native-maxchi-discovery-trace.md

Supplied-source trace: MaxChi exploratory discovery

This trace documents the Session 10 ordinary-triplet MaxChi discovery port. It uses only the supplied RDP5 manual, VB6 application source, DNA5 DLL source, and the companion dna.dll source. No alternate RDP implementation was consulted. The separate FastRecCheckMC2 confirmation path is covered in native-maxchi-recheck-trace.md.

Manual workflow represented

Section 8.4.1 of the supplied manual describes the exploratory triplet workflow:

1. discard sites that are not variable across the selected sequences;
2. slide a centrally partitioned window and calculate all three pairwise χ² plots;
3. select the highest peak, grow its window one site at each end until the score declines;
4. compare the far-side left and right χ² values to choose the recombinant tract side; and
5. repeat with the next peak until no significant peak remains.

The browser keeps this discovery stage separate from later event reconciliation. Method-labelled RDP and MaxChi signals enter the same strongest-first round, two-shared-identities/tract-overlap support grouping, three-role evidence, erasure, fragment re-entry, and complete rescan cycle.

Active supplied call path

Supplied location	Native responsibility	Port location
VB Source/Module5.bas, MCXoverF, around 49606–50910	Active ordinary triplet scan, probability gates, peak loop, event construction, destruction, and retry exits	wasm/src/maxchi.cpp::maxchi_discover; RdpScanner::scan_triplet
Module5.bas, around 49684–49822	Variable-site compression, window selection, missing/erased and linear-end bans, three raw χ² profiles	build_variable_profile, choose_discovery_window, make_banned_windows, calculate_chi_profiles
DNA5/MathFuncsDll.cpp, FindMChiP, around 4467–4492	Strict global raw maximum over boundary outer/pair inner loops	one-time HeapPeak priority ordering with lazy destroyed-cell rejection
DNA5/MathFuncsDll.cpp, SmoothChiValsP, around 5008–5047	Smoothed profiles used only to find destruction basins	source_smooth_chi
DNA5/MathFuncsDll.cpp, GrowMChiWinP2 and GrowMChiWin2P2, around 4359 and 6508	Symmetric window growth and missing-boundary stop	grow_discovery_peak
dna.dll/threshold.CPP, FindSide, around 12927	Far-side left/right χ² comparison	find_side_chi
threshold.CPP, OptLeftBPMC / OptRightBPMC, around 27085 / 27177	One-boundary optimization with scaled χ² comparison and failure limit	optimize_left_breakpoint, optimize_right_breakpoint
threshold.CPP, DestroyPeaks, around 12792	Remove the smoothed basin around a completed tract	destroy_completed_peak_region
DNA5/MathFuncsDll.cpp, DestroyPeakP, around 18662	Remove the smoothed basin around a rejected grown peak	destroy_rejected_peak
Module5.bas, around 50411–50433 and 50870–50910	Ensure the source peak lies in a completed destruction interval; clear the exact raw maximum; stop after three wasted cycles	include_source_peak_in_destroy_region, exact-cell clear, wasted_attempts

Module31.bas contains an older/sibling copy with disabled optimization branches. Session 10 follows the active Module5.bas path, where both breakpoint optimizers are called.

Index and ordering contract

The source mixes zero- and one-based positions. Session 10 retains these distinctions deliberately:

• raw ChiVals peaks are searched at indices 0 ... LenXoverSeq - 1;
• sequence match and missing-boundary positions are treated as 1 ... LenXoverSeq;
• MaxX = 0 becomes 1 for growth, while the original raw index is retained for reporting and exact peak destruction;
• FindMChiP updates only on strict >; equal scores therefore prefer the lowest raw boundary, then pair 0, pair 1, pair 2;
• the accepted-interval adjustment uses ordinary numerical absolute differences, not shortest circular distance, exactly as the VB branch does; and
• alignment coordinates are emitted one based from the compressed variable-site map.

The smoothing code also preserves a supplied implementation quirk. SmoothChiValsP sums positions -5 ... +6—twelve terms—but divides by eleven. Its terminal ChiVals[LenXoverSeq] access is a zero-filled padding cell, not an alias for raw boundary zero. The port allocates that padding cell and keeps the resulting asymmetry because it can change destruction basins.

Candidate probability lifecycle

For each raw maximum, the port keeps the three native scopes separate:

• raw one-degree-of-freedom ChiPVal2P-shaped tail;
• within-triplet position/profile correction, raw × LenXoverSeq / min(initial,grown window) × 3;
• optional current-round triplet correction.

The raw maximum first passes the source pre-growth within-triplet gate. The peak is then grown using the active DLL's highest-χ² rule. A corrected-significant result enters side selection and tract construction. A nonsignificant grown result has its smoothed peak basin destroyed and counts as a wasted attempt. A stored candidate resets the wasted counter. Three consecutive wasted attempts end the triplet scan; a separate 100-attempt cap bounds the native Redox loop.

Safe performance adaptation

The supplied VB loop calls FindMChiP after every destruction, rescanning all three raw profiles. The port constructs the same raw priority order once with linear-time heap construction. Each extraction is logarithmic; entries whose raw cells were cleared by a destruction basin are discarded lazily. Raw χ² construction itself uses rolling left/right totals on all three pair tracks, avoiding a half-window sum at every boundary. In the combined scan, RDP categories, MaxChi pair matches and variable prefixes, and the triplet missing/erasure map are populated during one alignment-byte pass. Capacity-retaining workspace vectors are reused across triplets.

The original destruction functions contain index expressions that can address outside their logical profile range under malformed/edge state. The browser port uses bounded circular traversal and a zero padding cell; it does not reproduce undefined memory reads. This is a safety adaptation and a golden-comparison item, not a claim of bit-for-bit identity.

Event-role boundary

The native scan initially stores a tract and later uses desktop event state, FindDaughter, existing event counts, and StoreLPV to arbitrate recombinant and parents. Those global desktop structures do not have a one-call equivalent at the raw discovery boundary. Session 10 derives a provisional role from inside-versus-outside changes in the three pair similarities, then immediately subjects all three roles to the shared supplied-source late evidence and weighted consensus. The raw peak pair, provisional role, final recommendation, and any user edit all remain auditable. Native golden role fixtures are required before this part is called parity validated.

Deliberately unresolved modes

• MaxChi permutation significance modes;
• the manual pair/doublet workflow whose variable positions are derived from a larger alignment;
• native lookup-table and floating-point rounding parity;
• full desktop method-battery interactions with MaxChi role voting; and
• native golden comparison of smoothing aliases, destruction basins, missing-boundary edges, and combined RDP/MaxChi event order.

These limitations are exposed in result metadata, documentation, and the validation plan. They do not prevent the ordinary enabled triplet workflow from reaching scan, reconciliation, ordered review, project checkpoint, CSV, and accepted-alignment export in source.