# Performance and real-database delivery Accepted scope: v0.0.7 performance evidence, v0.0.8 graph/allocation improvements, v0.0.9 bounded optimized completion and v2 evidence, v0.0.10 large reduction, v0.0.11 actual PostgreSQL capture, then v0.1.0 API stabilization. The October 2026 official public page requires MoonBit as the primary language, traceable public development, runnable examples/tests, substantive new work, open-source compliance and explainable AI assistance: https://moonbitlang.github.io/Hackathon2026/ These performance goals are our engineering goals, not organizer scoring weights. No award, organizer acceptance or registration is inferred from a release. ## Reproduce measurements `npm run benchmark -- --baseline PATH_TO_V006_CORE` measures the same fixture with full JSON input and full JSON output on both engines, with five warmups, 30 observations and three independent process batches in alternating order. The original compiled v0.0.6 core can be extracted from its public portable ZIP. The local default is `artifacts/baseline-v0.0.6/core.mjs` (not redistributed). `--quick` is a smoke check and cannot satisfy the performance release gate. Receipts in `verification/releases/VERSION/` record fixture hashes, raw samples, runtime/CPU/OS, cold module load, first analysis, P95 and process peak RSS. Phase hooks time parse, dependency inference (including validation/coverage), SCC/shortest cycle, legacy completion, reduction, and full report serialization independently, with already-prepared input outside each timed section. They are development hooks, not stable SDK interfaces. Rendering in v0.0.7 uses its original structured serializer; later versions measure their shipped streaming serializer. Phase results on a no-cycle report measure the specified isolated operation, not a complete user-visible result. Targets: graph group geometric mean >=2x and RSS ratio <=0.7; search and reduction >=2x; control P95 regression <=max(10%,2ms). Summary output must not be compared with a full baseline report. A correctness pass does not mean all performance targets passed. Raw failed measurements are retained. ## Stable boundary Versioned point reads/writes only. Preserve existing default budgets and v1 evidence traversal. Advanced profiles and v2 evidence carry explicit effective budgets and algorithm identifiers. Exhaustion and a negative proof are distinct. The author must still perform the documented installation/demo/explanation exercise; registration, group membership and organizer decisions remain separate. ## v0.0.8 implementation Edge equality and hashing cover every provenance field, including the reason. No transitive dependencies are dropped. SCC and BFS use integer endpoints with stable adjacency/root order and epoch-marked reusable buffers. Coverage and model comparisons reuse the writer index. Full dependency arrays, per-key coverage and version maps are streamed with official JSON string escaping and a per-report escaped-string cache. Duplicate-key scanning still decodes escaped keys and uses the official JSON parser for syntax validation. The v0.0.7 exhaustive-failure capsule is retained as a legacy compatibility sample. Performance receipts include intermediate attempts that missed the target; do not select the older smoke receipt as final release evidence. ## v0.0.9 results The full balanced search group is 3.69x faster than v0.0.6 in this local measurement; the 8-transaction hard case is 10.36x. All 10,320 histories are independently checked through both legacy and balanced entry points, plus eight 9-12 transaction proof fixtures. This search release does not claim all roadmap performance goals are complete; its balanced graph group measured 1.92x, and large reduction comes next. ## v0.0.10 accepted measurement The final full protocol in `verification/releases/0.0.10/release-gate-v010.json` measured graph 2.665x, search 5.023x and reduction 2.178x geometric-mean speedup against v0.0.6, with graph peak RSS ratio 0.6972 (30.28% lower). Control P95 regressions stayed within the stated allowance. All full semantic comparisons passed. Earlier misses are retained, including the initial ASCII-encoder attempt; it was removed rather than using that attempt as successful release evidence. RT ordered pairs are unique by validated transaction IDs and no longer allocate redundant hash-set entries. SCC, shortest cycle and topological order share one immutable integer adjacency. Unchanged initial-only reads skip unnecessary local maps, while all transactions remain in graph and RT accounting. ## v0.0.11 actual-database acceptance The local PostgreSQL 18 suite completed 20 RR rounds and 20 Serializable server-abort/fresh-retry rounds, including a second run on a newly initialized cluster with a clean hash-verified optional driver. Both histories, 100 observed transactions, SQL events, two-transaction reduction and three v2 capsules were checked. The owned temporary cluster was stopped afterward. Independent Linux PostgreSQL CI is a required separate public release check. The analyzer algorithms are unchanged from v0.0.10; this version includes a smoke-only benchmark and references the accepted v0.0.10 full protocol rather than claiming a new full run. ## v0.1.0 convergence The frozen balanced profile passed the full protocol: graph 2.820x, search 5.557x and reduction 2.215x; graph peak RSS ratio 0.69977. Full scenario tables and limits are in performance.md. Legacy entry points remain unchanged; the portable demo recommends balanced and exports v2. Fixed fixture JSON, pinned public baseline fetching and immutable timestamped measurement receipts make retesting self-contained after the documented toolchain setup. All five small versions have public download and registry-consumer proof. The v0.0.11 first Linux workflow initialization failed; its immutable source was subsequently tested by a repaired workflow with separately attested source and workflow commits. This history is retained. The current publisher waits for Windows/Linux, Native and PostgreSQL CI before package publication. Participant installation/explanation and organizer states remain independent requirements.