Convert all Figma theme backgrounds to dark/colored, dedupe again

The 34 kept themes were split 24 light-mode / 18 dark-mode - the light
ones all shared the same washed-out near-white background (L=0.95),
too bright per feedback. Added a force_mood override to build_theme()
and regenerated all 24 as dark instead, same accent character, colored-
dark background rather than near-white.

That conversion collapsed a lot of previously-distinct light backgrounds
into similar dark neutrals, so re-ran the same color-distance dedup pass
across the full set: 8 more turned out to be near-twins once everything
converged to dark (including a 3-way orange-on-dark cluster trimmed to
one). Net: 53 candidates -> 34 kept, all dark/colored, zero white
backgrounds - down from the 42 (24 light/18 dark) shipped last commit.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

tools/generate_website_scheme_themes.py

@@ -1,5 +1,5 @@
 """
-One-off tool that generated the 53 "website scheme" theme entries in
+One-off tool behind the 34 "website scheme" theme entries in
 THEME_PALETTES/THEME_DISPLAY_NAMES (offlineu_core.py), sourced from Figma's
 "53 Unique Website Color Schemes" resource page:
 https://www.figma.com/resource-library/website-color-schemes/
@@ -8,22 +8,36 @@ That page has no raw hex data - each scheme is a rendered mockup
 screenshot, not a swatch grid - so website_scheme_swatches.json (checked in
 alongside this script) holds dominant colors already extracted from those
 53 images via canvas pixel-histogram sampling in a browser, one entry per
-scheme: {n, name, colors: [{hex, pct}, ...]}. This script turns that raw
+scheme: {n, name, colors: [{hex, pct}, ...]}. build_theme() turns that raw
 material into an actual UI palette per scheme (bg/text/accent, matching
-THEME_PALETTES' shape), then nudges lightness as needed so every result
-clears the same contrast bars this file's checks apply: text-vs-background,
-accent-vs-background, and accent-vs-white (buttons always use white text -
-see .btn in course_dashboard.html - so a too-bright accent needs catching
-even when it reads fine against the background alone).
+THEME_PALETTES' shape), then nudges lightness/saturation as needed so every
+result clears the same contrast bars this file's checks apply:
+text-vs-background, accent-vs-background, and accent-vs-white (buttons
+always use white text - see .btn in course_dashboard.html - so a too-bright
+accent needs catching even when it reads fine against the background
+alone), and so accents don't land at the full saturation that reads as
+neon in the mid-lightness band.
 
-Not a live pipeline - re-running it regenerates the exact same 53 themes
-from the same frozen swatch data. To add more schemes, extract their
-dominant colors the same way (see the canvas-sampling approach used
-in-session; not scripted here) and append to the JSON, then rerun and
-splice the output into offlineu_core.py by hand.
+Running this file directly (see __main__ below) auto-detects each scheme's
+own mood from its swatches and writes all 53 as candidates - that's NOT
+what's actually shipped. The source material is mostly light-mode
+marketing mockups, so the first pass over-represented near-white
+backgrounds; the shipped 34 all use build_theme(entry, force_mood='dark')
+instead (every one of the 53 converted to a dark/colored background), then
+had duplicates pruned by measuring real color distance (hue + lightness +
+saturation, background and accent both) between every pair, tight enough
+to only catch genuine near-twins - not something this script's __main__
+does for you. That curation was one-off analysis, not captured as a single
+rerunnable command; to redo it, call build_theme(entry, force_mood='dark')
+per scheme, then de-duplicate the results the same way before splicing
+into offlineu_core.py by hand.
+
+To add more schemes: extract their dominant colors the same way (see the
+canvas-sampling approach used originally; not scripted here) and append to
+the JSON, then rerun.
 
 Usage: python3 generate_website_scheme_themes.py
-Writes generated_themes.json with the full computed palette per scheme.
+Writes generated_themes.json with all 53 candidates, auto-detected mood.
 """
 import colorsys
 import json
@@ -93,7 +107,7 @@ def tame_accent_saturation(s, l):
     return min(s, 0.85)
 
 
-def build_theme(entry):
+def build_theme(entry, force_mood=None):
     name = entry['name']
     swatches = entry['colors']
     parsed = []
@@ -107,8 +121,13 @@ def build_theme(entry):
     if not content:
         content = parsed
 
-    darkest_l = min(p['l'] for p in content)
-    is_dark = darkest_l < 0.28
+    if force_mood == 'dark':
+        is_dark = True
+    elif force_mood == 'light':
+        is_dark = False
+    else:
+        darkest_l = min(p['l'] for p in content)
+        is_dark = darkest_l < 0.28
 
     accent_candidates = [p for p in content if p['s'] > 0.25]
     if accent_candidates: