feat(cdp): add Network.setBlockedURLs and --block-urls request blocking

[mvanhorn]

What

Add request-time URL blocking from two entry points that share one blocklist:

• --block-urls <LIST> — comma-separated *-wildcard patterns, matched case-insensitively against the full URL, on serve / fetch / mcp, next to the existing --block-cidrs.
• Network.setBlockedURLs — the CDP method, matching Chrome's semantics.

Why

Lightpanda's Network CDP domain implemented no request-blocking method (src/cdp/domains/network.zig dispatched enable, disable, setCacheDisabled, header/cookie methods, and getResponseBody, nothing else), and the fetch CLI had no URL blocklist. So a Puppeteer/Playwright script that blocks ad/tracker requests against Chrome via page.setRequestInterception / page.route(..., r => r.abort()) — backed by CDP Network.setBlockedURLs — silently lost that behavior when pointed at Lightpanda. For a high-volume scraper those dropped requests are the point: they cut connections, latency, and bandwidth an agent never reads (Browserless markets the same request/ad blocking as a scraping speedup).

Change

• New src/network/UrlBlocklist.zig, modeled on IpFilter.zig: compiles and owns a set of wildcard patterns once, and answers isBlocked(url) with an allocation-free two-pointer glob match.
• HttpClient holds an optional UrlBlocklist built from config.blockedUrlPatterns() at init. Enforcement sits at the existing request-start gate that already handles robots and IP/CIDR blocking, and also on the redirect target, so a blocked URL never opens a socket and fails via the same async path. Internal transfers (e.g. robots.txt) are exempt, matching the robots / IP-filter exemption. setBlockedUrls() replaces the live list at runtime for the CDP path (owning its own copies).
• A blocked request fails with error.UrlBlocked, which maps through Notification.RequestFail.blocked_reason to Network.loadingFailed with blockedReason: "inspector" — the reason Chrome reports for setBlockedURLs hits — so CDP clients see the expected event.

Demo

Simulated demo (illustrative, not a live capture): --block-urls "*doubleclick*,*analytics*,*://*/*.png" drops the matching ad/tracker/image requests while index.html and the non-matching pixel.gif still load.

Test

• src/network/UrlBlocklist.zig: wildcard-matcher unit tests — case-insensitivity, backtracking, consecutive/trailing *, empty and * patterns, and pattern ownership (mutating the source after init leaves matches intact). These run standalone under the leak-detecting test allocator and pass.
• src/network/HttpClient.zig: setBlockedUrls owns/replaces/clears, and the internal-transfer exemption.
• src/cdp/domains/network.zig: a blocked request emits Network.loadingFailed with blockedReason: "inspector", for both the initial request and a redirect target.
• zig fmt --check clean; zig ast-check clean on all touched files. The full make test needs the V8 toolchain build.

[mvanhorn]

Reworked onto urlPatterns in b573eb1: each pattern carries its block/allow flag (first match wins), the blocklist owns the compiled patterns plus flags, and the legacy urls path stays for back-compat. The three standalone UrlBlocklist tests pass; the full Zig 0.15.2 build runs in CI since my local toolchain is 0.16.

[krichprollsch]

merged via #3040
Thanks!