native-siscan-discovery-trace.md

Supplied-source SISCAN discovery and confirmation trace

This note records the source path used for the browser port. It was traced only from the supplied RDP5 Visual Basic project, dnaDLLSource, dna5DLLSource, and manual. No alternate RDP implementation was consulted.

Native entry points and defaults

The ordinary scanner calls Module3b.bas::SSXoverC. The equivalent longer implementation remains in Module3.bas, while Module30.bas contains the manual/plotting path. Default setup in Module5.bas is:

• SSFastFlag = 1;
• SSGapFlag = 0 (strip gaps);
• SSVarPFlag = 2 (one-, two-, and three-variable categories);
• SSOutlyerFlag = 1 (nearest fourth sequence);
• SSWinLen = 200, SSStep = 20;
• SSNumPerms = 1000 for the final region and SSNumPerms2 = 100 for scanning;
• random seed 3 in the standard configuration.

The desktop program leaves primary SISCAN discovery off in the normal automated selection but uses it as an ordinary later confirmation method. The browser therefore defaults primary SISCAN off and fixed-region confirmation on; both are explicit settings.

Call mapping

Supplied source	Browser implementation	Retained behavior
Module3b.bas::SSXoverC	siscan_discover	Window screen, pair-switch region construction, boundary shrink, final region scoring, role geometry, and threshold gate
threshold.CPP::MakeDistanceBakB plus RDP tree preparation and threshold.CPP::GetSSOL	build_source_wpgma_context and nearest_source_outlier	Direct similarity matrix, source-shaped WPGMA/cophenetic context, and nearest eligible fourth sequence
threshold.CPP::Get3Score	build_triplet_categories	Gap-stripped triplet pattern categories
threshold.CPP::GetPScores2	pattern_category and count_patterns	Fourth-sequence conversion into the exact 16 partition categories
Module2.bas::SetUpSiScan	compile-time category tables in siscan.cpp	Exact Seq34Conv, partition groups, and summed groups used by default variable-pattern mode
threshold.CPP::MakeVRand	ensure_vertical_random_prefix	Microsoft CRT rand() stream and flat byte-prefix template
threshold.CPP::DoPerms3 / dna5 MathFuncsDll.cpp::DoPerms3P	permute_patterns	Category-constrained vertical randomization with the scan/final permutation counts
threshold.CPP::MakeZValue2	source_z_score	Population variance and zero-variance behavior for partition and summed scores
threshold.CPP::DoSums	summed-score construction in permute_patterns	Exact 15 summed categories from the 16 partition counts
threshold.CPP::QuickCheckB	quick_check_window	Fast-screen control flow, including the supplied missing-braces quirk
threshold.CPP::FindMaxZ	strongest_window_pair / strongest_region_score	Highest-pair comparison and eligible alternative partition/summed score
threshold.CPP::ShrinkRegionC	shrink_region	Conversion from consecutive switched windows to nucleotide bounds
Module30.bas::NormalZ	source_normal_z	Supplied polynomial normal-tail approximation
late SSXoverC calls from FinalTrim	siscan_recheck	Nearest-outlier full fixed-region score without changing reconciled coordinates

Probability stages

The browser preserves the desktop stages as separate evidence rather than collapsing them:

1. normalTailPValue = NormalZ(maxZ);
2. regionLengthAdjustedPValue = normalTailPValue × alignmentLength / regionLength;
3. windowAdjustedPValue = regionLengthAdjustedPValue × alignmentLength / SSWinLen;
4. when project correction is active, multiply by the frozen initial scan-plan opportunity count;
5. clamp the reported probability to [10^-300, 1] so a finite nonzero audit value survives ordinary floating-point underflow.

This separation is important when comparing very small values. The port evaluates the polynomial and products in double and retains a positive subnormal where possible. A desktop build may round or underflow at a different stage; such a numerical difference is not treated as evidence that the breakpoints changed.

Caches and cyclic invalidation

The desktop supplies the distance/tree context and vertical-randomization template to repeated SSXoverC calls. The browser keeps the same lifetimes explicitly:

• the direct and WPGMA/cophenetic matrices are built once per cyclic round and reused by every SISCAN triplet;
• the seeded MakeVRand byte stream is an extensible flat prefix, so repeated windows and later rounds reuse identical bytes rather than regenerating them;
• accepting an event invalidates only the state-dependent distance/tree context;
• starting another analysis in the same browser context resets cache counters while retaining the deterministic prefix, so progress never includes telemetry from the preceding dataset;
• v1alpha19 project restoration round-trips the WPGMA build/comparison/merge and generated-random counts as audit data; cached matrices themselves are deliberately reconstructed when work resumes;
• the existing XOverList/BestXOList-style triplet shortlist skips SISCAN entirely for an unchanged triplet whose method summary can be replayed.

This keeps the expensive context at O(N²L) + O(N² log N) per affected round rather than per triplet. Per-window work remains proportional to retained category sites times the requested permutations.

Browser safety normalization

RDP5 can re-enter fragments created by previous erasure. The browser excludes any proposed fourth sequence whose original-sequence identity matches one of the three triplet members. This is the TraceSub intent made explicit and prevents a fragment from serving as its own outlier. Disabled original sequences are also ineligible. The source scans linear SISCAN windows even for circular datasets; circularity is retained for event geometry and wrapped fixed-region confirmation, not by inventing an extra origin-crossing scan window.

The desktop manual plot exposes all 15 partition categories and 9 category sums. The ordinary browser review endpoint is intentionally smaller: for each of the three sister pairs and each window it selects the eligible partition/summed value with greatest absolute Z and preserves that value's sign. This provides a signed diagnostic without claiming full manual-plot parity.

Validation boundary

scripts/verify-siscan-core.cpp deterministically plants a sister-pair switch and checks outlier selection, roles, region coverage, probability stages, cached WPGMA reuse, bit-stable random-prefix replay (including the first 20 Microsoft-CRT seed-3 bytes), fixed-region confirmation, review plotting, round invalidation, and same-context analysis restart. It is a port regression, not native parity evidence. Authorized native saved-output fixtures are still required for exact window maps, category Z scores, final bounds, and every probability stage.