MPDX-10133 Stop error link crash that left pages stuck loading

.claude/review/learnings/feedback.jsonl

@@ -45,3 +45,6 @@
 {"ts":"2026-09-29T15:44:38.863Z","reviewId":"address-2092","id":"f3","signature":"2e44759de6ee","agent":"architecture","category":"pattern-consistency","severity":4,"file":"src/components/HrTools/PdsGoalCalculator/Setup/SetupStep.tsx","message":"The goal year is now passed by hand into three separate PDS calls to useGoalCalculatorConstants (SetupStep, MonthlyReimbursableSection, usePdsSummaryData). GoalCalculator does this differently: its context owns one year-scoped constants result and tells consumers not to call the hook themselves, so a future PDS consumer that calls the hook with no year will quietly get current-year rates.","outcome":"accepted"}
 {"ts":"2026-09-29T15:47:11.190Z","reviewId":"address-2092","id":"f7","signature":"9caaa653a14a","agent":"financial","category":"date-windows","severity":3,"file":"src/components/HrTools/PdsGoalCalculator/Shared/PdsGoalCalculatorLayout.tsx","message":"The past-year comparison reads DateTime.local().year fresh on every render instead of capturing it once, unlike the repos established now = useMemo(() => DateTime.now(), []) pattern for long-lived pages. On a session left open across a year boundary the past-year banner/chip can appear or disappear mid-session even though nothing about the goal changed.","outcome":"accepted"}
 {"ts":"2026-10-05T20:43:32.007Z","reviewId":"address-2123","id":"f1","signature":"53773db921b4","agent":"architecture","category":"state-consistency","severity":6,"file":"src/components/HrTools/AdditionalSalaryRequest/Shared/useAdditionalSalaryRequestForm.ts","message":"The new validator guard checks the sticky stableNonBackpayTotal (falls back to lastValidNonBackpayTotalRef), while useSalaryCalculations checks the live nonBackpayTotal, so the two exceedsCap answers split apart. An over-cap staffer who types a non-backpay amount and then clears it back to backpay-only sees \"Optional Comments\" and an enabled Submit, but Formik still fails additionalInfo as required.","outcome":"accepted"}
+{"ts":"2026-10-07T21:11:54.597Z","reviewId":"address-2127","id":"f1","signature":"52aed5594f03","agent":"testing","category":"test-cannot-fail","severity":5,"file":"src/components/Reports/StaffExpenseReport/StaffExpenseReport.test.tsx","message":"`expect(calls.report).toBe(0)` can never fail: `calls.report` only increments inside `failCallsWhen`, and this test passes no `failReportCall`, so the counter is 0 even if the report query runs. The test does not prove the report query is skipped when HCM errors.","outcome":"accepted"}
+{"ts":"2026-10-07T21:17:39.702Z","reviewId":"address-2127","id":"f2","signature":"5189466fa1ce","agent":"architecture","category":"pattern-consistency","severity":4,"file":"src/lib/apollo/client.ts","message":"The non-array `graphQLErrors` guard is fixed only in the browser error link; the sibling SSR error link in src/lib/apollo/ssrClient.ts still calls `graphQLErrors.map(...)` behind a truthiness check, so a string or object `errors` throws inside the link the same way.","outcome":"accepted"}
+{"ts":"2026-10-07T21:42:15.282Z","reviewId":"address-2127","id":"f1","signature":"d423e73efc0c","agent":"architecture","category":"pattern-consistency","severity":4,"file":"src/components/Reports/StaffExpenseReport/StaffExpenseReport.tsx","message":"The new HCM failure path folds every Hcm error into the generic \"Staff Expense report could not be loaded. Please try again later.\" alert, while the repo already has a dedicated convention for the Hcm query failing with HCM_UNAVAILABLE (isHcmUnavailableError + HcmUnavailableAlert). When HCM is overloaded, this page tells users something different from every other HCM-backed page.","outcome":"dismissed","dismissalReason":"intentional","dismissalDetail":"it's no longer possible to get an HCM_UNAVAILABLE error on the Staff Expense Report except in an extreme edge case that isn't worth having special logic for"}