Cursor real SDK proof for https://github.com/pingdotgg/t3code/pull/17462
Reviewed/published head: aa4605cf7d0964b6db02d3d80850a11ffb0c4fa2
Linux, Node 24, real Cursor SDK 1.0.35, production T3 adapter and runner.
Local executor intercepts the outgoing SDK request before model network traffic.
Catalog is synthetic, with a 1m default. No authenticated remote Cursor request or UI change.

Before main: runtime context capacity is undefined, native integrated assertion expected 1000000 and failed.
The original PR also mislabeled explicit 500k as 1m; rewritten driver/runtime regression catches that mismatch.
After: same integration scenario passes, 1 test in 1 file, 4.98 seconds.
400000-history estimate; reported capacities and handoff allowances match outgoing context selections below.
500k is a stale/arbitrary value here: SDK serialization is verified, backend acceptance is not claimed.
All 125 Cursor and 19 handoff-budget tests pass. PR-added tests 2 -> 1 integrated driver/runtime case.

Captured outgoing SDK models and computed budget receipts (synthetic fixture values):
{
  "requests": [
    {
      "id": "claude-opus-5-5",
      "params": [
        {
          "id": "context",
          "value": "1m"
        }
      ]
    },
    {
      "id": "claude-opus-5-5",
      "params": [
        {
          "id": "context",
          "value": "500k"
        }
      ]
    },
    {
      "id": "claude-opus-5-5"
    }
  ],
  "proofRows": [
    {
      "context": "1m",
      "capacity": 1000000,
      "budget": 16000
    },
    {
      "context": "500k",
      "capacity": 500000,
      "budget": 0
    },
    {
      "capacity": 1000000,
      "budget": 16000
    }
  ]
}
Limits: no actual Cursor account/backend model execution, browser, macOS, Windows, iOS, Android or Electron validation.