fix: harden LLM providers, UTF-8 handling, and webhook/batch reliability

- webclaw-llm: add explicit request + connect timeouts to the reqwest
  client in every provider (anthropic, openai, ollama) with a shorter
  timeout on the ollama health check, so a stalled provider fails fast.
- webclaw-llm: fix a panic when truncating a provider error body that
  contains multibyte characters near the 500-char cut (char-safe take).
- webclaw-core: snap the endpoint-scan budget cut to a UTF-8 char
  boundary so oversized scripts with non-ASCII content no longer panic.
- webclaw-core: rewrite js_literal_to_json to copy raw bytes instead of
  `byte as char`, preserving multibyte UTF-8 in SvelteKit string values
  rather than producing Latin-1 mojibake.
- webclaw-cli: have fire_webhook return its JoinHandle and await it at
  the crawl/batch/batch-llm call sites, removing the fixed 500ms sleeps.
- webclaw-mcp: drop the up-front DNS pre-validation loop in batch that
  aborted the whole request on one bad URL; the fetch layer already
  applies the same SSRF guard per URL and reports per-URL errors.
- webclaw-fetch: include the port in the warmup homepage URL so hosts
  on a non-default port are warmed correctly.

Adds regression tests for the UTF-8 endpoint-scan and SvelteKit cases.

crates/webclaw-mcp/src/server.rs

@@ -323,9 +323,10 @@ impl WebclawMcp {
         if params.urls.len() > 100 {
             return Err("batch is limited to 100 URLs per request".into());
         }
-        for u in &params.urls {
-            validate_url(u).await?;
-        }
+        // No up-front DNS pre-validation: it aborted the whole batch on a
+        // single unresolvable URL. The fetch layer applies the same SSRF
+        // guard (validate_public_http_url) per URL, so bad entries surface
+        // as individual per-URL errors below instead of failing the batch.
 
         let format = params.format.as_deref().unwrap_or("markdown");
         let concurrency = params.concurrency.unwrap_or(5);