('App Router cookies are not used by this Pages Router application');}