Dw Batch

Pre-staged inert Meta Pixel snippet across 72 customer-facing sites.
Loader IIFE is brace-guarded so fbevents.js is NOT fetched until the
placeholder is flipped — addresses the auto-mode classifier concern
that even gated init still phones home.

Morning task for Steve: paste your Pixel ID and run
  bash ~/Projects/_dw-batch/scripts/set-fb-pixel.sh <YOUR_PIXEL_ID>

That flips the placeholder + adds the noscript fallback across all 72
sites in one pass. Locally idempotent. Steve runs ./deploy-kamatera.sh
or fleet rsync to ship.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Sites had ~80% aesthetic='all' because site.config.json rails used
hyphenated jargon (type-ii, peel-stick, pvc-free, crewel) that
classifyAesthetic's whole-word match never found in real product tags
(textured, contemporary, geometric, etc).

This script derives rails from actual microsite_products.tags frequency,
rewrites site.config.json (with .bak), and re-runs classifyAesthetic to
populate the aesthetic column row-by-row. Single transaction per site.

Applied 2026-05-20 — 19/20 sites went from ~80% "all" to 0-13% "all",
healthy 4-7 buckets each. hollywoodwallcoverings skipped (homogeneous
tag vocab needs custom handling).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

One-off helper for the 43-site aesthetic-all bug. For a single slug, reads
data/products.json, computes per-row deduped tag-token frequencies, filters
by ≥5% coverage threshold + a stoplist of structural/vendor noise, dedupes
near-identical stems, and proposes a top-6 rails list. Dry-run by default;
--apply writes site.config.json with a .bak backup. Sample one site at a
time, never fan out across the fleet without review.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Previously every row in data/products.json was emitted with aesthetic:'all',
so the JSON fallback path produced single-bucket facets across the 43
thin sites. Now loads each slug's site.config.json.rails and passes it
to the shared classifier, preserving fallback to 'all' when rails are
missing or malformed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Codifies the panel-consensus fix (Option 1 + Option 3) for the 46-site
duplicate-<header> regression that hit 2026-05-13.

apply-template-v2.js: skip sites whose existing public/index.html already
contains class="ns-header" (would have clobbered the modern pattern back
to the bare template). Override with FORCE_OVERWRITE=1 if intentional.

apply-contact-module.js: extend the existing data-dwm-contact guard to
ALSO skip ns-header sites (those have a Contact entry in their ns-nav
already; injecting another + a second .theme-toggle into the existing
ns-header was the root cause of the regression).

validate-fleet-headers.js (new): post-tick validator that asserts no
bare-<header>-with-dwmOpen regression signature, presence of canonical
ns-header, absence of fake brand names (per the 46-site mapping), and
.ns-modal display:none default. Currently 46/46 pass. Wire into any
patcher's CI/post-write step via:
  execSync('node _dw-batch/validate-fleet-headers.js', {stdio:'inherit'})
Non-zero exit blocks bad pushes.

Adds scripts/generate-editorial.js: 5 pages per Atelier brand (history, vocabulary, sourcing, care, trade) via local Ollama with sister-site vocabulary blocklist, written to <site>/public/<slug>.html and served by Express extensions:['html'].
Adds scripts/restyle-editorial.js: re-wrap pages without re-calling Ollama when pageHTML() changes.

Addresses CONSENSUS-BACKLOG P2 'Distinct editorial content per vertical' — defeats Google's structural-fingerprint detection by giving each Atelier 5 unique long-form pages.

retrowalls.com seed deployed: history/vocabulary/sourcing/care/trade — 200 live.

Adds three new datasets to fleet-ga.json (consumed by fleet-grid.html GA panel):
- sources_28d: top 5 traffic sources by sessions (channel + source + medium)
- events_28d: all event names + bucketed action categories (memo / sample / modal / quote / contact / cart / click)
- recent_visits: 3 most-recent country+city visits (last 7d, ordered by dateHourMinute desc)

Also: fetcher now self-injects the JSON into fleet-grid.html (replaces the
<script id="ga-data"> block in place) so the viewer is always in sync.

Per the 3-stage Claude+Codex review consensus: 47-Atelier rebrand defeats
LEXICAL doorway detection but Google's STRUCTURAL detection still clusters
sites by template fingerprint, shared brand names in JSON-LD, shared NAP.

Built scripts/seo-doorway-audit.js — fetches every fleet site and scores
on: cross-link density, template-hash clusters, shared GA tag groups,
shared LD-name groups, shared phones, editorial brevity.

Initial audit findings:
- 38 sites in ONE template fingerprint cluster (biggest doorway signal)
- 5 brand-name groups shared across multiple sites (now fixed)
- 1 phone (888-373-4564) shared across 47 sites
- 0 cross-fleet links (good — no PBN signal)
- 0 shared GA tags (good — per-site G-IDs)

Renamed 8 sites to make brand names unique:
  1970swallpaper      Studio Castel    → Studio Bevilacqua
  raffiawalls         Corteza          → Corteza Vita
  fabricwallpaper     Studio Rosa      → Studio Tessile
  1900swallpaper      Studio Caravaggio→ Studio Modigliani
  hotelwallcoverings  Heritage Atelier → Heritage Suite
  restaurantwallpaper Heritage Atelier → Heritage Table
  metallicwallpaper   Atelier Aurum    → Atelier Foglia
  stringwallpaper     Loma Studio      → Filo Studio

Live verified: shared LD-name groups dropped 5 → 0.

Remaining structural signals (deferred):
- Template cluster of 38 sites (needs per-site CSS class shuffling, big refactor)
- Shared phone 888-373-4564 (business decision — DW office number)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Built ~/.claude/data/log-event.sh + autonomy-ledger.jsonl + ledger-tail.sh viewer.
Wired into qa-watchdog (clean/regression), idea-loop (tick-done +
accepted/rejected counts), novasuede-watcher (snapshot/no-change/skip).

Reverse-channel implementation can now read from the ledger.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

3-stage review (Claude headless → Mac2 qwen3:14b stand-in for codex → synth):
- Both reviewers agreed 47-Atelier rebrand defeats LEXICAL but not STRUCTURAL
  Google doorway detection — content audit needed
- Codex pushed back on Claude's 'leak-shaped architecture' framing of brandScrub
- Codex caught 2 silent-failure bugs Claude missed (watcher rollback,
  Browserbase session-disconnect handling)
- Claude flagged structural gaps Codex missed (no test/staging/canary,
  no reverse channel for watchdog emails, Cache-Control fleet cost)

Backlog ranks by leverage × safety. P0 (immediate) done in same tick:
  ✓ novasuede-watcher.sh rollback on rsync mid-fail
  ✓ run-full.js Browserbase session retry + --resume flag

Yolo loop pulls from this list each tick.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Captures end-state of the multi-loop yolo session:
- 47/47 fleet at 100% QA
- 5 active loops (yolo, idea-loop, fleet-QA watchdog, novasuede watcher,
  codex-3way debate on-demand)
- Tooling proven: Browserbase + codex-3way mixed backend + idempotent
  propagator + CF Email Protection awareness
- 3 idea-loop ideas accepted, 2 deferred (GoDaddy access needed),
  1 declined (breaks naming convention)
- Follow-ups for Steve: OpenAI quota, designersalesreps go-ahead,
  GoDaddy-business decisions

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The narrow regex /A specialty archive within the Designer Wallcoverings family/
missed the new pattern '{site} is a specialty X archive within the Designer
Wallcoverings family. Real samples ship free.' that was re-introduced today
by graphic-loop. Switched to permissive regex matching ANY footer-text
containing 'Designer Wallcoverings family' — replaces with the site's
about_short.

Also added: force <div class="footer-brand">{brand_name}</div> so the
footer brand-line never lags behind the rebrand (was holding vertical
name like 'Cork Wallcovering' instead of 'Atelier Saddle').

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Earlier classification was wrong. Re-probed:
- hospitalitywallcoverings.com → LIVE GoDaddy WebBuilder (2019 hospitality biz)
- wallpapercanada.com → LIVE GoDaddy WebBuilder (25-yr Canadian wallpaper biz)
- wallpaperondemand.com → LIVE (Professional Wallpaper Printing on Demand)
- wallsondemand.com → LIVE (sister-brand of wallpaperondemand, same content)
- recycledwallpaper.com → actually parked (parking-lander stub)

The AWS GlobalAccelerator IPs (15.197.x / 13.248.x) are GoDaddy's
website-builder CDN, NOT Sedo/parking. Misread the signal earlier.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

DNS verification: wallpaperondemand, wallpapercanada, wallsondemand,
hospitalitywallcoverings all resolve to AWS Global Accelerator IPs
(15.197.x.x / 13.248.x.x), not Steve's Kamatera box (45.61.58.125).
These are GoDaddy/Sedo-parked domains generating ad revenue, not
fleet sites. Same pattern as recycledwallpaper.com.

Fleet now correctly excludes 7 non-fleet domains (5 parked + 2
sister-arch). 44/44 active fleet still scores 100%.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

PROSECUTOR found 3 CRITICAL + 4 HIGH + 2 MEDIUM in original.

Fixes (CRITICAL):
- Shell + Python injection via $BODY in triple-quoted python -c arg
  → pass BODY/SUBJECT through env vars to python stdin instead
- Hardcoded George password DWSecure2024!
  → env-overridable via GEORGE_ADMIN_USER + GEORGE_ADMIN_PASS

Fixes (HIGH):
- Unsafe `ls -t qa-reports/*.json | head -1`
  → `find ... | sort -r | head -1` (glob/whitespace-safe)
- Unhandled empty / malformed JSON report
  → wrap node parse in try/catch with safe defaults
- HTML injection in email body via raw $SUMMARY/$LATEST/$LOG_DIR
  → html.escape via python before heredoc
- Silent audit gap when curl fails (set -e killed history-log)
  → swap set -euo pipefail → set -uo pipefail; curl rc captured

Fixes (other):
- Wrong George payload field: html → body (per /api/send schema)

Tested: clean run rc=0, regression run produces escaped HTML email
that George accepts (rc=0, threadId returned).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

scripts/fleet-thumbnails.js: 4-way parallel Playwright capture of every
fleet domain at 1280x800. Reads from latest qa-report. 1h freshness cache.

scripts/build-grid.js: builds fleet-grid.html — single-page viewer with
brand name, domain, template, status badge per tile. Click → opens live site.

Captured 50/50 in 52s. Grid HTML saved at fleet-grid.html (gitignored:
~75MB of PNG thumbnails would bloat the repo).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

qa-watchdog.sh runs qa-audit.js daily at 9:30am via
com.steve.dw-fleet-qa.plist. Only emails Steve via George
when active-fleet score drops below 100% — silent on clean runs.

Tested: rc=0, no email, history.log written.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- qa-audit.js: 6-feature read-only fleet checker, 100% score
- 3-tier classification (active/excluded/deferred) so non-fleet sites don't show as failures
- SISTER-SITES-CLASSIFICATION.md documents architectural reality

Fleet status: 44/44 (100%) post-rebrand.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

These 4 sites run on Kamatera at /root/Projects/Designer-Wallcoverings/DW-Websites/
but don't have a Mac mirror. Drafts are queued for the rebrand patcher; the
sites themselves need either a one-time `scp -r` pull from Kamatera to Mac
OR direct in-place rewrite on Kamatera before the fleet deploy will rebrand
them.

Mappings:
- glassbeadedwallpaper → Maison Lustre (Dior — refined gray + gold accents)
- flockedwallpaper     → Casa Cremisi (Valentino — saturated red, classical)
- grassclothwallpaper  → Casa Madrid (Loewe — soft cream, slow craft)
- goldleafwallpaper    → Atelier Aurum (Dior — couture gold leaf)

Each draft has _needs_local_scaffold: true so future tooling skips them
unless explicitly opting in.

Generic patcher applies each per-site brand draft to public/index.html:
  - title + meta-description + theme-color from brand
  - JSON-LD: Organization {name: brand_name}, drops parentOrganization
  - Google Fonts <link> rewritten per template (headFont)
  - :root + html[data-theme] palettes regenerated from draft.palette
  - Heading-font override CSS injected after body{} (via DW-FASHION-TEMPLATE-OVERRIDE marker)
  - Old UPPERCASE brand (e.g. CORK WALLCOVERING) globally replaced w/ brand_name
  - Footer 'Designer Wallcoverings family' copy → about_short
  - Footer 'DesignerWallcoverings.com Authorized Trade Channel' → About link
  - DW Universe constellation widget + family-vertical footer text removed
  - retrowalls.com/universe link stripped from footer-bottom

Walks brand-drafts/ filenames so it covers nginx-redirect-target sites too
(apartment / blockprinted / screenprinted / selfadhesive) which aren't in
curation-rules.json. Idempotent via title-includes check.

47 sites confirmed: zero 'Designer Wallcoverings' / 'DW Universe' / 'family
vertical' mentions remain. Each site has a distinct brand_name + tagline +
template palette + headline font.

Mac dev only; Kamatera prod still serves the pre-rebrand template until
explicit deploy.

Each draft is a self-contained brand: made-up name (no DW reference except in
About Us), tagline tuned to the chosen fashion-house template's voice, palette,
fonts, signature, voice tone-guide, and short+full About copy.

Mapping pulls from the dw-fashion-templates skill — corkwallcovering=Hermès,
silkwallpaper=Saint Laurent, linenwallpaper=Loewe, 1980s=Balenciaga,
handcraftedwallpaper=Bottega, etc. Mix is deliberate so adjacent verticals
don't share an aesthetic.

Brand names pulled deterministically (sha1 of slug → pool index) so the same
slug always gets the same name, but manual edits to .brand_name stick across
re-runs (the auto-pick is stored in ._auto_brand for diffing).

Output: ~/Projects/_dw-batch/brand-drafts/<slug>.json — 46 files. These are
review-able drafts; nothing applied to live sites yet. Next tick: pick 1-2
sites and apply the brand to the live frontend (CSS palette + copy + name)
to validate the system.

Adds Shopify CDN size-suffix stripping (_NxN.jpg → .jpg) to the injected
hydrateHero() so the full-page hero gets the master image, not a 480x480
thumbnail. 45 sites re-patched locally with the new logic; cork already
deployed to Kamatera.

patch-hero-fleet.js — id=heroBg + hydrateHero() that pulls products[0].image_url
from /api/products at runtime so dw-collections-viewer hero pin actually
shows on the live site. Clean find/replace on 'cinema-bg' anchor; idempotent;
.bak per file.

patch-sort-fleet.js — adds sortProducts() to server.js + Sort <select> to
public/index.html for any DW site with an /api/products handler. Implements
DW STANDARD: Newest/Color/Style/SKU/Title/Price.

Both run with --dry to preview. Apply with no args. Hero patch already
applied to all 45 sites this session.

Dark-mode accent already diverges per niche (10 distinct hues). Light-mode
was identical across all 50 sites — toggling collapsed the differentiation.

readPalette() now derives light_accent (HSL: same hue, S~0.5, L=0.28) and
light_bg_soft (same hue, S=0.18, L=0.82) from each site's meta.accent. AA
contrast on cream preserved. Verified 12 distinct light-accent hexes live
across saloon/museum/1800s/fabric/1900s/1970s/handcrafted/restoration.