fix(vinpin): cap Rpartstore acquire at ~1 open via 14s sub-budget

Rpartstore's launch-error modal OCR-detection is unreliable across renderings,
so a DOWN Rpartstore was grinding all 3 reopens (~54s) + dirtying the seat →
Dialogys fallback couldn't finish inside the decode budget → not_found.
Lower acquireBudgetMs 90s→14s so the loop bails after the first open+poll (~22s)
straight to Dialogys on a still-clean seat. Healthy Rpartstore loads on the first
open and is used as before; maxOpens kept so tests still exercise the reopen path.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

apps/api/src/integrations/vinpin/vinpin.constants.ts

@@ -399,8 +399,20 @@ export const VINPIN_DECODE_BUDGET_MS = 180_000;
  * to Dialogys rather than burning the entire decode budget on reopens.
  */
 export const VINPIN_RPARTSTORE = {
+  // Rpartstore has two failure modes: a genuine born-stuck spinner (reopen can
+  // recover it) and a hard server-side launch error "Ошибка запуска каталога"
+  // (reopen NEVER helps — Rpartstore is DOWN). Detecting the tiny centered error
+  // modal by OCR proved unreliable across renderings, and burning reopens on a
+  // DOWN Rpartstore (a) wastes ~54s and (b) dirties the RDS desktop so the
+  // Dialogys fallback then can't finish inside the budget.
+  // `maxOpens` still bounds reopen attempts (a real transient spinner can recover
+  // on a reopen), but `acquireBudgetMs` is the hard lever: one open+load-poll is
+  // ~22s, so a 14s wall-clock sub-cap makes the loop bail after the FIRST open if
+  // it hasn't loaded — straight to Dialogys on a still-clean seat (~25–52s decode)
+  // — instead of grinding 3 reopens. A healthy Rpartstore loads within the first
+  // open's poll and is used as before; the cap only prevents wasted reopens.
   maxOpens: 3,
-  acquireBudgetMs: 90_000,
+  acquireBudgetMs: 14_000,
 } as const;
 
 /** Per-key delay (ms) when typing a VIN into a focused canvas field. A focused