fix(agents): merge divergent docs-writer duplicates, keep corrected facts

.claude/agent-memory/ux-designer/MEMORY.md

@@ -23,3 +23,4 @@
 - Cannot `gh pr review --request-changes` **or** `--approve` on your own PRs (GraphQL rejects both, not just request-changes) — post via `gh api repos/.../issues/{N}/comments` with an explicit "Verdict: APPROVED/CHANGES_REQUIRED" line in the body instead.
 - `Intl.NumberFormat` enables thousands-grouping by default; `toFixed()`/manual string-building never did. When a PR swaps a `toFixed()` call for a locale formatter, script-verify the _old vs new_ output isn't just per-locale-correct but also identical for large values (≥1000) — grouping separators are an easy-to-miss 3rd English-output regression beyond whatever the PR explicitly discloses. See PR #1845 in [pr-review-findings.md](pr-review-findings.md).
 - `formatDateForAria`-style aria-labels that interpolate localized weekday/month into a hardcoded English sentence order (`"{weekday}, {month} {day}, {year}"`) render grammatically odd in German (`"Dienstag, Februar 24, 2026"` vs correct `"Dienstag, 24. Februar 2026"`) — this is an accepted, deferred, non-blocking gap across the app, not a new bug to flag each time.
+- When a PR claims a Badge/shared-component consolidation is "zero visual delta," build a before/after table (padding/font-size/weight) per migrated component — the shared base class is usually extracted from only a subset of sources, so other migrated components can silently inherit a different size than they had. See PR #1848 in [pr-review-findings.md](pr-review-findings.md).

.claude/agent-memory/ux-designer/pr-review-findings.md

@@ -74,6 +74,16 @@ Full remediation of 370 latent stylelint violations, plus wiring `stylelint` int
 
 Verification method worth reusing: `awk '/^\+/ && !/^\+\+\+/' <diff> | grep -Ei '#[0-9a-fA-F]{3,6}|rgba?\('` and equivalent greps for `z-index:\s*[0-9]`/`font-weight:\s*[0-9]` against the full diff (excluding `tokens.css`) — fast way to confirm zero hardcoded-value regressions slipped past stylelint across a large multi-file remediation PR, rather than spot-checking a handful of files.
 
+## PR #1848 — Modal focus trap + Badge consolidation (Issue #1816) (APPROVED via --comment, Medium finding non-blocking)
+
+Reviewed `Modal.tsx`'s new Tab-cycling focus trap (`getFocusableElements()` re-queried per keypress against `contentRef.current`, which wraps header+body+footer) across all 18 consumers, plus 4 components (`BudgetHealthIndicator`, `CriticalPathCard`, `UpcomingMilestonesCard`, `SubsidyPipelineCard`) migrated from hand-rolled badge spans to shared `Badge` + `BadgeVariantMap`. Zero/single/mid-list focusable-element edge cases all correctly handled and test-covered (6 new `Modal.test.tsx` cases). `KeyboardShortcutsHelp` correctly inherits portal/Escape/focus-trap/dialog semantics by switching to `<Modal>`.
+
+**RECURRING PATTERN TO CHECK: "Badge consolidation = zero visual delta" claims need script/table verification, not eyeballing.** When N pre-existing hand-rolled badge implementations get unified onto one shared `.badge` base class, each source component's own `padding`/`font-size`/`font-weight` gets silently overridden by whatever the *first* group of components (the ones the base class was originally extracted from) happened to use — even though every value on both sides is already a valid design token. Build a before/after table (padding v/h, font-size, font-weight) for every migrated component, not just the ones the author explicitly compared. In PR #1848: `SubsidyPipelineCard` and `CriticalPathCard`/`UpcomingMilestonesCard` only drifted ±2px padding (harmless), but `BudgetHealthIndicator` dropped from 14px/semibold/12px-h-padding to 12px/medium/10px-h-padding — a real, visible size/weight downgrade on an important status badge, not disclosed in the PR description. No WCAG contrast issue (colors unchanged), so Medium/non-blocking, not Critical — but worth flagging explicitly rather than trusting a "ported verbatim" claim.
+
+Also confirmed: dark mode routes correctly for all new Badge variant classes (checked every consumed token has a `[data-theme='dark']` override); `CriticalPathCard`'s div→span tag swap is a genuine no-op (both sides compute to `display: inline-flex`, div's default block was already being overridden); `BudgetHealthIndicator`'s `<span role="status"><Badge/></span>` wrapper correctly preserves live-region semantics on change without double-applying `role`/`aria-live`. Token-snap acceptable: `KeyboardShortcutsHelp`'s `td` padding 0.875rem(14px)→`var(--spacing-4)`(16px) — no 14px spacing token exists in `tokens.css` (only `--spacing-3`=12px/`--spacing-4`=16px), so rounding up to the nearest token is correct under a token-only system.
+
+Informational, not blocking: migrating a component off its own local aria-label (e.g. `KeyboardShortcutsHelp`'s old `keyboardShortcuts.closeAriaLabel`) in favor of a shared component's built-in one (`common:aria.closeDialog`) orphans the old i18n key in both locale files — worth a follow-up i18n-orphan sweep, not a per-PR blocker.
+
 ## Process notes
 
 - Cannot `--request-changes` on your own PRs — use `--comment` instead, and note this in the review body