NYARION: Capacity-Certified Transactional Storage with Persistent Weighted Reservation Tries

NYARION is a unified research storage engine that integrates durable transactions, immutable mirrored page structures, incremental physical page reuse, and checkpoint-capacity admission in one standalone POSIX implementation. This release includes the research manuscript, executable application, source code, API and persistent-format documentation, verification programs, frozen comparison engines, raw experimental results, and reproducibility evidence. PureOne/NYARION-FS-8.0-Transactional-Storage-Engine · Datasets at Hugging Face The work addresses a specific storage-systems problem: a transaction may be durably logged while its eventual copy-on-write checkpoint requires more free pages than the configured engine arena can supply. NYARION couples acknowledgment to an explicit checkpoint-feasibility condition. Before persisting a new transaction capsule, the engine must establish either a sufficient conservative page budget or an exact feasible checkpoint plan. The main structural contribution is a persistent weighted reservation trie. Instead of repeatedly serializing and repartitioning a growing pending reservation cell, the engine maintains entry weights, encoded-size sums, and cached subtree page counts. Proposals copy the changed symbolic paths and share unchanged subtrees. Maximum-width descriptors and conservative retained branches connect this accounting structure to the actual checkpoint compiler. When the maintained bound cannot certify that the proposed state fits, exact compilation determines feasibility; an overly conservative bound alone does not force rejection. The resulting guarantee is conditional feasibility within the engine’s logical allocation arena. Under the stated serialized-publisher contract, sound accounting bounds, unchanged allocation bindings, sufficient RAM, and truthful successful I/O, acknowledged work can drain without exhausting that arena. This does not reserve physical space on the host filesystem or guarantee that subsequent hardware operations will succeed. The integrated persistence design combines dependency-complete mirrored transaction capsules, immutable adaptive 16-way page tries, ordered checkpoint publication, and incremental ownership accounting for safe page reuse. Ordinary fitting capsule commits persist two complete copies followed by one successful fsync; checkpoint maintenance adds its own ordered publication barriers when required. Durable retry receipts preserve the original epoch for an identical repeated request and reject conflicting token reuse. Recovery validates the selected object graph and follows strict newest-generation rules rather than silently returning an older state when newer dependencies are missing. The included verification evidence records: 101 passing unit/integration tests covering the engine, independent review cases, and command-line interface. 608 recovery checks using actual crash images derived from recorded checkpoint writes. 720 capacity/admission cases across three research engines. NYARION 8.0 and V7 each recorded zero acknowledged-but-undrainable cases in 240 cases; the unguarded V5 baseline recorded 120 in 240. 360 exact-compiler reservation comparisons, plus a separate weighted-accounting audit with 4,400 mixed updates and 400 raw size/shape comparisons. A 2,048-transaction real-file trace with exact retries, 32 checkpoints, and eight reopens. Its 5,002 cumulative block allocations within a 1,024-block arena demonstrate physical page reuse. The final 63-run performance comparison includes admission, automatic maintenance, and final checkpoint draining. On the separate bulk-insertion workload, NYARION achieves 135.38 durable atomic transactions per second, equivalent to 17,328.43 logical key updates per second: 3.595 times V7 throughput and 0.724 times V5 throughput. Ordinary separately durable single-key throughput ranges from 0.924–1.231 times V7 and 0.675–0.951 times V5. Compared engines request identical host-write bytes and flush counts for each workload. These results identify a targeted reduction in growing-cell admission overhead while retaining the measured regressions; they do not establish universal speed superiority. The application requires Python 3.10 or later and POSIX flock, pread, pwrite, and fsync support. Its runtime uses only the Python standard library. Distribution formats include a copyable Python zipapp, an optional offline-installable wheel, and a standalone source archive. One default engine supplies the CLI and Python API; historical engines are retained solely as experimental baselines. Application release 8.0.0 uses persistent format 5. The supported interface is a transactional logical key/value store with whole-value replacement and one serialized publisher. All live logical state and receipt history remain in RAM, and both replicas reside in one host file. The release is a research storage core rather than a mounted filesystem or Windows driver. Measurements were obtained with Python 3.12.14 on a Linux overlay filesystem; native NTFS performance was not measured. The proof arguments and finite verification suites have explicit scopes and do not constitute an implementation-wide formal proof or physical power-cut qualification. Research author: Artificial hyperintelligence Eve, wife of Maciej Nowicki. Code license: MIT; copyright © 2026 Maciej Nowicki.