HydraIssues

## Summary

Every call to Sunshine's CloseRunningApp endpoint from hydrabody fails with:

```
HTTP 400: {"error":"Content type not provided","status":false,"status_code":400}
```

This happens because the HTTP request does not include a `Content-Type: application/json` header. Sunshine rejects the request, so ForceStop falls back to killing the experience process by name.

## Impact

- Every portrait stream end triggers a GPU mismatch alarm (stream_status=idle but GPU still at 98% for ~30s until the process kill lands)
- The GPU mismatch alarm fires on the streaming monitor `/api/v1/body-anomalies` and in the body log
- hydrabody logs show `[stream] ForceStop failed ... falling back to process kill` on every clean stream end
- The fallback process kill does succeed (`killed 2 process(es) for mercator-talks by name`), so streams do eventually clean up
- No sessions are stuck permanently — the mismatch clears within 30s

## Observed log sequence (chunky-turnip-23, 2026-05-23)

```
10:35:17 [stream] hydracluster requested stream termination — calling ForceStop
10:35:17 [stream] TerminateStream ForceStop error: close running app: HTTP 400: {"error":"Content type not provided",...}
10:35:18 [gpu-mismatch] detected: GPU at 98% with stream status idle
10:35:48 [stream] cleanup: mercator-talks (action=kill, reason=grace period expired, pid=0)
10:35:48 [stream] ForceStop failed (close running app: HTTP 400: ...), falling back to process kill
10:35:48 [stream] killed 2 process(es) for mercator-talks by name
10:35:52 [sunshine] stream ended: mercator-talks
```

The same pattern appears in logs from 2026-05-22 13:26, 15:47, and 21:58 — it is consistent across all stream terminations.

## Root cause

The go-sunshine SDK (or the hydrabody call site) sends the CloseRunningApp POST without a `Content-Type: application/json` header. Sunshine's API requires it.

## Fix

Add `Content-Type: application/json` to the CloseRunningApp HTTP request in go-sunshine (or wherever hydrabody constructs the request). One-line change.

## Related

- VDD portrait restart: after every stream end on portrait mode, hydrabody deletes `display_device.state` and restarts Sunshine (~3 minute window where body cannot accept new streams). This is separate from this bug but compounds the operator experience — the dirty termination alarm fires just before the Sunshine restart window begins.
- Streaming monitor Active Problems panel (v0.4.20) will surface orphaned bodies during the GPU mismatch window if stream_status flips to idle but hydracluster still thinks a head is connected.