Beagle SCM

GET-020: be get file:<wt> leaves an incomplete checkout + SNIFFFAIL

The worktree-form clone be get file:/home/gritzko/beagle (the Issues worker setup: a store-backed wt sharing ~/beagle/.be) intermittently exits SNIFFFAIL with a HALF-POPULATED tree: abc/ empty (submodule not mounted), woof/ and 3 test/tree/ files absent — yet most of the tree is present, so it looks done. The DIFF-003 worker hit this, then silently restored the missing build-only dirs from the canonical tree to make cmake work — a workaround that should have been THIS ticket, not a silent patch (Issues / CLAUDE §17). Distinct from GET-018 (atomic rollback ON failure): here GET reports failure yet leaves partial state.

Input

Context

• be get file:/home/gritzko/beagle into a fresh $HOME/todo/DIFF-003/ → SNIFFFAIL; abc/ empty (sub unmounted), woof/ + 3 test/tree/ files missing; cmake config then failed on the gaps.
• DIFF-004 worker (2026-06-15) found it DETERMINISTIC: be get file:<wt> checks out 1171/1200, an IDENTICAL 29-file miss across fresh re-runs (all of woof/, some test/spot/07-10, test/tree/), SNIFFFAIL each time; re-running does NOT converge. So NOT mere contention (earlier theory).
• Objects ARE in the store: per-file be get <path> materializes every missing file (→ 1200 ok). So this is a BULK-checkout defect, not missing data — a specific set of paths is silently skipped by the multi-file walk.
• Recovery used (legit, not a silent patch): complete the tree via per-file be get <path> (real store checkouts), verify 1200 ok, then build.
• 2026-06-16 PTR-sweep batch (9 wts off one ~/beagle/.be): 7 of 9 truncated to 474/528 test/ files, a CLEAN tree-order cutoff at test/put/05 (ALL of test/spot+test/tree+test/put/05… dropped, mis); woof/+abc/ present here, so the dropped SET is NOT fixed across runs.
• The 2 complete wts (528) are exactly the ones whose worker re-ran be get → a second pass filled the tail. So re-run makes PROGRESS (474→528), not the "doesn't converge" earlier theory; cutoff is a moving truncation point.
• Live repro preserved at $HOME/todo/PTR-001 (do NOT delete) — truncated wt, be status shows the 54 mis. Consequence: truncated wts ran ctest over a REDUCED set — 286 vs 298 in complete wts (~12 spot/tree/put cases silently never registered); their "green" excluded those.

Goals

• be get either produces a COMPLETE checkout (every tree file + mounted subs) or fails leaving the wt visibly unusable — never a half-populated tree that reads as done.
• A SNIFFFAIL mid-checkout must surface WHICH step failed and not silently drop files / skip the sub mount.

Constraints

• Shares ground with GET-018 (atomic wt GET) — the fix may reuse a stage-to-temp + atomic-swap. The empty abc/ ties to the submodule mount step (SUBS); both must be covered by the complete-or-fail guarantee.
• No abc/ edits; this IS the dogfood worker-clone path.

WIP

Design decisions

• ROOT CAUSE FOUND (2026-06-16, fix-ready): keeper/WALK.c::walk_tree_dive recurses into subdirs with a PLAIN C call while each level a_carves a 1 MiB tbuf off ABC_BASS — bare recursion never rewinds, so ~1000 dirs pile up to the 1 GiB arena cap; the next carve returns BNOROOM mid-walk → WALK return → SNIFFFAIL. Trace: BASS climbs 21→1023 MiB, breaks exactly at test/put/05 with free=0, vo=BNOROOM, ctx.error=OK (no file-write error). Same class as BLAME-004's blame-bignoroom. NOT the atomic-swap hypothesis.
• FIX: wrap the recursion try(walk_tree_dive,…); vo=__; — try() snapshots BASS after the parent's tbuf carve and rewinds per subtree, so only the live root-to-leaf chain stays carved. One line; covers WALKTreeLazy (get) + WALKTree (blame/diff). ("empty abc/" was downstream of the truncation — NOT a 2nd bug; abc mounts once the walk completes.)
• Updated dx (2026-06-16): the miss is a CLEAN tree-order TRUNCATION (writes files in order, stops at a point), NOT a scattered fixed set — cutoff moves by run/store but is identical across same-batch wts, and a re-run resumes from there (474→528). Strong tell of a count/cursor/batch-cap stopping the walk after N entries (or a buffer fill / swallowed mid-walk error), not the pool.
• Hunt the bulk-checkout walk (sniff/GET.c checkout + any ROWS/HUNK batch cap it drains through); per-file be get <path> works → compare the single-path path vs the bulk walk for where the tail is dropped + why SNIFFFAIL is the only signal.

TODOs

• Repro: single be get truncates at test/put/05; BASS climbs to 1023 MiB then BNOROOM. Repro preserved at $HOME/todo/PTR-001 (read-only).
• Fix: try()-wrap the walk_tree_dive recursion (BASS rewind per subtree); fresh be get now 1441 tracked / 0 mis, abc mounted, exit 0.
• Test: WALKtest6 (1300-wide sibling tree) — FAILS pre-fix (BNOROOM), PASSES post-fix; full suite 298/298.
• Land: gate the diff (keeper/WALK.c + keeper/test/WALK.c), then ff-post.

Blockers and bummers

• Intermittent / likely concurrency-dependent → repro may need parallel invocations against a shared store.

Outcome

• LANDED 49039f36 (ff, parent 229d122b): keeper/WALK.c try()-wrap of the walk_tree_dive recursion + keeper/test/WALK.c::WALKtest6. 298/298; a fresh be get now checks out the COMPLETE tree (1204 ok, 0 mis).
• Spun off: sibling-sniff PATH-fallback masks which binary runs when bin/sniff is absent (HOMEResolveSibling) — distinct dev-ergonomics hazard, to be filed as its own ticket.