wcag: linearize below the cited 0.04045 knee, not the older 0.03928

dag/extdeps/accessibility/wcag_contrast.dag

@@ -13,9 +13,10 @@ data extdeps_external_authority_anchor: ExternalAuthority = ExternalAuthority {
 }
 
 // Cited upstream: W3C WCAG 2.2, definitions of relative luminance and contrast ratio (retrieved
-// 2026-07-21). Relative luminance L = 0.2126 R + 0.7152 G + 0.0722 B where each channel c (sRGB,
-// 0..1) linearizes as: c <= 0.03928 -> c/12.92, else ((c+0.055)/1.055)^2.4 (the 0.03928 knee is
-// WCAG's published constant). Contrast ratio = (L_lighter + 0.05) / (L_darker + 0.05); thresholds
+// 2026-10-05). Relative luminance L = 0.2126 R + 0.7152 G + 0.0722 B where each channel c (sRGB,
+// 0..1) linearizes as: c <= 0.04045 -> c/12.92, else ((c+0.055)/1.055)^2.4 — the 0.04045 knee is
+// the cited constant (the definition's Note 2 records 0.03928 as the threshold before May 2021;
+// 2.2 aligned it with sRGB). Contrast ratio = (L_lighter + 0.05) / (L_darker + 0.05); thresholds
 // 4.5:1 (AA normal text), 3:1 (AA large text and non-text graphical objects, per SC 1.4.11), 7:1
 // (AAA). Arithmetic here is integer at scale 10000 (one unit = 0.0001). The single approximated
 // kernel is the exponent 2.4, computed as x^2.4 = x^2 * (x^2)^(1/5) with the fifth root by integer
@@ -74,8 +75,10 @@ fn pow_2_4_scaled(x_scaled: Int) -> Int {
   }
 }
 
+// The cited knee is 0.04045; c_scaled is an integer at scale 10000, so c <= 0.04045 is
+// c_scaled <= 404 (404.5 is not representable, and 404 is the largest integer at or below it).
 fn linearize_channel_scaled(c_scaled: Int) -> Int {
-  if c_scaled <= 392 {
+  if c_scaled <= 404 {
     (c_scaled * 100) / 1292
   } else {
     pow_2_4_scaled(x_scaled: ((c_scaled + 550) * 10000) / 10550)

dag/gunbc/recurring_failure_mode/a_citation_attests_a_decision_rule_the_code_does_not_implement.dag

@@ -15,6 +15,8 @@ data a_citation_attests_a_decision_rule_the_code_does_not_implement: RecurringFa
 
     "**RECEIPT (vivid-ibex-684, 2026-10-04, gunbc#13258, side-chat review `5406002589`; defect live in the tree).** `dag/extdeps/accessibility/wcag_contrast.dag` cites WCAG 2.2 (www.w3.org/TR/WCAG22/#dfn-relative-luminance) and its comment asserts the linearization knee is the page's published constant at 0.03928 -- and the code does use 392/10000. But the cited definition specifies the threshold 0.04045 and notes 0.03928 as the older value, so the citation attests a definition the code does not implement. The module was authored 2026-07-21 with the comment noting the page as retrieved that day, so the pairing was already false when authored -- there is no upstream-drift story here, only a misattestation that name-level checks (live URL, subject ownership) do not see. Pulled from #13258 per that review rather than fixed; the defect is live as of this receipt.",
 
+    "**REPAIR RECEIPT (vivid-lynx-377, 2026-10-05, frontier; wcag half).** `dag/extdeps/accessibility/wcag_contrast.dag`'s linearization knee is now the cited 0.04045 (integer `c_scaled <= 404` at scale 10000) and the comment cites the definition's Note 2 recording 0.03928 as the threshold before May 2021 (live page re-read 2026-10-05). The threshold half of the #13258 receipt was checked against the live page and already matched: 450/300/700 transcribe SC 1.4.3 (4.5:1 normal, 3:1 large-scale), SC 1.4.11 (3:1 graphical objects) and SC 1.4.6 (7:1 AAA); pinned by `witness_cited_knee_threshold_rows_match_criterion_text`, no change made. The knee's discriminating red samples the only window where the two constants disagree -- no 8-bit channel at unity white lands there (10/255 = 0.03922 is below both knees), which is why the module's earlier pins never saw it -- and pins the outcome flip: `linearize_channel_scaled(393)` returns 30 under the cited knee (29 under the defective one), and the colour pair (393,0,500)/(8000,1000,8000) FAILS AA normal text under the cited knee while the defective knee called it sufficient. Both green after the fix (dag/test/claim/wcag_contrast_witness_test.dag, remote claim_batch --claim-run). The semver half of this row is repaired in a separate PR (#13412).",
+
     "**WHY NO EXISTING WALL SEES IT, stated as a boundary rather than a gap to be closed by the neighbour.** The extdeps scope placement gate adjudicates the diff-added pairing of a subject declaration to a citation anchor at the level of NAMES: it can see that a subject is not the cited upstream's surface (tool CLI vs standard), because that mismatch is visible in what the module declares and what the anchor names. The lexical-vs-precedence difference lives in the code the declaration dispatches through, not in any name, module path, or anchor field; DESIGN section 4c makes annotations invisible to semantic passes, and a scope block's citation is an annotation about authority, not a type on the code path. So a fully green placement gate and a green namespace wave are both positive evidence about ownership and binding, and NO evidence about whether the decision the code implements is the one the standard defines.",
 
     "**THE SAME FAILURE WITHOUT A CITATION HUNK.** The false attestation propagates: any caller that takes a module's compared/verified value as standard-ordered inherits the gap, and any new comparison path added under an already-cited module inherits the citation without touching prose. Neither edit produces a citation hunk the recognition rule below would see, which is why the rule names the code path, not only the anchor.",

dag/test/claim/wcag_contrast_witness_test.dag

@@ -2,10 +2,13 @@ module test.claim.wcag_contrast_witness
 
 import extdeps.accessibility.wcag_contrast {
   linearize_channel_scaled, relative_luminance_scaled,
-  contrast_ratio_x100,
+  contrast_ratio_x100, contrast_at_least_hundredths,
+  rgb_scaled_luminance,
   hsl_contrast_ratio_x100, hsl_contrast_at_least,
   wcag_aa_normal_text_hundredths, wcag_aa_large_text_and_graphics_hundredths,
+  wcag_aaa_normal_text_hundredths,
 }
+import extdeps.color.srgb { RgbScaled }
 import gunbc.design.theme {
   Theme, assign, theme_contrast_ok,
   night_theme, day_theme, night_canvas,
@@ -36,6 +39,43 @@ test fn witness_gamma_kernel_pinned() -> Bool {
     && low == (300 * 100) / 1292
 }
 
+// WCAG 2.2 #dfn-relative-luminance (https://www.w3.org/TR/WCAG22/#dfn-relative-luminance),
+// cited by the module: for the sRGB colourspace, R = R_srgb/12.92 if R_srgb <= 0.04045,
+// else ((R_srgb+0.055)/1.055)^2.4 — same for G and B. (Note 2 there records that the
+// threshold was 0.03928 before May 2021; the current threshold is 0.04045, aligned with sRGB.)
+//
+// The cited knee and the code's knee disagree only on channels between 0.03928 and 0.04045
+// (scale-10000: 393..404) — no 8-bit channel at unity white lands there, so the existing
+// pins never sample the window. A channel at 0.0393 (393) does.
+
+// 0.0393 <= 0.04045 -> the cited formula linearizes with c/12.92: (393*100)/1292 = 30.
+// The defective 0.0392 knee sends 393 to the gamma curve instead (29).
+test fn witness_cited_knee_linearizes_below_threshold() -> Bool {
+  linearize_channel_scaled(c_scaled: 393) == (393 * 100) / 1292
+}
+// The verdict FLIPS at a real WCAG threshold for a fresh input: A = (393, 0, 500) scaled —
+// nearly black, red channel inside the knee window. Cited: L_A = (2126*30 + 722*39)/10000 = 9.
+// Defective knee: L_A = (2126*29 + 722*39)/10000 = 8. B = (8000, 1000, 8000) scaled: L_B = 1789
+// either way. AA normal text is 4.5:1 (450 hundredths); 100*(1789+500) = 228900:
+//   cited knee:     450*(9+500) = 229050 -> 228900 < 229050 -> contrast is NOT sufficient;
+//   defective knee: 450*(8+500) = 228600 -> 228900 >= 228600 -> (wrongly) sufficient.
+test fn witness_cited_knee_flips_aa_normal_verdict() -> Bool {
+  let lum_a = rgb_scaled_luminance(rgb: RgbScaled { r: 393, g: 0, b: 500 })
+  let lum_b = rgb_scaled_luminance(rgb: RgbScaled { r: 8000, g: 1000, b: 8000 })
+  lum_a == 9
+    && lum_b == 1789
+    && !contrast_at_least_hundredths(lum_a_scaled: lum_a, lum_b_scaled: lum_b, threshold_hundredths: wcag_aa_normal_text_hundredths)
+}
+
+// Threshold half of the citation: SC 1.4.3 (AA) 4.5:1 normal text and 3:1 large-scale text;
+// SC 1.4.11 (AA) 3:1 for graphical objects; SC 1.4.6 (AAA) 7:1. The module's rows must keep
+// transcribing those numbers.
+test fn witness_cited_knee_threshold_rows_match_criterion_text() -> Bool {
+  wcag_aa_normal_text_hundredths == 450
+    && wcag_aa_large_text_and_graphics_hundredths == 300
+    && wcag_aaa_normal_text_hundredths == 700
+}
+
 test fn witness_reviewer_oracle_night_text() -> Bool {
   let r = hsl_contrast_ratio_x100(hue_a: 35, sat_a: 14, light_a: 78, hue_b: 226, sat_b: 30, light_b: 7)
   r >= 1100 && r <= 1200