Handoff — commit & push everything in battlebox's WSL before the rebuild

Summary

Bitbucket SSH auth from the box is confirmed working — ssh -T git@bitbucket.org returns "authenticated via ssh key". Keys are ~/.ssh/id_ed25519 and ~/.ssh/id_ed25519_bitbucket_battlebox.

Critical find — do this one first

Step 0 — cheap safety net before touching anything

Copy every .git directory out of the VHDX to the D: drive first. This is full-fidelity — it preserves reflogs and all stash entries, which a git push does not — and it is small, since the working trees (the bulk of the 129 GB) stay behind.

mkdir -p /mnt/d/wsl-rescue
cd ~/dev
for d in */; do
  [ -d "${d}.git" ] || continue
  rsync -a --info=progress2 "${d}.git" "/mnt/d/wsl-rescue/${d%/}/"
done
du -sh /mnt/d/wsl-rescue

Also note this copy is only a stopgap. /mnt/d/wsl-rescue survives the distro deletion, but it is not a substitute for pushing to Bitbucket — finish the pushes.

What's at risk, by tier

Not all 14 repos matter equally. Sorted by what is actually lost if the distro is deleted right now:

Per-repo detail

Tier 1 Unpushed commits

The last one also has a deleted data/03_raw/.gitkeep and two untracked output dirs (data/04_raw/, data/interim_overlays) — decide whether the .gitkeep deletion is intentional before committing.

Tier 2 Uncommitted source

Tier 4 Junk, ignore-candidates, and one to check

Stashes — 36 across 4 repos

To promote one: git stash branch <name> 'stash@{n}' checks out its parent commit, applies it, and drops the stash — then commit and push the branch.

Pragmatic suggestion for lanefinder: 33 stashes going back to 2024 is not worth triaging one by one. The .git copy from Step 0 preserves all of them. Propose to Miro: promote the two named/deliberate ones, keep the rest only in the /mnt/d/wsl-rescue archive, and move on.

Work order

1. Step 0 safety net — rsync all .git dirs to /mnt/d/wsl-rescue. Verify with du -sh.
2. tablecloth — confirm the Bitbucket URL with Miro, git remote set-url origin, push main. Nothing else proceeds until these 9 commits are on a real server.
3. asphaltedge (5) and linebitmapimagerasterizer (1) — straight git push, remotes are fine.
4. guardrails, visualizationoverlays, trajectoryfilter — review each diff with Miro, commit under the right AI3D- ticket, push. Prefer a feature branch for guardrails rather than committing straight to main.
5. rasterizer stash — promote; it's current work.
6. lanefinder — promote the two deliberate stashes; commit the job script and .gitignore/uv.lock changes; leave the other 31 in the archive.
7. orchestrator yaml and windowchecker obj.py — inspect, then commit or discard.
8. Dead repos — pytorchinstancesegmentation (2023), pdfpropertyai (2023), windowchecker (2024): confirm with Miro they're archive-only, then discard the untracked output.
9. Final gate: re-run the audit script (below) and confirm every repo reports zero unpushed commits and zero unexplained dirt. Only then is the rebuild safe to start.

cd ~/dev
for d in */; do
  d=${d%/}; [ -d "$d/.git" ] || continue
  st=$(git -C "$d" status --porcelain 2>/dev/null)
  un=$(git -C "$d" log --branches --not --remotes --oneline 2>/dev/null)
  stash=$(git -C "$d" stash list 2>/dev/null | wc -l)
  [ -z "$st" ] && [ -z "$un" ] && [ "$stash" = "0" ] && continue
  echo "$d: $(echo "$st" | grep -c .) dirty, $(echo "$un" | grep -c .) unpushed, $stash stashes"
done

Notes for the next session

Conventions observed in these repos

• Commit subjects are AI3D-<n>: <summary> or AI3D-<n> <summary> (both forms appear); older repos use AI-<n>.
• Default branch is inconsistent — master in some repos, main in others, feature branches in the rest. Check per repo; don't assume.
• All real remotes are git@bitbucket.org:ioholding/<repo>.git.
• Merges arrive via Bitbucket PRs ("Merged in … (pull request #N)") — so a PR may be the expected route for the Tier 2 work rather than a direct push.

Environment gotchas

• Over ssh battlebox "bash -s", never use sudo — the password prompt hangs the session to timeout. None of the work here needs it.
• Long git push operations over the Czech VPN can drop. Consider running on-box instead; the rebuild handoff compares the runner options (codex-bb, the cursor-agent already installed in this WSL, or ssh from Battlebook).
• /mnt/d is mounted and writable from WSL — confirmed 3.7 TB, ~744 GB free.
• The ~/dev working trees total 129 GB; linebitmapimagerasterizer alone is 63 GB and lanefinder 21 GB. Don't try to copy working trees wholesale.

Suggested skills

• battlebox-ssh — required if driving remotely.
• codex-bb-runtime or cursor-cli-runtime — if running on battlebox itself.

Open questions for Miro

1. tablecloth: does a Bitbucket repo exist for it, or does one need creating? What is the URL?
2. Tier 2: should guardrails / visualizationoverlays / trajectoryfilter go via Bitbucket PRs or direct pushes? And is the ~2.5-month-old trajectoryfilter change still wanted?
3. visualizationoverlays: the pyproject.toml + uv.lock change looks like an unreleased version bump after v0.2.0 — intended as a release?
4. lanefinder: confirm it's acceptable to keep 31 of the 33 stashes only in the /mnt/d/wsl-rescue archive.
5. Dead repos: confirm pytorchinstancesegmentation, pdfpropertyai and windowchecker are archive-only and their untracked output can go.