test: timing budgets + CI hardening (artifact v3 pin, npm cache retry)

Closes three items from the outstanding e2e/CI gap list:

- Timing budgets: timings.measure() only ever recorded durations for the
  (still-unwired) Pushgateway export — nothing failed CI when a step got
  slow. Add e2e/timing-budgets.ts (expectWithinBudget + shared BUDGET_MS
  buckets) and wire it into every measure() call site across the suite.
  Mail-wait steps are deliberately left unbudgeted (external mail-trap
  delivery latency, not a code performance signal).

- actions/upload-artifact@v4 doesn't work against this Gitea/act runner's
  artifact backend — pin to v3 for the e2e failure-report upload.

- Shared act_runner npm cache has corrupted platform-native tarballs before
  (@next/swc-linux-x64-musl) and reds viewer-unit/admin-unit/e2e with no
  product bug involved. All three npm ci steps now retry once after
  `npm cache clean --force` on first failure.

Verified: full local suite green against DEV (37 passed, 6 skipped, no
budget assertion failures) before wiring into CI.

e2e/tests/api.health.spec.ts

@@ -1,5 +1,6 @@
 import { z } from 'zod';
 import { test, expect } from '../fixtures';
+import { BUDGET_MS, expectWithinBudget } from '../timing-budgets';
 
 const HealthSchema = z.object({ status: z.literal('ok') });
 
@@ -8,6 +9,7 @@ test.describe('api @smoke', () => {
     const res = await timings.measure('api_health', () =>
       api.get('/health', { expectedStatus: 200, schema: HealthSchema }),
     );
+    expectWithinBudget(timings, 'api_health', BUDGET_MS.api);
     expect(res.data).toEqual({ status: 'ok' });
   });
 });