feat(tools): theme design audit + item 28 punch list (slice a)

tools/audit-theme-design.py (report-only, maintainer-run): for every
themes/*.json it checks OKLCH lightness architecture (raised bg stack,
fg hierarchy), WCAG contrast beyond the guarded pairs, status-color
hue families, accent harmony, protanopia/deuteranopia
distinguishability (Machado matrices), and ANSI slot semantics the
greeter/tty now rely on. Ran it over all 21 palettes + a visual pass
over representative previews; findings curated identity-aware (lumon/
vantablack/white/hackerman exemptions are identity, not defects) into
the P1-P3 punch list now recorded in BACKLOG item 28: two palettes
with literally invisible subtext (summer-day, flexoki-light — the
item-25/27 trap at its source), two inverted bg stacks (miasma,
kanagawa), muted/warn floors, a systemic "status is never color-only"
rule, the light-theme ANSI convention decision, and bar grouping
rhythm notes.

Verified: V0 (py_compile; the tool is report-only, deliberately not a
checks.* gate until slice b adopts rules + exemptions).

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

agent/BACKLOG.md

@@ -49,14 +49,61 @@ Bernardo has offered image-expert models for designed graphics
 (wallpapers, logos, glyphs, preview/branding art): an agent slice that
 needs artwork must STOP and ask ("say the word") instead of shipping
 programmer-art.
-Slices: (a) the audit — render every theme (VM themed-screenshot
+Slices: (a) ✓ the audit (2026-07-05, iteration #31 — findings below);
-recipe or real captures), review against the rules above, append a
+(b) palette-level fixes + the extended contrast/harmony checker;
-per-surface punch list to this item; (b) palette-level fixes + the
+(c) per-surface polish slices;
-extended contrast/harmony checker; (c) per-surface polish slices;
 (d) the whole-swap themes brought to the same bar. Visual work is
 V3-bound by nature — screenshots are the review artifact, hardware
 confirms.
 
+**Audit findings (slice a — `tools/audit-theme-design.py` over all 21
+palettes + visual pass over previews; identity-aware curation):**
+- **P1 · invisible-secondary-text palettes:** summer-day and
+  flexoki-light have `subtext == base` (contrast 1.0 — literally
+  invisible) and `text == surface`. Items 25/27 fixed the *consumers*
+  (swaync/bar stopped using them); the palettes are still traps — any
+  new surface consuming subtext/surface re-trips. Fix the palettes:
+  real mid-tone subtext, raised surface.
+- **P1 · inverted bg stacks:** miasma's surface is #000000 under a
+  much lighter base; kanagawa's surface (L .14) sits below base
+  (L .24). "Raised" surfaces render sunken/black. Retune.
+- **P2 · muted floor:** muted-on-base < 2.0 on gruvbox (1.27),
+  everforest (1.55), lumon (1.68), nord (1.69), white (1.82), latte
+  (1.91), retro-82 (1.93) — dimmed workspaces/inactive items barely
+  render (visible on the lumon preview). Adopt ≥ 2.0 and retune.
+- **P2 · warn barely reads:** warn-on-base ≈ 2.0–2.3 on rose-pine,
+  summer-day, latte, flexoki-light — this is the bar's 25%-battery
+  color. Floor 2.5 (or see the color-only rule below).
+- **P2 · text-on-surface < 4.5:** ristretto 3.88, latte 4.39 (menus
+  and chips draw text on surface).
+- **P3 · accentAlt-on-base < 3:** latte 2.34, summer-day 2.83.
+- **Design rule (systemic, beats palette surgery): status is never
+  color-only.** good/warn collapse under protanopia on ~half the
+  themes and good/bad under deuteranopia (59 pair-findings) — that's
+  inherent to red/green/yellow, not fixable per-palette. Battery and
+  recording already carry glyph changes; sweep the remaining
+  color-only indicators (updates count, notification dot, VPN shield
+  states) for shape/glyph redundancy.
+- **Identity exemptions, made explicit:** vantablack + white
+  (deliberately hueless statuses), lumon (all-blue), hackerman
+  (matrix green statuses), matte-black/retro-82 (off-family hues) are
+  identity, not defects — slice (b)'s checker gets a per-theme
+  exemption list instead of silently lowered floors.
+- **ANSI / greeter notes:** latte, rose-pine and white ship
+  `ansi[0]` LIGHTER than `ansi[15]` — decide the light-theme ANSI
+  convention (terminals conventionally keep black dark) before the
+  greeter/tty theming ships to them. tokyo-night/summer-night
+  `ansi[7]-on-ansi[0]` ≈ 2.8–3.0 = the greeter's muted text; fine as
+  muted, watch on hardware.
+- **Visual pass (480×270 previews — too small for type/spacing; the
+  deep per-surface pass needs 1080p captures):** the generated bar's
+  right cluster runs ~9 modules with uniform spacing — no grouping
+  rhythm (consider micro-gaps between tray/net · stats · toggles ·
+  battery/notify groups); latte's preview shows a DARK terminal on
+  the light desktop — verify ghostty consumes the latte palette (or
+  the preview is stale); summer-day's island-pill whole-swap bar and
+  lumon's globe identity are strong — preserve through any polish.
+
 ### 14. Automated lock bumps — confirm the first real run
 Slices b+c shipped 2026-07-05 (`.gitea/workflows/bump.yml`): weekly
 schedule (Mon 05:17 UTC) + `workflow_dispatch` as the security

agent/JOURNAL.md

@@ -17,6 +17,30 @@ Template:
 
 ---
 
+## 2026-07-05 — theme design audit (iteration #31, item 28 slice a)
+- **Task:** BACKLOG NEXT#28 slice (a) — the design-theory audit
+  (human-requested; Bernardo asked to run it now).
+- **Did:** tools/audit-theme-design.py (new, report-only): OKLCH
+  lightness architecture, extended WCAG pairs, status-hue families,
+  accent harmony, Machado protan/deutan simulation, ANSI slot
+  semantics — over all 21 palettes; plus a visual pass over 4
+  representative preview.png captures. Findings curated
+  identity-aware into a P1–P3 punch list appended to item 28: 2
+  invisible-subtext palettes (summer-day, flexoki-light), 2 inverted
+  bg stacks (miasma, kanagawa), 7 muted-floor misses, warn-barely-
+  reads ×4, a systemic "status is never color-only" rule (CVD
+  collapse is inherent, glyph redundancy is the fix), explicit
+  identity exemptions, light-theme ANSI convention decision, bar
+  right-cluster grouping rhythm, latte dark-terminal verify.
+- **Verified:** V0 (py_compile; report tool, not wired into checks).
+  Analysis is palette-math + thumbnail previews; per-surface
+  spacing/type audit needs 1080p captures (slice c prep).
+- **Pending:** slice (b) — palette retunes + checker extension w/
+  exemption list; needs Bernardo's eye on the P1 retunes (visual
+  taste calls). Artwork needs none so far — no image-model ask yet.
+- **Next suggestion:** slice (b) P1s (summer-day/flexoki subtext+
+  surface, miasma/kanagawa stacks) — small hex edits, big payoff.
+
 ## 2026-07-05 — greeter theming from the JSON (iteration #30, item 16)
 - **Task:** BACKLOG NEXT#16 — tuigreet themed from theme-state.json at
   system rebuild (the Plymouth model). Item deleted.

tools/audit-theme-design.py

@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+"""Item 28 slice (a): design-theory audit of every themes/*.json.
+
+Reports (never fails): OKLCH lightness architecture, extended WCAG
+contrast pairs, hue-family sanity for status colors, accent harmony,
+CVD (protanopia/deuteranopia) distinguishability, ANSI slot semantics.
+"""
+import json, math, sys
+from pathlib import Path
+
+# ── color math ───────────────────────────────────────────────────────
+def srgb_to_linear(c):
+    c /= 255.0
+    return c / 12.92 if c <= 0.04045 else ((c + 0.055) / 1.055) ** 2.4
+
+def hex_rgb(h):
+    h = h.lstrip("#")
+    return tuple(int(h[i:i+2], 16) for i in (0, 2, 4))
+
+def hex_lin(h):
+    return tuple(srgb_to_linear(c) for c in hex_rgb(h))
+
+def oklab(h):
+    r, g, b = hex_lin(h)
+    l = 0.4122214708*r + 0.5363325363*g + 0.0514459929*b
+    m = 0.2119034982*r + 0.6806995451*g + 0.1073969566*b
+    s = 0.0883024619*r + 0.2817188376*g + 0.6299787005*b
+    l, m, s = l ** (1/3), m ** (1/3), s ** (1/3)
+    return (0.2104542553*l + 0.7936177850*m - 0.0040720468*s,
+            1.9779984951*l - 2.4285922050*m + 0.4505937099*s,
+            0.0259040371*l + 0.7827717662*m - 0.8086757660*s)
+
+def oklch(h):
+    L, a, b = oklab(h)
+    C = math.hypot(a, b)
+    H = math.degrees(math.atan2(b, a)) % 360
+    return L, C, H
+
+def de(h1, h2):  # OKLab euclidean distance
+    a, b = oklab(h1), oklab(h2)
+    return math.dist(a, b)
+
+def wcag_lum(h):
+    r, g, b = hex_lin(h)
+    return 0.2126*r + 0.7152*g + 0.0722*b
+
+def ratio(f, b):
+    hi, lo = sorted((wcag_lum(f), wcag_lum(b)), reverse=True)
+    return (hi + 0.05) / (lo + 0.05)
+
+# Machado et al. 2009, severity 1.0, applied in linear RGB.
+PROTAN = [(0.152286, 1.052583, -0.204868), (0.114503, 0.786281, 0.099216), (-0.003882, -0.048116, 1.051998)]
+DEUTAN = [(0.367322, 0.860646, -0.227968), (0.280085, 0.672501, 0.047413), (-0.011820, 0.042940, 0.968881)]
+
+def cvd_hex(h, M):
+    r, g, b = hex_lin(h)
+    sim = [max(0.0, min(1.0, M[i][0]*r + M[i][1]*g + M[i][2]*b)) for i in range(3)]
+    def enc(c):
+        c = 12.92*c if c <= 0.0031308 else 1.055 * c ** (1/2.4) - 0.055
+        return max(0, min(255, round(c*255)))
+    return "#%02x%02x%02x" % tuple(enc(c) for c in sim)
+
+def hue_in(H, lo, hi):
+    return lo <= H <= hi if lo <= hi else (H >= lo or H <= hi)
+
+ROLES = ["base","mantle","surface","overlay","text","subtext","muted","accent","accentAlt","good","warn","bad"]
+
+findings = {}
+def note(slug, cat, msg):
+    findings.setdefault(slug, []).append((cat, msg))
+
+themes = sorted(Path(sys.argv[1] if len(sys.argv) > 1 else "themes").glob("*.json"))
+for tf in themes:
+    t = json.loads(tf.read_text()); slug = tf.stem
+    c = t["colors"]; mode = t.get("mode", "dark"); dark = mode == "dark"
+    L = {r: oklch(c[r])[0] for r in ROLES}
+    C = {r: oklch(c[r])[1] for r in ROLES}
+    H = {r: oklch(c[r])[2] for r in ROLES}
+
+    # A. lightness architecture: bg stack monotonic away from base
+    stack = ["base", "surface", "overlay"]
+    diffs = [L[b] - L[a] for a, b in zip(stack, stack[1:])]
+    want = 1 if dark else -1
+    for (a, b), d in zip(zip(stack, stack[1:]), diffs):
+        if d * want < 0.005:
+            note(slug, "lightness", f"bg stack not raised: L({b})={L[b]:.3f} vs L({a})={L[a]:.3f} ({mode})")
+    # fg stack: muted < subtext < text in |L - L(base)|
+    dist = {r: abs(L[r] - L["base"]) for r in ("muted","subtext","text")}
+    if not (dist["muted"] <= dist["subtext"] + 0.005 and dist["subtext"] <= dist["text"] + 0.005):
+        note(slug, "lightness", f"fg hierarchy off: |dL| muted={dist['muted']:.2f} subtext={dist['subtext']:.2f} text={dist['text']:.2f}")
+
+    # B. contrast pairs (audit thresholds)
+    for f, b, mn, why in [
+        ("text","base",4.5,"body text"), ("text","surface",4.5,"text on chips/menus"),
+        ("subtext","base",3.0,"secondary text"), ("accent","base",3.0,"indicators"),
+        ("accentAlt","base",3.0,"alt accent"), ("base","accent",3.0,"selected rows"),
+        ("good","base",2.5,"status glyph"), ("warn","base",2.5,"status glyph"),
+        ("bad","base",2.5,"status glyph"), ("muted","base",2.0,"dimmed text"),
+    ]:
+        r = ratio(c[f], c[b])
+        if r < mn:
+            note(slug, "contrast", f"{f} on {b} = {r:.2f} (< {mn}, {why})")
+
+    # C. hue families for status roles + harmony
+    for role, lo, hi, fam in [("good",120,180,"green"),("warn",50,120,"yellow/orange"),("bad",0,45,"red")]:
+        h = H[role]
+        ok = hue_in(h, lo, hi) or (role == "bad" and hue_in(h, 330, 45))
+        if C[role] < 0.03:
+            note(slug, "hue", f"{role} is near-grey (C={C[role]:.3f}) — status color carries no hue")
+        elif not ok:
+            note(slug, "hue", f"{role} hue {h:.0f}° outside {fam} family")
+    dh = abs(H["accent"] - H["accentAlt"]); dh = min(dh, 360 - dh)
+    harm = "analogous" if dh < 60 else "triadic-ish" if dh < 150 else "complementary"
+    if 90 <= dh <= 150 and de(c["accent"], c["accentAlt"]) > 0.12:
+        note(slug, "harmony", f"accent/accentAlt Δhue={dh:.0f}° (square-clash zone; {harm})")
+
+    # D. CVD distinguishability of status pairs (small glyphs)
+    for M, name in [(PROTAN, "protan"), (DEUTAN, "deutan")]:
+        for r1, r2 in [("good","bad"),("good","warn"),("accent","bad")]:
+            d = de(cvd_hex(c[r1], M), cvd_hex(c[r2], M))
+            if d < 0.09:
+                note(slug, "cvd", f"{r1}/{r2} nearly identical under {name} (dE={d:.03f})")
+
+    # E. ANSI slot semantics (greeter/tty rely on them)
+    ansi = t.get("ansi", [])
+    if len(ansi) == 16:
+        for idx, lo, hi, fam in [(1,330,45,"red"),(2,100,180,"green"),(3,50,110,"yellow"),
+                                 (4,220,290,"blue"),(5,290,350,"magenta"),(6,160,230,"cyan")]:
+            for slot in (idx, idx+8):
+                Ls, Cs, Hs = oklch(ansi[slot])
+                if Cs >= 0.04 and not hue_in(Hs, lo, hi):
+                    note(slug, "ansi", f"ansi[{slot}] hue {Hs:.0f}° not {fam}")
+        if oklch(ansi[0])[0] > oklch(ansi[15])[0]:
+            note(slug, "ansi", "ansi[0] lighter than ansi[15] (inverted slots)")
+        rt = ratio(ansi[7], ansi[0])
+        if rt < 3.0:
+            note(slug, "ansi", f"ansi[7] on ansi[0] = {rt:.2f} (tuigreet text=gray on container=black)")
+
+print(f"audited {len(themes)} themes\n")
+cats = {}
+for slug in sorted(findings):
+    print(f"── {slug}")
+    for cat, msg in findings[slug]:
+        cats[cat] = cats.get(cat, 0) + 1
+        print(f"   [{cat}] {msg}")
+clean = [t.stem for t in themes if t.stem not in findings]
+print(f"\nclean: {', '.join(clean) if clean else '(none)'}")
+print("totals:", dict(sorted(cats.items())))