fix: early polyfill injection via onCommitted + CSP bypass

[madmath]

Issue

I observed that the polyfill doesn't inject soon enough in the page lifecycle. Which doesn't guarantee tools would register.

Summary

• Inject the WebMCP polyfill via webNavigation.onCommitted + chrome.scripting.executeScript with files:[] and injectImmediately: true, guaranteeing it runs before page scripts and bypasses strict page CSP
• Keep the manifest content_scripts entry (world: MAIN, document_start) as a belt-and-suspenders fallback; the polyfill's own guard prevents double-init
• Remove noisy duplicate log when both injection paths fire

Test plan

• Load extension in Chrome, open a page with strict CSP (e.g. GitHub), verify navigator.modelContext is available in console
• Serve the inline-script timing test page (<script> checking 'modelContext' in navigator at parse time), confirm available: true
• All 473 unit/integration tests pass

🤖 Generated with Claude Code

[igrigorik]

@madmath huh, do you have a ~reliable way to repro this? Intuitively, the strategy in this PR should be strictly worse..

According to Google docs content scripts declared via manifest are the first to be injected, and should be executing before any executeScript call, etc. Manifest scripts are injected by the browser engine synchronously with document creation vs. much laxer contract for injectImmediately. This doesn't smell right.

[madmath]

You can repro with https://peroh.com, it has a script in the head which will look for window.navigator.modelContext and prior to this change, it's not available. Previous discussion online talks about this approach, which is probably why Claude picked it up. Let me know what you want to do!

[igrigorik]

I can't reproduce this, the tools are coming up reliably here. Are you testing on local preview / staging?

@madmath sanity check: are you using latest version from main?

[madmath]

Annnnd... I can't reproduce. I was on latest (last commit Feb 17), not sure what happened. Glad it's fine now!

[igrigorik]

\o/