fix: strict plan filtering in model-aware account routing

When mixing free and team/plus accounts, requesting a plan-restricted
model (e.g. gpt-5.4) could fall back to incompatible free accounts,
causing HTTP 400 "model not supported". Now acquire() returns null
instead of degrading to incompatible accounts when the model's plan
requirements are known but no matching account is available.

Also reduces model-fetcher initial delay from 5s to 1s to populate
the plan map faster at startup.

Fixes #54

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/auth/account-pool.ts

@@ -105,8 +105,10 @@ export class AccountPool {
         const matched = available.filter((a) => a.planType && planSet.has(a.planType));
         if (matched.length > 0) {
           candidates = matched;
+        } else {
+          // No account matches the model's plan requirements — don't fallback to incompatible accounts
+          return null;
         }
-        // else: fallback to all available (graceful degradation)
       }
     }
 

src/models/model-fetcher.ts

@@ -14,7 +14,7 @@ import type { ProxyPool } from "../proxy/proxy-pool.js";
 import { jitter } from "../utils/jitter.js";
 
 const REFRESH_INTERVAL_HOURS = 1;
-const INITIAL_DELAY_MS = 5_000; // 5s after startup
+const INITIAL_DELAY_MS = 1_000; // 1s after startup (fast plan-map population for mixed-plan routing)
 
 let _refreshTimer: ReturnType<typeof setTimeout> | null = null;
 let _accountPool: AccountPool | null = null;