fix: keep startup alive when gradio's localhost probe times out on large UIs

[saintorphan]

Problem

WanGP can crash at startup with:

ValueError: When localhost is not accessible, a shareable link must be created.
Please set share=True or check your proxy settings to allow access to localhost.

This was reported by a Pinokio/Windows user who had just enabled two plugins (ImageSuite + Prompt Library); disabling the plugins in wgp_config.json made it start again. Nothing was wrong with their network — the trigger is UI size.

Root cause

Gradio 5.29's launch() verifies localhost by probing HEAD / with httpx.head(timeout=3), 5 attempts. But serving / deep-copies the entire Blocks config and filters components per page with component["id"] in config["page"][page]["components"] — membership tests against a plain list, i.e. O(N²) in component count (gradio routes.py, still present on gradio main).

Measured per-request cost of the / handler on a fast Linux box (gradio 5.29.0):

components	deepcopy	id-filter	first response
8k	0.08s	0.13s	~0.3s
16k	0.20s	0.53s	~0.8s
32k	0.32s	2.22s	>3s → launch() raises

A plugin-free synthetic Blocks app at ~32k components reproduces the exact crash on a fast machine; slower Windows machines fail at far lower counts. WanGP's base UI plus a couple of plugin tabs is enough to cross the threshold — which is why "disable plugins" looks like a fix.

The same mechanism also explains intermittent sightings on fast machines without plugins: the probe allows five attempts of exactly 3s over a ~17s window at the busiest moment of startup. A borderline first response (WanGP's base UI alone costs a second or two) plus one unlucky burst of contention in that window — first-request template compile, model preloading, antivirus scanning freshly-touched checkpoints — pushes all five attempts over. Deterministic for big-UI setups, a race for everyone else.

Fix

Patch gradio.networking.url_ok just before demo.launch(): run the stock probe first, unchanged; only when it gives up, retry loopback URLs only once with a 60s timeout and trust_env=False (a system HTTP proxy must not intercept a localhost probe — proxies are the other classic trigger of this same error).

Properties:

• Normal startups are byte-for-byte identical: the stock probe runs first and passes as before.
• A genuinely dead localhost still fails fast — connection-refused returns immediately in both passes; only "alive but slow to serve the first page" is rescued.
• The share-tunnel polling loop (while not networking.url_ok(self.share_url)) is untouched: patience applies to loopback URLs only.
• Idempotent, and coordinates via sentinel attributes with the identical shims already shipped in the ImageSuite and Prompt Library plugins (so exactly one patch installs regardless of load order). A host-side fix is still the right place: any sufficiently large plugin reproduces this, not just the two that patched around it.
• If gradio's internals ever move, the patch logs a note and leaves everything untouched.

Verified against gradio 5.29.0: the 32k-component app that crashes stock launch() starts successfully with the patch, and both plugin shims correctly back off when the host patch is present.

The real long-term fix is one line in gradio's routes.py (set(...) around the page-membership lists); I'll file that upstream separately. This patch keeps WanGP startable in the meantime and remains harmless once gradio fixes it.