v0.25.0 — performance, memory, and runtime hygiene
v0.25.0 banks the first post-correctness performance and repository-hygiene tranche.
Highlights:
- Pending-parent child entries now carry exact packed field metadata, and GSS dedupe verifies pending descendants recursively instead of trusting coarse hashes.
- Java large-file accepted-error retries are exact-blob certified: the useful cap-16 merge pass remains, while two redundant cap-64 passes are skipped. The rebased DLBF witness improved wall time by 21.3%, with identical Go output and non-regressive realistic RSS.
- Current main re-certifies the 3,447,275-byte Poppler witness with exact deep C parity and a clean hard-2GiB run at 1,708,712 KiB max RSS.
- Retired zero-only no_alias telemetry and two test-only wrappers are removed.
Compatibility note:
This minor release intentionally removes the exported zero-only ReduceChildNoAlias and related ArenaBreakdown and PerfCounters fields. They had no production producer and always reported zero.
Still open:
Poppler remains 3.50x C, and JavaScript remains pending on throughput. Profiling identifies rejected full-node construction as the next memory lever; the existing pending/transient coexistence modes are not suitable replacements.