fix: keep GPU quota errors visible in crowd count results

AGENTS.md

@@ -24,3 +24,5 @@ Type: code
 ## Known failure modes
 - **crowd.skyebrowse.com showed a login form / "TypeError: Load failed" (resolved 2026-07-02).** The custom-domain proxy forwards `x-forwarded-host` but not `x-forwarded-proto`, so Gradio computed `root_url = http://crowd.skyebrowse.com` and the https page mixed-content-blocked every fetch; Gradio 6 SSR then rendered its login-form fallback (SSR also dropped `launch(head=...)` meta entirely). Fix, both required: `ssr_mode=False` in `launch()` + the `_https_get_root_url` monkeypatch in app.py (upgrades http→https for non-localhost hosts, because CSR still server-renders `theme.css`/`iframeResizer` URLs from the computed root). Do not remove either when upgrading Gradio; re-verify https://crowd.skyebrowse.com in a browser after any Gradio bump.
 - **HEIC (iPhone) uploads crashed with `PIL.UnidentifiedImageError` in Gradio preprocessing (resolved 2026-08-04, commit ab675ad).** Gradio's `image_utils.preprocess_image` calls `PIL.Image.open` even with `type="filepath"`, and stock Pillow has no HEIC decoder. Fix: `pillow-heif` in requirements.txt + `register_heif_opener()` at import time in app.py (registers globally in PIL). Verified via live-Space API call with a real .heic.
+
+- Browser failure diagnosis (2026-09-11): the native queue SSE returned `You have exceeded your ZeroGPU runs limit.`; the later HF `by-subdomain` 401 was a secondary status lookup, not proof that counting requires public source. Capture SSE errors before diagnosing from HTTP errors alone. Authenticated counting passed (stadium 1,345 people, 4.66s; cached recount 1.10s). Preserve `gr.Error` text in the count field so quota rejection is visible instead of generic Error; keep prior cached candidates on rejection. Protected visibility was requested through HF settings; anonymous app/config stayed accessible and source stayed inaccessible. No reset time or numerical run cap was provided by HF.

app.py

@@ -1,6 +1,8 @@
 import json
 import hashlib
+import html
 import os
+import re
 import threading
 import time
 
@@ -175,8 +177,9 @@ def _count_crowd(image_path, threshold, prediction_state, save_images):
         if save_images:
             save_images_async(image_path, overlay, total)
         return overlay, f"{total:,} people", prediction_state
-    except gr.Error:
-        raise
+    except gr.Error as exc:
+        message = html.unescape(re.sub(r"<[^>]+>", "", str(exc.message))).strip()
+        return None, message, prediction_state
     except Exception as exc:
         print(f"count_crowd failed: {exc}", flush=True)
         return None, "Crowd counter failed. Please try a smaller image or try again in a minute.", None