test: stabilize final PR 40 validation gates

tests/css_snapshot/README.md

@@ -92,6 +92,20 @@ same bytes, so the capture:
 - injects the theme and density classes into `<html>` *before* first paint
   rather than toggling them afterwards, so no CSS transition is ever
   mid-interpolation while `getComputedStyle` runs;
+- removes `autofocus` before parsing: focus states are outside this inventory,
+  and the browser's asynchronous autofocus step otherwise races the capture;
+- pauses CSS animations at time zero and finishes CSS transitions before each
+  measurement, including newly revealed modals and newly mounted bench nodes.
+  Animation and transition declarations are still captured; the harness does
+  not inject `animation: none` or `transition: none`;
+- pins Chromium's standard font preference to `Times New Roman` via CDP,
+  without overriding any author declaration;
+- canonicalizes only the `BlinkMacSystemFont` family token to `"system-ui"`,
+  the spelling Chromium uses for that alias on macOS. Other family names and
+  their order remain significant;
+- measures the `custom-system-prompt` element's `max-height` in `lh`, as opted
+  into by its inventory entry. Its authored `30lh` resolves to different pixel
+  heights with different fallback fonts; the line count remains significant;
 - aborts images, fonts and media, which cost time and change nothing in the
   pinned property set;
 - hides scrollbars, so a platform's scrollbar width cannot change the width
@@ -111,11 +125,11 @@ same bytes, so the capture:
 - **JS-applied classes.** State the app adds at runtime (collapsed sidebar,
   open panels, active tabs) is not represented beyond what the served markup
   and the bench selectors already carry.
-- **Cross-platform equality has not been measured.** The baseline was recorded
-  on macOS. The self-hosted Fira Code face means text metrics should not differ
-  from CI's Linux Chromium, and the layout-derived properties are excluded, but
-  until a Linux run confirms it, treat a CI-only drift as a possible harness
-  artifact and diff the dumps before assuming the CSS moved.
+- **Browser upgrades and additional platforms.** The original macOS baseline
+  and Linux captures were compared property by property through exact hash
+  recovery; the proven platform differences are now controlled above. A new
+  capture on macOS has not been performed. New browser serialization changes
+  still need investigation rather than automatic baseline regeneration.
 - **The stylesheet is only one of the inputs.** `static/login.html` styles
   itself from an inline `<style>` block; it is in the inventory so the
   hand-mirrored token values there are pinned too.
@@ -127,3 +141,13 @@ Add an entry to `inventory.json` - an `{key, selector}` object under a page's
 include custom properties), or a selector string under `bench` - then
 re-record the baseline. `test_baseline_covers_every_inventory_entry` fails if
 the two go out of sync.
+
+`test_capture_is_independent_of_elapsed_time_and_font_metrics` perturbs capture
+timing and font metrics, checks autofocus suppression, and proves that animation
+keyframes, metadata and relative line counts still affect measurements. The
+existing cascade-order self-test still detects a reordered declaration.
+
+The PR #40 canonical baseline and its exact justification are documented in
+[`pr40-validation.md`](pr40-validation.md). All 122 properties and 676 inventory
+elements remain covered; only one element/property opts into line-relative
+measurement.