The noise floor is per-page: pinned tolerances are load-bearing on CI

.okf/build/test-gates.md

@@ -120,7 +120,9 @@ the gate was blind to it by 50x - and passed, which is worse than failing
 page is ~1e-6, not 0 (three runs of `services/_technologies`: 0.013838252,
 0.013838252, 0.013837770 - about 2 px of 2,073,600), so the floor has to clear
 that; 0.0001 (~207 px) sits 100x above it, fails that copy edit, and goes green
-again when it is reverted. The noise claim is
+again when it is reverted. **That 1e-6 is the macOS-local figure and does NOT
+generalise - see the per-platform noise section below before you lower any
+tolerance.** The noise claim is
 checkable without re-running the copy edit: two independent runs of the same
 test reported difference_levels identical to 16 decimal places (0.6893909143518518
 for `services/_footer` every time), and 2 of the 7 screenshots stayed GREEN at
@@ -759,3 +761,31 @@ and confirm the page loads its own fingerprinted `.min.<hash>.css` before
 trusting any computed value (caught 2026-08-21 during the pilot column-fix
 verification; the earlier "inconclusive" production check had exactly this
 cause).
+
+# The noise floor is per-PAGE, not per-repo (measured 2026-08-22)
+
+The ~1e-6 above was measured on macOS, locally, on a static page. CI amd64 is
+different, and the difference decides whether a tolerance may be lowered.
+
+**Method**: two `update-baselines` dispatches of the SAME commit, ~12 hours
+apart, compared file by file. Anything that moves between two recordings of
+identical content is render noise by definition.
+
+**Result**: **135 of 147 Linux baselines were byte-identical.** CI is
+deterministic for ~92% of pages. The remaining 12 split cleanly:
+
+| Pages | difference_level | What they are |
+|---|---|---|
+| 11 | 0.0003 - 0.002 | mermaid diagrams, syntax-highlighted code, course + about + clients pages - font and SVG rasterisation |
+| 1 | 0.0425 (max_delta 238) | `mobile/blog/index/_pagination` - a real content difference, not noise |
+
+**The consequence that matters: every one of those 11 noisy pages already pins
+`tolerance: 0.03` at its call site.** Those pins are LOAD-BEARING on CI, not
+legacy padding left over from the blind era. Lowering them toward the 0.0001
+default - which an earlier plan proposed as a tidy-up - would redden CI
+permanently on pages that have no defect.
+
+**Rule.** Before lowering any pinned tolerance, measure that page's noise on
+the platform that will judge it, by recording it twice on the same commit. The
+floor for a page is its measured noise, not the repo default. A tolerance that
+looks like slack may be the only thing absorbing a rasteriser.

.okf/log.md

@@ -51,6 +51,34 @@ make it green: restructure same-day entries under one heading, and add
 `timestamp` to the 23 concepts missing it (anchored to each file's last commit
 time, which is verifiable - never invented).
 
+## 2026-08-22 - the noise floor is per-page, and the pinned tolerances are load-bearing
+
+Chasing the last red Linux key produced a better finding than the fix. Four
+hypotheses died on the way - arm64-vs-amd64 drift (the files measured
+pixel-identical against a control), #560's tolerance drop (that test pins its
+own 0.03, so the default never applied), the PR-merge-commit checkout, and a
+date-gated post - and each one cost a cycle because it was reasoned rather
+than measured.
+
+What the measurement says: two `update-baselines` dispatches of the SAME commit
+twelve hours apart left **135 of 147 Linux baselines byte-identical**. CI is
+deterministic for ~92% of pages. Eleven font/SVG-heavy pages (mermaid, syntax
+highlighting, course, about, clients) moved by 0.0003-0.002, and the twelfth,
+`mobile/blog/index/_pagination`, moved by 0.0425 with a max channel delta of
+238 - a real content difference, which is why it was the only failure.
+
+**Every one of those eleven pins `tolerance: 0.03` at its call site.** Those
+pins are not legacy padding from the blind era; they are the only thing
+absorbing a rasteriser on CI. An earlier plan to measure and lower the section
+tolerances toward the new 0.0001 default would have reddened CI permanently on
+pages with no defect. That plan is now scoped out, and the rule is written
+where the tolerance lives: measure a page's noise on the platform that judges
+it - by recording it twice on one commit - before lowering anything.
+
+Also corrected: the ~1e-6 noise figure in test-gates and in the code comment
+was macOS-local and read as universal. Same overclaim shape the #566 reviewer
+caught one layer down.
+
 ## 2026-08-22 - the site sells one thing, and title claims are now a canon rule
 
 Paul changed the positioning: the site sells an embedded team of senior

test/application_system_test_case.rb

@@ -78,7 +78,12 @@ class ApplicationSystemTestCase < Minitest::Test
   # the baseline fossilized. Measured run-to-run noise is ~1e-6 (three runs of
   # services/_technologies: 0.013838252 twice, then 0.013837770 - about 2 px of
   # 2,073,600), so the floor only has to clear that: 0.0001 is ~207 px. Calls
-  # needing real slack (font-swap, animation) still pin their own tolerance.
+  # needing real slack (font-swap, animation) still pin their own tolerance -
+  # and on CI those pins are LOAD-BEARING, not padding. That 1e-6 is macOS
+  # local; two CI recordings of the same commit left 135 of 147 baselines
+  # byte-identical but moved 11 font/SVG-heavy pages by 0.0003-0.002, every one
+  # of which pins 0.03 here. Do not lower a pinned tolerance without measuring
+  # that page's noise on the platform that judges it (.okf/build/test-gates.md).
   DEFAULT_SCREENSHOT_CONFIG = {tolerance: 0.0001}.freeze
 
   private