native-maxchi-recheck-trace.md

Supplied-source trace: MaxChi strongest-peak recheck

This document records the Session 9 MaxChi translation made only from the supplied DNA/DNA5 DLL source, VB application source, and RDP5 manual. No alternate implementation or reference-code project was consulted. The implementation is intentionally labelled source-shaped and unvalidated. This document describes the bounded corroboration kernel; the subsequently added ordinary exploratory discovery path is documented separately in native-maxchi-discovery-trace.md.

No source in this checkpoint was compiled or executed.

Native call path

The supplied workflow establishes the following path:

Supplied source	Native responsibility	Session 9 mapping
VB Source/Module30.bas, GetCriticalDiff, around 9598–9690	Derive a cheap absolute match-count threshold from LowestProb / 6, with a 0.0001 floor	source_critical_difference; strict screen remains separate from χ² evaluation
DNA5/MathFuncsDll.cpp, FindSubSeqCP / FindSubSeqMCPB2	Retain sites where all three sequences are usable and at least one state differs; map alignment ↔ variable-site coordinates	build_variable_profile, with one coordinate vector and three compact pair-match tracks
DNA5/MathFuncsDll.cpp, MakeWindowSizeP, around 18577–18623	Start from the configured fixed window, shrink for short profiles, reject half-windows below six	fixed_window_sites = 70 and the same fixed-window fallback arithmetic
DNA5/MathFuncsDll.cpp, WinScoreCalcP, around 4631–4705	Build circular rolling match totals for all three sequence pairs	strongest_peak, maintaining six totals and updating them in constant time per variable site
DNA5/MathFuncsDll.cpp, MakeBanWinP, around 4554–4629	Ban windows starting, ending, or traversing MissingData; retain distinct boundary indices zero and length	make_banned_windows, including zero/length aliases and the trailing origin ban
DNA5/MathFuncsDll.cpp, CalcChiVals4P3 / CalcChiValsP, around 6358–6407 and 6675–6712	Apply the strict critical-difference screen, calculate three 2×2 χ² tracks, and find the maximum	chi_square plus the rolling strongest_peak pass
DNA5/MathFuncsDll.cpp, NormalZ, around 4210–4259; ChiPVal2P, around 7494–7516	Convert one-degree-of-freedom χ² to the supplied tail approximation, including its extreme fallback	source_normal_z and source_chi_p_value
DNA5/MathFuncsDll.cpp, MakeTWinP, around 18644–18665; GetACP; GrowMChiWin2P2 / GrowMChiWinP2	Start near one quarter of the original half-window, expand both sides, retain non-decreasing χ² maxima, stop at missing boundaries or after failures	grow_peak, preserving the MaxX == 0 → 1 entry quirk and missing-edge ordering
DNA5/MathFuncsDll.cpp, FastRecCheckMC2, around 19215–19510	Gate the triplet, scan/smooth peaks, grow one candidate, correct its probability, construct/destroy candidates, and retry	strongest admissible peak, growth, and probabilities only
DNA5/MathFuncsDll.cpp, AlistMC3, around 23138–23220	Schedule MaxChi over analysis lists and feed persistent event records	representative/final-list scheduling is mapped to bounded recheck evidence; ordinary exploratory discovery uses the separately traced MCXoverF lifecycle

The RDP5 manual's distinction between discovery and later corroboration remains intact. The kernel in this document cannot create or move an event. Separately, the active ordinary MCXoverF port can author a method-labelled discovery signal that enters the shared cyclic event workflow.

Variable-site and match profiles

For each triplet (s0, s1, s2), the kernel walks the compact working alignment once. A site enters the variable profile only when all three states are nonzero and the three states are not identical. It records the original 1-based alignment coordinate and three bytes:

Pair track	Stored value
0	s0 == s1
1	s0 == s2
2	s1 == s2

The profile also retains a prefix map from every alignment coordinate to the current variable-site index. That map is used for native-shaped MissingData bans without searching the coordinate list.

Missing data, prior erasures, and linear ends

The caller constructs one compact triplet mask from two sources already present in the RDP cycle:

1. immutable input MissingData produced by the supplied ten-consecutive-low-state rule; and
2. coordinates that were present in the original row but are now zero in a working representative, covering prior tract erasure and the unavailable side of a re-entered fragment.

MakeBanWinP treats boundary zero and boundary LenXoverSeq as distinct array cells even though they describe the same circular join. The port retains both cells. For each mapped missing site it bans the two native half-window ranges, records both adjacent MDMap edges when available, and applies the source trailing-window ban when either origin alias is marked. Linear sequences add the same first/last boundary markers and trailing ban used by FastRecCheckMC2.

The result reports input/erasure filtering and linear-edge filtering as separate booleans. It does not claim that a missing-data-filtered profile is an ordinary negative MaxChi result.

Rolling χ² scan

The initial half-window is int(70 / 2 + 0.51) = 35, then follows MakeWindowSizeP for short profiles. The GetCriticalDiff probability is max(0.0001, cutoff / 6). The kernel finds the smallest absolute extreme match difference whose χ² exceeds that threshold and subtracts one, because the scan itself uses a strict difference > CriticalDiff comparison.

For each pair track, the first left and right half-window totals are calculated once. Advancing one boundary removes one match byte and adds one match byte on each side. Consequently, scanning all three tracks is O(V) for V variable sites and uses constant rolling state; it does not perform O(V × H) repeated half-window summation and does not retain the complete χ² profile.

The statistic for left matches A, right matches C, and half-window H is:

[ \chi^2 = \frac{2(AD-BC)^2}{H(A+C)(B+D)},\qquad B=H-A,\ D=H-C. ]

Ties retain the earlier boundary/pair because the source maximum finder updates only on a strict increase.

Growth and probability scopes

If the initial within-triplet probability passes the project cutoff, growth starts from int(H / 4 + 0.51), clamped to [6, H] and half the variable profile. Both sides expand together; a non-decreasing χ² resets the failure count. A new window is evaluated before its MDMap edges stop further traversal, matching GrowMChiWin2P2's ordering.

Three values remain distinct in JSON, CSV, and the review UI:

1. localPValue = ChiPVal2P(maximumChiSquare);
2. withinTripletPValue = min(1, local × V / min(initial H, grown H) × 3); and
3. correctedPValue, which additionally applies the current valid distinct-origin triplet count when Bonferroni correction is selected.

sourceRecheckHit compares the final corrected value with the ordinary project cutoff. A profile with no peak above the critical-difference screen is still a completed profile and is represented separately from an unavailable profile.

Integration points

• UniqueEvent.maxchi_triplet_recheck stores a recheck of the current recombinant/major/minor representatives. It is recalculated when role hypotheses are refreshed.
• DistanceCorrelationEvidence.post_group_maxchi_recheck stores the same strongest-peak evidence for each finalized nonrepresentative distance-list row against its role's two representatives.
• The own-role representative follows the supplied final-list loop's representative skip; its event-level representative evidence remains available above the table.
• MaxChi evidence is serialized into ordinary results, reloadable projects, and event CSV. The UI shows profile availability, bans, pair, peak, χ², windows, and all three probability scopes.
• This strongest-peak recheck never writes the cyclic discovery catalogue; the separate exploratory MaxChi kernel can author ordinary method-labelled discovery signals.

Deliberate boundary and deviations of this recheck

The following behavior is not claimed by this bounded recheck. Several items are represented by the separate exploratory kernel, but they are deliberately not folded into corroboration results:

• SmoothChiValsP, ordered local-peak selection, DestroyPeakP, the up-to-100 retry loop, and its WasteOfTime behavior inside the recheck;
• MaxChi breakpoint-pair/event construction or writes back to the event catalogue from a recheck;
• full AlistMC3 late dispatcher/event reconstruction and every enabled-method combination;
• native lookup-table rounding (ChiTable2); the port evaluates the same formula in double;
• growth beyond half the variable profile. The browser stops before the two half-windows overlap, while the supplied loop relies on failure/boundary termination and has no explicit overlap guard;
• desktop-vs-browser golden numeric validation.

Those boundaries are machine-readable as eventDiscoveryApplied: false, maxChiKernelStatus: "source-shaped-strongest-peak-unvalidated", and nativeMaxChiFullRecheckComplete: false.

Next native-parity fixtures

Before calling either MaxChi path validated, compare the supplied desktop/DLL and this kernel at full precision for: MaxX = 0; exact critical-difference ties; zero/length BanWin aliases; missing sites before the first variable site; linear end windows; a grown maximum smaller than, equal to, and larger than the initial half-window; correction factors near the native cap; and multiple peaks requiring destruction/retry. The full corpus is listed in validation-plan.md.