fix: scrape.sh incremental auto-save + chmod permission fix + branch threading

[tamara-builds]

Summary

• Adds a 30-minute background auto-save loop to scrape.sh (mirroring actions/extract's existing pattern) so a runner crash or lost connection mid-scrape only costs the last ~30 minutes of progress instead of the entire run. Prompted directly by FL's 21-hour self-hosted scrape that lost all progress to "runner lost communication with the server."
• Fixes a chmod permission bug: the tarball-packaging step was running chmod -R 755 against the Docker-root-owned source directory instead of the already-copied runner-owned output. On Linux runners this fails with Operation not permitted and (under set -e) silently kills the rest of the script — surfacing as a false "📦 Nightly fallback" job summary with all-N/A stats, even when real data landed and got committed fine. Found live on WI and SD.
• Adds a "Commit Content" section to the job summary showing insertions/deletions from the actual commit, so "re-scraped the same data with fresh UUIDs, no net new content" (e.g. an out-of-session state) is distinguishable at a glance from "genuinely new content" or "lost data" without having to dig through commit stats after the fact.
• Threads branch through as a new argument to scrape.sh so the new auto-save loop pushes to the correct branch instead of assuming main.

Test plan

• Dispatch a small, fast state (not FL) after merge to confirm the auto-save loop and chmod fix work end-to-end in a real run
• Confirm the "Commit Content" summary renders correctly
• Only then point FL (and its tinyproxy run) at these changes

🤖 Generated with Claude Code