fix: retry agent-browser open - close/open race kills ~50% of clone runs

[gkastanis]

What

agent-browser close tears down the daemon's unix socket, and the next open can arrive before it is recreated, producing Could not configure browser: Failed to connect: No such file or directory (os error 2) and a dead run - most often at the mobile-screenshot step, since job 01 ends with a close and 01b starts with an open.

I measured roughly a 50% failure rate on a bare close-then-open sequence. There are nine open call sites across eight job files, so patching one site just moves the failure. This adds lib/browser.js with a browserOpen() helper that retries with linear backoff (300/600/900ms) and routes all nine call sites through it.

It also switches the invocation to execFileSync argv form (no shell), which fixes 03b-extract-assets.js interpolating the target URL into the shell unquoted - a URL containing $(), backticks, or spaces breaks or executes today.

How I verified

• Before: 3 dead runs out of 6 full pipeline runs against a real site.
• After: 0 out of 6; the race still fired on 2 of them and the first retry recovered both.
• Shell-safety: opened a URL containing $() and backticks through the new helper; it opens as data, nothing executes.
• node --check passes on all touched files.