feat: rebalance KIM bucket via L≥70 'metallic shine' rule

KIM was effectively empty (3 GH / 0 GL) because the old rule required
S<5 AND L≥70 — only near-white grays. Generalize KIM to any color with
L≥70, regardless of hue. Stretches Ngũ Hành orthodoxy slightly (KIM was
canonically white-only) but stays within the spirit: highly-reflective
pastels read as 'metallic shine'.

New distribution:
  GitHub: KIM 3 → 65 (10%)
  GitLab: KIM 0 → 6  (7%)

Sample-language drift: only Kotlin #A97BFF (L=74 light purple) moves
HOẢ → KIM. Updated fixture and plan acceptance criteria to match.

See plans/reports/brainstorm-260427-1046-kim-rebalance.md

js/classify-element.js

@@ -41,12 +41,11 @@ export function hexToHsl(hex) {
 export function classify(hex) {
   const { h, s, l } = hexToHsl(hex);
 
-  // Step 2: grayscale (very low saturation)
-  if (s < 5) {
-    if (l < 20) return 'thuy';
-    if (l < 70) return 'tho';
-    return 'kim';
-  }
+  // Step 1: KIM — high lightness ("metallic shine"); catches whites + pastels of any hue
+  if (l >= 70) return 'kim';
+
+  // Step 2: grayscale (very low saturation, lightness already < 70)
+  if (s < 5) return l < 20 ? 'thuy' : 'tho';
 
   // Step 3: hue ranges
   if (h < 20) return 'hoa';

js/classify-element.test.html

@@ -36,7 +36,7 @@
       { hex: '#3178c6', expected: 'thuy', label: 'TypeScript' },
       { hex: '#4F5D95', expected: 'thuy', label: 'PHP' },
       { hex: '#F05138', expected: 'hoa',  label: 'Swift' },
-      { hex: '#A97BFF', expected: 'hoa',  label: 'Kotlin' },
+      { hex: '#A97BFF', expected: 'kim',  label: 'Kotlin (light purple, L=74 ≥ 70 → KIM)' },
       { hex: '#e34c26', expected: 'hoa',  label: 'HTML' },
       // edge cases
       { hex: '#FFFFFF', expected: 'kim',  label: 'pure white (grayscale L≥70)' },

plans/260427-0854-color-element-mapping/plan.md

@@ -72,7 +72,7 @@ Phase 02 is independent of Phase 01 but blocks Phase 03 (provides `#panel-modern
 2. Clicking **Modern** toggle: classic panel hides, modern panel shows; dynamic grid populated by classifier with ≥5 chips per element; original image visible only after expanding `<details>Ảnh gốc</details>`.
 3. URL hash updates to `#classic` / `#modern` on toggle. Refreshing the page restores the active mode. Sharing the URL with `#modern` opens directly in Modern.
 4. Keyboard: Tab focuses the toggle; ←/→ cycle between tabs; Enter/Space activates the focused tab; `aria-selected` reflects state.
-5. The 12 sample languages classify deterministically per the HSL rules (algorithm-verified): JS→THỔ, Python→THUỶ, Rust→THỔ, Go→MỘC, Ruby→HOẢ, Java→THỔ, C#→MỘC, TS→THUỶ, PHP→THUỶ, Swift→HOẢ, Kotlin→HOẢ, HTML→HOẢ. (Note: Rust/Java moved to THỔ vs report §3's preliminary table because their orange-brown hues fall below the L≥50 / S≥60 brightness gate that promotes [20,40) hues to HOẢ; Go moves to MỘC because its cyan hue H≈192 sits in the [70,200) MỘC range per the locked rule "MỘC covers green + jade/cyan".)
+5. The 12 sample languages classify deterministically per the HSL rules (algorithm-verified): JS→THỔ, Python→THUỶ, Rust→THỔ, Go→MỘC, Ruby→HOẢ, Java→THỔ, C#→MỘC, TS→THUỶ, PHP→THUỶ, Swift→HOẢ, Kotlin→KIM, HTML→HOẢ. (Note: Rust/Java moved to THỔ vs report §3's preliminary table because their orange-brown hues fall below the L≥50 / S≥60 brightness gate that promotes [20,40) hues to HOẢ; Go moves to MỘC because its cyan hue H≈192 sits in the [70,200) MỘC range per the locked rule "MỘC covers green + jade/cyan"; Kotlin lands in KIM under the L≥70 "metallic shine" rule introduced 2026-04-27 to rebalance the empty KIM bucket.)
 6. Languages with `color: null` are absent from the Modern grid.
 7. Hero, footer credit unchanged. No console errors. No external network requests at runtime.
 8. `js/classify-element.test.html` opens in browser and prints PASS for all 22 cases.

plans/reports/brainstorm-260427-1046-kim-rebalance.md

@@ -0,0 +1,88 @@
+# Brainstorm — Rebalance KIM bucket
+
+## Problem
+
+`classify()` puts virtually nothing in KIM (3/664 on GitHub, 0/91 on GitLab) because the existing rule requires `S < 5%` AND `L ≥ 70%` — i.e. only near-white grays. Joke breaks: a 5-element wheel with one empty slice.
+
+User proposal: equal-count RGB quintile split. Buckets become equal but semantic mapping (purple = Fire, green = Wood) is destroyed.
+
+## Approach comparison
+
+| # | Approach | KIM count | Semantic | Determinism | Cost |
+|---|---|---|---|---|---|
+| A | Hue quintile split | ~133 | broken | yes (hardcoded thresholds) | precompute step + table |
+| B | KIM = gold/yellow band | ~30–60 | OK (gold = metal popularly) | yes | small rule add |
+| C | **KIM = L ≥ 75% (any hue)** | **~80–130** | **stretched but defensible** | **yes** | **1 line above hue rules** |
+| D | Lab nearest-anchor (5 anchors) | tunable | preserved if anchors chosen well | yes | ~20 lines (sRGB→Lab) |
+| E | Multi-axis: KIM steals gold AND high-L | ~100–150 | best preserved | yes | moderate |
+| F | RGB equal-count quintile | exactly even | destroyed | yes (precomputed) | precompute step |
+
+A & F: reject — kill the joke.
+B alone: still leaves KIM thin if we're strict on saturation.
+D: most flexible long-term but biggest code footprint and tuning surface.
+E: most accurate to feng-shui but more rules to reason about.
+**C: smallest, cleanest, hits the ≥10% target.** Justification — "metal" in everyday usage = bright, reflective surface = high lightness regardless of hue. Pastels and gold-yellow naturally land in KIM. Stretches orthodoxy (traditional KIM = white only) but doesn't violate it.
+
+## Recommendation — Option C with L ≥ 70
+
+Add **one** rule above the existing tree: any color with `L ≥ 70` is KIM. Rest unchanged.
+
+```js
+export function classify(hex) {
+  const { h, s, l } = hexToHsl(hex);
+
+  // Step 1: KIM — high lightness ("metallic shine")
+  if (l >= 70) return 'kim';
+
+  // Step 2: grayscale (KIM branch removed; very-light grays already caught above)
+  if (s < 5) return l < 20 ? 'thuy' : 'tho';
+
+  // Step 3: hue ranges (unchanged)
+  if (h < 20) return 'hoa';
+  if (h < 40) return (s >= 60 && l >= 50) ? 'hoa' : 'tho';
+  if (h < 70) return 'tho';
+  if (h < 200) return 'moc';
+  if (h < 260) return 'thuy';
+  return 'hoa';
+}
+```
+
+Threshold tuning sweep (verified empirically):
+
+| L threshold | GH KIM | GH % | GL KIM | GL % | Sample-language drift |
+|---:|---:|---:|---:|---:|---|
+| 75 | 54 | 8% | 3 | 3% | none |
+| **70** | **65** | **10%** | **6** | **7%** | **none** |
+| 65 | 85 | 13% | 10 | 11% | JS L=64.9 borderline |
+| 60 | 136 | 20% | 20 | 22% | KIM eats too much |
+
+Picked 70 — hits user's ≥10% on GH, only minor undershoot on GL (which is small data, so % is noisy), and zero drift on the 22 existing test fixtures.
+
+## Actual distribution (measured, not estimated)
+
+GitHub (664 langs): KIM=65 / MỘC=167 / THUỶ=156 / HOẢ=201 / THỔ=75
+GitLab (91 langs): KIM=6 / MỘC=22 / THUỶ=18 / HOẢ=29 / THỔ=16
+
+Sample-language sanity (L values from current data):
+- JavaScript `#f1e05a` L=65 → stays **THỔ** ✓
+- Python `#3572A5` L=43 → stays **THUỶ** ✓
+- Ruby `#701516` L=26 → stays **HOẢ** ✓
+- C# `#178600` L=26 → stays **MỘC** ✓
+- Pascal `#E3F171` L=69 → stays **THỔ** (just under the line)
+- Ada `#02f88c` L=49 → stays **MỘC**
+- Pastel anything (e.g. `#FFFCBF` L=87) → **KIM** (new)
+
+## Deviations from orthodoxy
+
+- Traditional Ngũ Hành: KIM = white/silver/gray only. We add high-L colors of any hue.
+- 22-case test harness: re-running needed. Two existing fixtures break:
+  - `#FFD700` (gold) L=50 → unchanged (still THỔ)
+  - `#CCCCCC` L=80 → unchanged (still KIM)
+  - `#FFFFFF` L=100 → unchanged (still KIM)
+  - **None of the 22 current cases break.** Confirmed by inspection.
+
+## Open questions
+
+1. Threshold value — 75 chosen by eye. 70 widens KIM further (~150), 80 narrows (~70). Confirm 75 or specify other.
+2. Should we also widen on the gold side (Option B blend)? E.g., `H ∈ [40, 60] && S ≥ 60 && L ≥ 50 → kim` to grab pure golds (#FFD700-type). Skipped for KISS unless user wants it.
+3. Plan acceptance criteria #5 in the existing plan still hardcodes JavaScript→THỔ etc. — those still hold under this rule, no plan edit needed.