Check out the default branch when the clone has no origin/HEAD

[thalida]

Loading https://github.com/PostHog/posthog rendered a completely empty city, and stayed that way on every reload.

What was wrong

The clone on disk had everything except a checkout:

HEAD                        → ref: refs/heads/.invalid   (dangling)
refs/remotes/origin/HEAD    → missing
refs/remotes/origin/master  → 7da1c469…   ✅
remote-tracking branches    → 13,114
working tree                → just .git/

refs/remotes/origin/HEAD is written by git clone and never by git fetch, so a clone that never got one has no local record of which branch is default. _resolve_default_branch read that missing ref as "the remote has no commits" and skipped the checkout — an inference that only holds for a genuinely empty remote. A repo with 13,114 branches took the same path as a brand-new empty one.

Skipping the reset isn't an error either, so the self-heal (update failed → discard clone → re-clone) never fired. Every reload re-fetched, re-decided "no commits", re-scanned 0 files.

Fixing that exposed a second bug. reset --hard moves whatever HEAD already points at; it does not repoint a broken HEAD. Files landed on disk while HEAD stayed on the dangling refs/heads/.invalid:

git rev-parse HEAD → fatal: unknown revision
git log            → fatal: your current branch appears to be broken

_run_git returns "" on failure, so the scan silently read zero commits — no dates, no ages, no trees, no timeline, while the city otherwise looked fine.

The fix

• Missing ≠ empty. When origin/HEAD is absent, re-derive the default: ask the remote what HEAD points at, else a conventional name among the branches already fetched, else the lone branch. Return None only when nothing was fetched at all.
• Record it back to refs/remotes/origin/HEAD so the repair is permanent.
• _reset_to() puts HEAD on a real branch before resetting when it isn't on a usable one.
• Guard the fresh-clone path — git clone exits 0 with nothing checked out, so the first load of such a repo would otherwise still be empty.

Verification

The regression test reproduces the exact on-disk shape (populated clone, origin/HEAD deleted as a symref, HEAD parked on refs/heads/.invalid, tree emptied). I disabled each half of the fix in turn to confirm both assertions bite — the checkout assertion and the git log assertion each go red on their own.

Against the real repo after the fix: branch: master, root created_min: 2020-01-23 (was the checkout timestamp), 42,969 files with 0 missing creation dates, only 38 falling back to mtime, spanning 2020–2026. The manifest grew 21.9 MB → 34.7 MB, all recovered history.

403 passed, coverage 90%. ruff clean. Frontend 2,992 passed.

Also here

One unrelated commit: the building lightness ramp's top end drops 50 → 40, so recently-modified files stop washing out toward white. Happy to split it out if you'd rather.

🤖 Generated with Claude Code