A snapshot is not byte-reproducible across processes

Index: Known gaps

Status (2026-10-31): NARROWED. SourceGraphSummary.finalize() (live and decoded graphs alike) and encode_graph_table now order node rows sorted by node id and edge rows sorted by relation_key() (both unique), so the interner indices and rows no longer follow in-memory insertion order. The producer of the insertion-order drift is still unidentified, but it can no longer reach the stored bytes. tests/test_snapshot_byte_reproducibility.py (bug class serialization.stored_bytes_depend_on_producer_iteration_order, family F9) checks every insertion permutation in-process and dumps a compiled corpus in three processes (PYTHONHASHSEED 1/4242/random, different cwd, SOURCE_DATE_EPOCH pinned) expecting identical bytes. Residual: the cross-process check runs on a generated ~300-entity corpus, not a 100k-node library like oneDAL; a different section with the same defect would only be caught at that scale. created_at differs by design unless SOURCE_DATE_EPOCH is set.

What was observed. Dumping the same library twice, with the same code, in two separate processes produces snapshots whose surface_graph.nodes differ in order: 9 positions out of 108,142 on oneDAL's libonedal_parameters. The node sets are identical (0 added, 0 removed), so no ABI conclusion moves; only the ordering does.

Why it matters anyway. It defeats comparing two snapshots by digest, and any content-addressed caching or reproducibility claim built on one. It is also a trap for anyone measuring a change by hashing output: a cross-process A/B will report a difference that the change did not cause. The haystack-deduplication measurement hit exactly that and had to be re-checked in one process (where the snapshot is byte-identical) before the "findings unchanged" claim could be made honestly.

Not investigated further here. The shape — a small number of adjacent positions permuted, stable within a process — points at set/dict iteration over values whose hashes vary with PYTHONHASHSEED, but the specific producer was not identified, and this is not a defect the surrounding performance work introduced: it reproduces with that work reverted.