PARITY.md

Routine parity ledger

Oracle means the pinned locally compiled DNA5.dll invoked by the pinned RDP5CL.exe. All discrete state comparisons are byte-for-byte. Probability outputs use the user's accepted platform-math tolerance when Windows and WASM pow differ only in final rounding.

Execution order	Routine	Oracle boundary	Current result
1	SuperDistP	first live Dataset0 row call from RDP5CL.exe	byte-identical result and distance state
2	MakeAListP2	synthetic input, Windows DLL vs WASM	byte-identical
3	CountNucs	chained synthetic state	byte-identical
4	RecodeNucs	chained synthetic state	byte-identical
5	DoRecodeP	chained synthetic state	byte-identical
6	MakeCompressSeqP	chained synthetic state	byte-identical
7	AlistRDP4	live Dataset0 call from RDP5CL.exe	byte-identical result, RL, and StoreLPV
8	FindSubSeqPB3	live Dataset0 call from RDP5CL.exe	byte-identical result, AH, and window state
9	XOHomologyP	live Dataset0 call from RDP5CL.exe	byte-identical result and homology state
10	FindNextP	live Dataset0 call from RDP5CL.exe	byte-identical result
11	DefineEventP2	live Dataset0 call from RDP5CL.exe	byte-identical result and all scalar output state
12	ProbCalcP2	live Dataset0 call from RDP5CL.exe	numerically equivalent; one final-bit pow rounding difference
13	FindSubSeqPB4	live Dataset0 call from RDP5CL.exe	byte-identical result and position maps
14	FindFirstCOP	live Dataset0 call from RDP5CL.exe	byte-identical result
15	ProbCalcP	live Dataset0 call from RDP5CL.exe	numeric result gate
16	CleanXOSNW	live Dataset0 call from RDP5CL.exe	byte-identical cleared state
17	MakeTestPVs	live Dataset0 call from RDP5CL.exe	byte-identical event-test state
18	FindBestRecSignalP2	live Dataset0 call from RDP5CL.exe	byte-identical selected event and state
19	UFDist	live Dataset0 call from RDP5CL.exe	byte-identical interval distances
20	SuperDistP2	live Dataset0 call from RDP5CL.exe	byte-identical distance matrices
21	CheckMatrixP	live Dataset0 call from RDP5CL.exe	byte-identical matrix filtering state
22	MakeNJTreesP2	live Dataset0 call from RDP5CL.exe	byte-identical tree-building state
23	MarkOutsides	live Dataset0 call from RDP5CL.exe	byte-identical event marking state
24	MakeSDMP2	live Dataset0 call from RDP5CL.exe	byte-identical regional distance state
25	FillRmat	all three live Dataset0 calls from RDP5CL.exe	byte-identical correlation-matrix state
26	CalCR	all three first-event calls, fresh local-RDP capture on all ten datasets	byte-identical RCorr, RInv, and tRCorr state
27	MakeProperRCorr caller tail	first-event MakeRList input, fresh local-RDP capture on all ten datasets	byte-identical filtered correlations, inversion categories, and warning mask
28	QuickDist6 / MakeDMatS	first-event MakeRList warning input on all ten datasets	byte-identical downstream warning state; literal inclusive-window source port
29	MakeGoodC	first-event MakeRList input on all ten datasets	byte-identical overlap eligibility
30	MakeINList / MakeACOR	first-event MakeRList input on all ten datasets	byte-identical role map and acceptable-correlation mask
31	MakeRList	first live call from each of ten fresh local-RDP runs	exact row bounds, candidate membership, list order, and inversion flags; accepted platform-math tolerance for derived probability cells
32	FindActualEvents / StripUnfound	all three first-event role calls from fresh local-RDP runs on all ten datasets	exact candidate inputs, found-slot sets, overlap scores, breakpoint matches, swap-last pruning, and structural OKSeq state; accepted p-value tolerance
33	StripDupInv	first live call from each of ten fresh local-RDP runs	exact pre/post row bounds, list order, inverse removal, and inversion penalties
34	MakeLDist / MakeRCompat	every first-event call (6–24 calls per run) from fresh local-RDP runs on all ten datasets	byte-identical list distances, compatibility/reverse-compatibility scores, and non-recombinant list bounds across primary, tie-break, alternate-list, and retrim families
35	MakeDoneThis3 / MakePhPrScore	all first-event score calls from fresh local-RDP runs on all ten datasets	exact filter/trace state and numerically equivalent correlation/subscores
36	MakeTrpGroups2 / MakeTrpScore2	all three role calls from fresh local-RDP runs on all ten datasets	byte-identical group membership/counts and numerically equivalent triplet tree scores
37	FindSets and pre-retrim compatibility orchestration	complete first-event path from fresh local-RDP runs on all ten datasets	exact alternate candidate sets and exact conditional 6-, 9-, or 24-call FAMat/FCMat/SAMat/SCMat family order
38	CheckPatternX / CheckPattern score panel	first live call from each of ten fresh local-RDP runs	byte-identical per-sequence pattern counts and completion mask; integrated VB proportion-panel update
39	FinalTrim candidate maintenance	first retrim path from fresh local-RDP runs on all ten datasets	exact active candidate bounds and membership on every dataset that retrims
40	MakeVarMap2 / MakeCntHit2 / CalcMatchY	native DLL trace plus connected first-event path	byte-identical hit products and breakpoint smoothing state; both native grouping-threshold branches ported
41	ConsensusOK	post-retrim candidate state from fresh local-RDP runs on all ten datasets	exact active candidate bounds, membership, and list order
42	second FinalTrim strict-group block	first-event path from fresh local-RDP runs on all ten datasets	exact active candidate bounds, list order, synthetic RDP group records, and RDP rescan tail
43	MakeRelevant / MakeCollecteventsC	both parent calls for the first event from fresh local-RDP runs on all ten datasets	exact relevant membership, nonblank RDP working records, per-row counts, and collected parent-event outputs; accepted platform-math tolerance for p-values
44	FillFSSGC / FindSubSeqGCAP7	every initial GENECONV triplet through the first MakeTestPVs boundary	exact compressed informative-site stream, difference counts, and coordinate maps on all ten datasets
45	GetFragsP / GetMaxFragScoreP / CalcKMaxP / GCCalcPValP2	targeted native peak traces plus the complete initial GENECONV scan	exact peak order and accepted-fragment topology, preserving the independent GCDimSize=2999 fragment stride
46	GCGetHiPValP / DelPValsP / MakeDeleteArrayP / GCXoverD caller loop	complete first MakeTestPVs state with only RDP and GENECONV enabled on all ten datasets	exact counts, rows, event identities, ordering, and p-values except one accepted platform-math difference on Dataset8

The live AlistRDP4 fixture covers 2,300 triplets and changes 709 redo states. Because the exact C++ implementation calls FastRecCheckPB internally, that passing outer boundary also validates its complete internal chain collectively; the later rows isolate the same routines at their separately exported calls.

The CLI now constructs Seqnum, the complete triplet list, and CompressSeq directly from the supplied Dataset0 FASTA. Those three arrays are byte-identical to the live state passed by RDP into AlistRDP4; they are no longer supplied to that preprocessing gate by a fixture.

The CLI also ports the scalar VB/DNA distance setup preceding the first SuperDistP: site categorization, the seven NucXX categories, RDP's unusual fixed-width sequence compressors, the sequential SuperDistP row loop, and the caller's diagonal/low-validity normalization. Starting from FASTA, the resulting complete 25-by-25 distance matrix is byte-identical to the live matrix passed into AlistRDP4. The legacy helper bodies are implemented but remain collectively gated at this shared-state boundary until their individual DNA.dll calls are captured.

The following UPGMA block now starts from that generated distance matrix and ports MakeDistanceBakB, MakeDistMapX, ShortestDistB, AddSeqToUPGMA, UpdateDistMapX, and TreeDist2 in their VB call order. The resulting complete TreeDistance matrix is byte-identical to the live matrix passed into AlistRDP4.

With those generated matrices substituted for captured state, AlistRDP4 still produces the byte-identical redo list and StoreLPV. The first redo triplet selected from that list drives generated CompressSeq through FindSubSeqPB3, XOHomologyP, the intervening VB average-homology role ranking and distance tie breaks, FindNextP, FindFirstCOP when the circular start branch applies, and DefineEventP2. Rejected intervals advance through the source NextPosX loop and the probability-table prefilter until the first literal ProbCalcP or ProbCalcP2 branch is reached. The scan then preserves the source's two role transitions and its OldX state across cycles until the first corrected-significant interval is found. Generated AH, XoverSeqNumW, XDiffPos, and XPosDiff at the resulting FindSubSeqPB4 call are byte-identical on all ten datasets (with the accepted platform-math tolerance for final p-values). The port then exhausts the remaining two role scans and matches the complete XoverSeqNumW input and cleared output at the final CleanXOSNW call on all ten datasets; run tools/check-all-pre-scan-parity.sh for that table.

The same source-order walk now runs every triplet selected by the generated AlistRDP4 redo list. It ports the late FindSubSeqPB4 map construction, compressed-coordinate conversion, the XPDDone branch of CentreBP, RDP's CurrentXOver/StoreLPV event-owner choice, and UpdateXOList3's ordinary append path. Across all ten datasets, all 16,431 generated active records have the exact native daughter, parents, centered breakpoints, program/SBP flags, and zero-initialized metadata at the first MakeTestPVs boundary. Every per-sequence CurrentXOver count also matches. The generated arrays then pass through the exact MakeTestPVs and FindBestRecSignalP2 routines and select the same first event on all ten datasets. That selection is now exposed as a fixture-independent library routine, including the source's fallback from topology-changing (DoneTarget=0) to non-changing (DoneTarget=1) signals. Run tools/check-all-event-state-parity.sh for the current table. Probability cells retain the accepted Windows/WASM final-rounding differences and are not part of the structural identity gate.

The initial mixed-method path now follows the RDP scan with the literal compressed GCXoverD path. Its generated event state matches the locally compiled oracle on all ten datasets: totals are 1,400, 663, 987, 1,144, 775, 3,366, 6,008, 3,799, 3,797, and 4,556 respectively, with every row count and all 26,495 structural event identities exact. Dataset8 has one accepted last-bit probability difference; every other probability cell is exact.

The selected event now continues through the first phylogenetic test. The generated full-alignment state and selected-region state match the native UFDist and SuperDistP2 boundaries on all ten datasets. The port then executes the installed DNA.dll!FinishDists evaluation order, the VB CheckMatrixX masking prepass, CheckMatrixP, sequence removal and trace-list construction, and MakeNJTreesP2. Both generated NJ tree strings and all discrete outputs match on all ten datasets. Tiny platform log differences in otherwise equivalent stored Jukes-Cantor cells are compared at 1e-6; the delayed-rounding implementation is required because rounding the apparent source float temporaries changes the native tied-tree merge order.

From the same generated event and alignment, the port now also executes the literal circular MakeBPosLR flank walk, constructs the five VB correlation regions and fixed triplet comparison matrix, and reaches MakeSDMP2 with the exact native input state. Its complete SDM and DistMat outputs and the three sequential FillRmat branches are byte-identical on all ten datasets. The resulting matrices then pass through a direct port of all three CalCR calls. Its complete RCorr, RInv, and five-permutation tRCorr state is byte-identical to a reproducible installed-DNA.dll oracle capture on all ten datasets.

The connected path now continues through the literal VB tail of MakeProperRCorr, including SDM suppression, both warning passes, six-decimal tested-correlation rounding, and inversion-category remapping. It then ports the old-DNA QuickDist6, MakeDMatS, and MakeGoodC loops, constructs the compact three-row matrix panels used by MakeINList/MakeACOR, and executes a direct source port of MakeRList. Fresh first-call traces from the locally compiled RDP match all structural inputs and outputs on all ten datasets. The same connected path now executes Module3's active FindActualEventsVB wrapper, the compiled FindActualEvents body, StripUnfound, and StripDupInv, preserving their swap-last ordering and two confirmed legacy indexing bugs. It then runs MakeLDist and directly checks every observed MakeRCompat call, including the optional FCMat/SCMat, alternate-list, and post-retrim families (6–24 calls per dataset). The connected score path also ports MakeDoneThis3, MakePhPrScore, MakeTrpGroups2, and MakeTrpScore2. Every routine boundary matches fresh oracle runs on all ten datasets; carried probability and correlation-score cells use the approved 1e-12 tolerance.

The orchestration preserves two otherwise easy-to-collapse state values: RDP passes the FASTA length to FindSubSeqPB3 but the FASTA length plus one to XOHomologyP, and its initial tree-selected HighHomol can differ from the average-selected HighHomol later passed to FindNextP.

The connected path now also enters FinalTrim when the native compatibility scores request a retrim. It ports the active candidate-maintenance loop, MakeVarMap2, MakeCntHit2, CalcMatchY, and ConsensusOK, including the configured conservative grouping thresholds. The resulting active RNum and RList ranges match the post-retrim MakeRCompat boundary on all four demo datasets that take this path; the other six correctly skip it. As in VB, cells beyond each RNum are stale backing-array storage and are not observable list members.

Immediately after that retrim path, the clean port now reproduces the active CalcMaxD prelude and its CMaxD2P3 boundary. It constructs the same representative and included-sequence masks, informative-site nucleotide maps, pairwise V-score matrix, and accumulated maximum-distance values. Every input and output cell at that boundary matches fresh locally compiled RDP traces on all ten datasets.

The connected first-event path now also ports MakeSSDistB, SimpleDist, MakeOUCheck, MakeEList, MakeListCorr, GetBadDists, the two post-retrim MakeRCompat families, and the active decision-tree path of MakeConsensusC. Every generated input statistic agrees with the native NN_inputs boundary on all ten datasets at its printed precision, and the raw consensus scores and selected recombinant role agree on all ten. The CSV oracle is used only as a boundary assertion; no value from it drives the connected consensus calculation.

The first-event path now continues through the second, strict-group FinalTrim, its synthetic group-member records and relaxed primary-RDP rescans, MakeRelevant, and both parent-role MakeCollecteventsC calls. Every discrete RDP record and collected event agrees across all ten datasets. The exact AddjustCXO redistribution boundary is now also a reusable library routine rather than fixture-runner-only code; its retained event matrix, done-state matrix, and pair-rescan schedule remain covered by the same ten dataset gate. The immediately following inner- and outer-pair rescan screening is now also fixture-independent library code. It preserves the source AlistRDP3 call, including the flattened triplet order and correction threshold, and remains exact through both post-rescan event sets, the second MakeTestPVs/selection, and the second resolved-round prefix on all ten datasets. The entire identification path for one selected event is now exposed as a fixture-independent production routine. Starting only from live scan state, the selected raw event, and the current event list, it reproduces the source compatibility-family ordering, PhylPro and tree support, literal retrim condition, consensus winner, and strict final candidate groups on all ten datasets. No fixture or precomputed consensus value is read by this routine. Native zero-filled slots created by disabled-method bookkeeping are excluded from the RDP-event fixture because they contain no event and cannot overlap a real interval. The enclosing post-event sequence-mutation and cyclic rescan loop remains the next parity boundary.

The initial method stream now also enters MaxChi using the native dispatcher’s captured triplet order. Runtime traces establish that this path uses the fixed 70-site window and FindAllFlag=0: the first GrowMChiWinP2 call gets TWin=10, not the TWin=36 implied by the supplied VB caller’s assignment of FindAllFlagX=1. Reproducing the observed flag, the source CalcChiVals4P3 lookup path, compressed informative-site map, and initial CentreBP branch reduces Dataset0 MaxChi output from 1,286 port records to 636 versus 634 in the oracle. A direct trace of the first remaining divergent growth call then showed that circular MCXoverF does not use that lookup path for its initial scan: it calls formula-based CalcChiValsP, preserving a full-precision double peak value. The port's rounded float-table peak compared equal to a later window and reset GrowMChiWinP2's failure count, while the native peak compared strictly greater. Restoring the source's circular/formula branch removes that event and advances the first count divergence from call 41 to call 341. Dataset0 is now 2,033 port records versus 2,034 in the oracle, with 24 of 25 event-bearing sequence rows exact. The remaining call-341 miss came from collapsing MCXoverF's explicit MaxX = 1 opposite-boundary branch into a generic circular-position helper. That changed the DestroyPeaks interval from (1203,1) to (1,1492) and erased an adjacent peak. Reinstating the literal source branch restores Dataset0 to 2,034/2,034 records, 25/25 row counts, and 2,034/2,034 probabilities. Breakpoint/window structural identity then reaches 2,034/2,034 only after the active program-3 branch of FindDaughter and MCXoverF's second CentreBP call are preserved. Despite its name, this branch does not change the selected roles: it walks the informative-site map according to MCMaxY and LoHiFlag, conditionally polishes both endpoints, and recentres them. Dataset0's complete initial RDP+GENECONV+MAXCHI stream is now exact in total, row counts, structural identity, and probability.

Fresh mixed-method fixtures now cover all ten demos and can be checked with tools/check-all-rdp-geneconv-maxchi-parity.sh. Dataset0, Dataset2, and Dataset7 are completely exact. Current signal totals are: Dataset0 2,034/2,034; Dataset1 1,061/1,062; Dataset2 1,457/1,457; Dataset3 1,741/1,744; Dataset4 1,103/1,104; Dataset5 4,310/4,313; Dataset6 8,045/8,045; Dataset7 4,672/4,672; Dataset8 5,318/5,326; and Dataset9 5,763/5,764. Dataset6 has equal totals but 8,010 of 8,044 comparable structural identities; the other non-passing datasets still have one to eight missing signals plus later coordinate differences.