Include Veloz through a general simple-fold set view
perfloop-oss/casei · UNCATALOGUED MECHANISM
https://perfloop.ai/t/oss/case_5j87wk3729
Verdict
VERIFIED · settled 2026-10-21 · merged as tsenart/casei#41
What happened: The assertion is satisfied with this change.
Hypothesis
Make the folded generality board include a general Veloz-backed comparison for substring and literal-set searches. The view must return the right first result or complete non-overlapping stream, so omitted competitors cannot flatter casei. Required coverage is all112 original folded Find, Each and IndexFold cells with nonempty validUTF8 patterns and input. Integrated conformance and charged scoreboard costs remain unproved.
At22ed84b, boardEntrant177–276 allows Veloz only on the ASCII-tier singleton route with its existing AVX2 availability gate. Pinned Veloze01f736adf9a333d0058c6a2a5bcfa8ddcb9f0f4 folds only ASCII letters while preserving high bytes. Minimum-scalar simple-fold keys contain no lowercase ASCII, making that byte fold identity on canonical UTF8. Retain one canonical literal per original ID, build the haystack view and exact boundary map inside each timed operation, merge earliest canonical starts and lowest original IDs, and project selected spans to original bytes. Resume native searches in canonical coordinates; refresh every pending start consumed by the selected endpoint.
The bound is112 Veloz entrant-cell obligations, with at most97additional after9current native cells and the6Veloz omissions already owned by hyp_hemyx8zyj8/case_mtwn2vqphj. Count overlaps once and preserve useful certified direct bindings. This is a named comparison-integrity prerequisite with zero removed casei CPU cost. hyp_qgng8cd31s/case_7n9hd97rqc owns RustAC's analogous view, not this backend's per-pattern search/merge proof. Share the canonical definition and available utility rather than add a separate proof-only helper Case; no result of either unmerged Case is assumed.
The parent independently tested all112 ordinary operations and bounded direct-orbit, scalar, byte-alias, stop, nested, concurrent and actual portable controls with Go1.26.6/Unicode15. Full conversion and maps can be expensive: single charged samples reach about91.8MB map storage and114.8MB total allocation. These are Research feasibility/resource observations, not native integrated timing, retained heap or a speedup. Construction, first-use and reuse costs must be disclosed; no silent expansion/resource cap is justified.
Only arena mode metadata, call-owned view/merge state, integration and tests are added. No public casei API, type or configuration changes and casei never delegates to Veloz. Keep exact cells unsupported by pinned Veloz, native sources unchanged, and raw ASCII semantics distinct from translated Unicode support. field.yaml21–35 permits timed adaptations; README3–11 and AGENTS145–191 describe active comparison work. Maintainer acceptance of this particular view is unknown. Do not globally declare other arena consumers supported without their own integration.
On nonempty validUTF8 inputs, minimum-orbit canonicalization makes pinned Veloz's ASCII-byte folding an exact search relation. A canonical-coordinate per-pattern merge and exact boundary projection preserve first-result priority, original IDs, non-overlap and source widths without borrowing invocation state.
Integrate the actual BenchmarkBoard adapter and independent support verifier on the original200cells with Go1.26.6/Unicode15 and verified digests. Require Veloz support for all112folded ordinary operations while preserving all other declared slots, all88exact cells and the existing availability policy. Independently exercise the translated view on all112cells even where a useful certified direct mode is selected. Strictly check complete ordered Each starts/original IDs/source widths and Find/IndexFold first results against an independent original-source orbit oracle; unreported quantities must not hide a quantity this view actually supplies. Require every projected endpoint to be a valid canonical rune boundary, including EOF, and keep canonical restart coordinates separate from source offsets. Required controls include all fold-orbit members, canonical duplicates, reversed prefix priority, Kelvin/long-s widths, sigma, four-byte folds, sharp-s without ss expansion, stale pending hits crossed by another pattern, dense/no-match and EOF streams. Check the fixed expected streams for patterns[ks,k,s] and[k,ks,s] on Kſk, [ab,ba,a] on aba and[ss,ß] on ßssẞ. Require stops1/256/257, repeated calls, nested callbacks and actual concurrent scratch ownership. Feature-off controls must preserve truthful unavailability rather than count a weaker field path. Charge every haystack conversion/map fill, native search, merge and projection; prohibit preconverted fixtures/results. Report constructor, first-operation, allocation, peak/retained scratch and reuse costs. Preserve current native sources/guards and exact exclusions. Independent verifier tests must reject missing supported slots, corrupted active ratios and unsupported flags without weakening largest-ratio/family/cell rules. Reject any language/order/width/stop/lifetime/integration/timed-cost failure, silent resource exclusion or row/seed/count gate. No benchmark victory is required.
These are private Research experiments, not implemented Case proof. Parent native/portable/race controls pass9directed,6000seeded cases(seed5693698),1,112,064scalar keys,65,536byte pairs at five lengths and8workers atGOMAXPROCS4. They are not exhaustive substring, malformed-input or protected-page assurance. An initial876-versus1 full-enumeration discrepancy was a probe-scope mistake: the retained Find fixture stores its first result only; first hits agreed. That diagnostic and the corrected ordinary-operation checks are retained. A missing goroutine brace and packaging shell expansion were corrected setup failures. The parent view also matches18Rebar stream digests, but Rebar integration is not this Case's proof. The adviser's separate Go1.27.2/Unicode17 probe reports remain explicitly unconfirmed as its implementation; no adviser timing is credited. Full-haystack conversion before an early Find or stop remains charged and may be slow. Empty/malformed/opaque patterns/input are outside this selected view, not changed casei semantics. Physical product/exclusivity, optimized scratch, integrated timings and final combined-revision leadership remain unknown.
Change to test: Add an arena-only, honestly labeled Veloz view that searches one canonical simple-fold needle per original pattern, merges first hits and projects source spans. Integrate folded-board support and independent verifier expectations, charging every conversion, per-pattern search, merge and projection while preserving useful certified direct modes and existing availability policy.
Where it lives
perfloop-oss/casei · arena/board_bench_test.go
Evidence
The assertion is satisfied with this change: `For all 112 original folded BenchmarkBoard cells with nonempty valid UTF-8 input and patterns, when Veloz's AVX2 route is available, the translated Veloz view matches an independent original-source SimpleFold oracle on first Find/IndexFold hits and complete leftmost, lowest-ID, non-overlapping Each streams, including exact source-byte widths and valid canonical rune-boundary projection, including EOF.`
Checks: 3 of 3 passed. Verification: no defect found.