fix: leave unbound guids in event payloads untouched instead of closing the connection

[pcockrill]

Problem

Since the guid-resolution rewrite, connection.Dispatch treats any map carrying a guid key inside event params as a reference to a bound protocol object, and if the lookup fails it calls cleanup(...) on the whole connection. From then on every call on that Playwright instance fails with target closed: failed to resolve event parameters for event: failed to resolve map key 'params': object with guid <uuid> was not bound in the connection.

Raw CDP payloads relayed through CDPSession (event with {method, params}) legitimately contain such keys: Chrome's Page.downloadWillBegin and Page.downloadProgress carry a guid that names a download, not a Playwright object. So the first download seen by any page with a CDPSession that has Page.enabled permanently kills the process's driver connection.

Reproduction (v0.5700.1, stock driver): NewCDPSession(page) → Page.enable → Page.setDownloadBehavior allow → page.Goto(url that serves Content-Disposition: attachment). Afterwards browser.NewContext() and even pw.Chromium.Launch() return the error above.

Fix

Mirror the reference JS client (_replaceGuidsWithChannels): a guid that names a bound object resolves to its channel; any other guid value is plain data and the map is walked like every other map. This is also how playwright-go behaved before the rewrite (v0.5200.0 never errored here).

Two unit tests drive Dispatch directly: one asserts an unbound download guid reaches the event listener intact and the connection stays open; the other asserts bound guids (top-level and nested in a slice) still resolve to channels.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Uotgs1HWeELTqn9fLggWoH

[cursor]

Follow-up I wanted on this branch: a live Chromium regression (TestCDPSessionDownloadDoesNotCloseConnection in tests/cdp_session_test.go) that does NewCDPSession + Page.enable + Page.setDownloadBehavior + a real attachment download, then asserts browser.NewContext() still works.

I could not push it here — cursor[bot] gets 403 on Holocron-Technologies/playwright-go even with maintainer edits allowed. The unit tests in this PR already cover the walk; the live test passed locally on this fix. Happy to land it in a follow-up after merge.

Bugbot on the branch (fix + that test) found no bugs. The change LGTM.