Beagle SCM

BE-061: a pre-seeded empty .be/ makes the FIRST jab get into a worktree throw ToBigInt in hashlet60FromBytes — a second get succeeds

Clone into a scratch worktree that already carries an empty .be/ directory — the shield test/lib/repo-setup.sh rs_fresh_wt/rs_shield plant in EVERY test worktree — and the first jab get dies with TypeError: Invalid argument type in ToBigInt operation from hashlet60FromBytes ← locate ← getObject ← commitTree ← fanoutWholeTree ← get. Re-run the SAME command in the SAME worktree and it succeeds, so the bootstrap leaves something half-written that the first pass reads as undefined. A worktree with NO .be/ at all clones correctly first time. This is load-bearing: rs_fresh_wt is THE shared repo-setup procedure, so anything built on it is one transport away from this. Found by the TEST-004 treadmill; method Issues.

Input

Context

• Repro (2026-07-27, jab build 22bec7d6), three lines: mkdir -p $HOME/tmp/wt/.be; cd $HOME/tmp/wt; jab get 'http://127.0.0.1:PORT/repo' against a smart-HTTP git upload-pack — first run throws, second run prints ?master#<sha> and checks out correctly.
• The stack, verbatim: TypeError: Invalid argument type in ToBigInt operation / BigInt@[native code] / hashlet60FromBytes@ / locate@ / getObject@ / commitTree@ / fanoutWholeTree@ / get@.
• Same command, worktree created as a BARE EMPTY DIR (no .be/): first run succeeds, tree correct. That is the only difference.
• test/lib/repo-setup.sh §"HOW this lib achieves isolation": an empty .be/ dir is documented as the fresh-bootstrap anchor — "the first be/sniff command bootstraps it into a real store in-place". That bootstrap is what misbehaves.
• test/get/http/run.sh does NOT hit it: getcase.sh makes $WORK/jT with no .be, so the passing ctest case takes the working shape.
• TEST-004 worked around it by making its scratch worktrees bare dirs and relying on the .be firewall FILE above them; that is a workaround, not a fix.

Goals

The first jab get into an empty-.be/-shielded worktree behaves exactly like the first get into a bare dir.

• hashlet60FromBytes never receives an undefined/short value — whatever locate hands it is a real byte range or a clean miss.
• The bootstrap-in-place path initialises the shard fully BEFORE fanoutWholeTree reads it, or locate reports "not in the store" instead of crashing.
• A user-facing failure here is plain words, not a raw JSC TypeError with a stack (the standing ruling on error wording).
• A repro test pins the empty-.be/ shape, since it is the shape every test worktree uses.

Constraints

• Do not "fix" it by retiring the empty-.be/ shield: it is the documented isolation anchor and the firewall depends on the pair.
• Do not paper over it with a retry — the second get succeeding is the SYMPTOM, not the cure.
• be/ JS only; the fault is in the JS read path, not in a binding.

WIP

Design decisions

• Unknown whether this is http-specific: the treadmill can only reach a vanilla git repo over the smart-HTTP wrapper on this box (no sshd by default, be:/file: want a beagle store). Re-run the repro over ssh before scoping the fix.

TODOs

• Minimal repro as a test first (the Issues rule): empty-.be/ worktree + first get.
• Find what locate returns on a freshly-bootstrapped-but-empty shard and why hashlet60FromBytes gets a non-BigInt-able value.
• Check whether the difference is a shard subdir that exists-but-is-empty vs absent (home_dir_no_subdirs reads the two differently).
• Confirm over ssh and over be:/file: whether the transport matters at all.
• Plain-words error on the genuine miss path, whatever the root cause turns out to be.
• Once fixed, revert TEST-004's bare-dir workaround to the standard rs_shield shape.

Blockers and bummers

• The second-run-succeeds behaviour makes this easy to miss in any harness that retries or that reuses a worktree across cases.

Outcome

Not started — filed 2026-07-27 from a TEST-004 treadmill finding.