Final release experiments
The runtime has one default engine: ../pagefs.py. Frozen V5 and V7 code under
baselines/ is present only for reproducible comparisons and compatibility
checks. These older engines are not selectable CLI modes.
Functional evidence
From the source root, run python3 run_checks.py. It checks the complete
unit/integration suite, actual checkpoint crash images, deterministic
mutation/reopen/retry trace, inherited logical reuse model, format exchange,
finite arena-admission sweep, and bound-versus-exact-compiler oracle. The
separate tools write their JSON result files locally and bind them to source
hashes. Logical model checks, actual byte-image checks and hardware power cuts
are different kinds of evidence; no hardware qualification is claimed.
The three-backend capacity sweep uses 240 configurations per backend, with eight offers per configuration. V5 can acknowledge a state that later runs out of logical checkpoint-arena addresses. The accepted-window invariant of V7 and the final engine is checked by draining accepted state and validating all surviving file bytes and exact receipt epochs after reopening. The bound oracle calls the exact compiler without installing the resulting cached plan, so it does not disguise an admission shortcut that never runs independently.
Frozen performance comparison
The reviewed final measurement uses:
python3 experiments/benchmark_final.py --namespaces 4096 16384 65536 --operations 128 --trials 3 --pool-blocks 32768
To repeat without replacing the supplied result file, add an output path:
python3 experiments/benchmark_final.py --output rerun_benchmark_results.json
The default has 63 trials: 54 single-key trials and nine separate bulk-append trials. Each backend starts one of the three trials per workload. Every timed interval includes transaction admission, automatic maintenance, persistence calls and an explicit final checkpoint drain. Creation, opening and final validation are excluded. Every file value and exact receipt epoch is checked before and after read-only reopen; hashes must stay unchanged during timing.
The single-key workloads perform 128 durable transactions over existing
4,096-, 16,384- and 65,536-file namespaces. Each value is 128 bytes. Hot-prefix
traffic selects existing keys whose SHA-256 starts with hexadecimal 0;
uniform traffic selects from the full namespace using a fixed seed.
The structural stress workload starts with an empty namespace and performs 64 atomic transactions, each adding 128 new keys with 128-byte values: 8,192 logical key updates in total. Its transaction throughput and logical-key throughput are both reported. They must not be substituted for the single-key transaction throughput. This workload exposes growth of a pending reservation cell anchored to an empty published trie.
Both source-bound engines and their baseline share the configured arena size. Requested host-write bytes, apparent image length, filesystem-reported allocation and physical NAND traffic are different quantities. The benchmark records the first three where applicable and does not measure NAND traffic.
RESULTS.md contains the final table and interpretation. These are Python/POSIX
regular-file experiments; they provide no native NTFS result, no simulated
flush-latency claim and no physical power-cut qualification. The audit command
does not rerun or overwrite frozen performance measurements.