[upstream-sync] fix(engine): scope GC-sweep read transaction to list_tombstones only (upstream #2004)

[github-actions]

Upstream Change Summary

Type: bug-fix
Difficulty: Easy
Recommendation: Adopt

Committer::launch_child_component_gc previously borrowed a read transaction from commit() and held it open through the entire spawn-and-await loop over child delete tasks. The fix opens a new read transaction scoped strictly to the list_tombstones call and drops it before spawning child tasks. With backends whose ReadTxn pins a shared resource (e.g., a fixed-size connection pool), holding the rtxn across async spawning could deadlock once cascade depth × per-level fanout exceeds capacity. LMDB's single-writer design doesn't surface this today, but the lifetime discipline is correct regardless and future-proofs the engine for alternative state-store backends.

Upstream References

• PR: refactor(engine): scope GC-sweep rtxn to list_tombstones only cocoindex-io/cocoindex#2004
• Upstream repo: https://github.com/cocoindex-io/cocoindex

Relevant Upstream Files / Areas

• rust/cocoindex/src/engine/ — Committer::launch_child_component_gc, commit()

Recoco Considerations

• Affected files: crates/recoco-core/src/engine/ (committer module, GC sweep)
• Small, targeted lifetime fix — open a fresh RoTxn inside launch_child_component_gc, use it only for list_tombstones, drop before the async spawn-and-await block
• No Python code to exclude; no feature-gating required
• Aligns with recoco's correctness-first philosophy and prepares for potential future non-LMDB state-store backends

Integration Notes

This is a 5–10 line change in the committer. Identify the equivalent launch_child_component_gc function in recoco's engine code and verify it has the same rtxn lifetime pattern before applying. The fix is purely a scope narrowing — no API or behavior change.