Changelog

Current alpha state · October 4, 2026

The product has only paper-derived basic and multilevel modes, both routed through SQLite FULL-WAL Durable and the thread-safe bounded Service. Atomic multi-update batches, same-sequence retries, bounded coherent read snapshots, backups, and deterministic operation-log recovery are implemented. The former native matching engine and its graph-checkpoint format are removed. Hosted Linux qualification now exposes unresolved memory failures: Multilevel startup exceeds the 512 MiB address-space cap; Basic completed one million updates (~9,054/s) and backup but failed disk-pressure recovery while rebuilding System indexes. The 10k/s target and production readiness remain unproven.

Reliability and storage

Some pressure and backup subtests pass, but the full paper-mode resource gate does not. Hardware power-loss qualification is deferred for this version. Billion-vertex capacity is not measured or claimed. See current contracts, raw evidence and limits.

Product surface

• Added native routes for getting started, the playground, concepts, modes, API reference, examples, architecture, research status, and contribution guidance.
• Removed the browser simulation and scroll animation scripts. Replaced mock terminal tabs with a verified durable Service example and an exact deletion/insertion/query/reopen walkthrough.
• Changed the default palette to ivory, charcoal and teal; fixed collapsed-menu spacing and added keyboard menu dismissal.
• Made the first-use path executable in CI, including installation, updates, matching inspection, partner lookup, statistics, and maximality checks.

Engineering boundary

• Canonicalized the public API around basic and multilevel; legacy names and compatibility wrappers are not supported.
• Added atomic update validation and rollback-safe state exposure across graph, matching, index, and hierarchy state.
• Hardened recursive schedules, graph identity, deferred deletion checks, and deterministic verification.
• Durable basic/multilevel integration is implemented; exact failure-boundary, skew, scale, and installed deployment qualification continues. Historical native throughput does not qualify the paper algorithms.

For source-level history, use the repository as an engineering reference after starting with the public documentation.