CHANGELOG.md

Changelog

0.26.0-session-26

• Reproduced cyclic RDP's post-threshold CheckSplit/SplitEvent path. A spanning candidate must first pass the ordinary unsplit cutoff, then the replay-only MissingData matrix is sampled at ten-base steps, the outer pieces are independently rescored, and only significant pieces are emitted. This includes the native inclusive missing-data bounds, left-piece informative-site exclusion, and omitted long-tract exponent; it removes complement-of-erasure calls while restoring native split events and improves exact event prefixes in Datasets2, 3, and 9.
• Preserved synthetic working fragments when only the public significant-signal catalogue is available. Native DropSeqs uses all internal XOverList hits, including sub-threshold ones; treating reportable signals as the complete retention ledger prematurely deleted Dataset8's B fragment, changed the following CheckMatrixP panel, and reversed the third event's role.
• Matched installed RDP 5.93's cyclic-erasure boundary behavior: both reported breakpoint sites remain in the background row and are omitted from the retained fragment. The supplied older C body is inclusive despite its two-site-penalty comment; captured Dataset7 tree inputs confirm the executable uses the strict tract interior and that the difference changes the second event's NJ topology and role choice.
• Restored the supplied primary-RDP candidate walk: DefineEventP2's decrement-before-check anchoring and two-site resume, FastRecCheckP's N-minus-one search/probability bound and shared cross-role OldX, and FindFirstCOP's circular-origin terminal-count behavior. Together these restore Dataset6's first nine supplied events by identity, roles, breakpoints, and order while preserving Dataset9's first five-event prefix.
• Added RDPbugsandquirks.md as a compatibility-phase ledger for supplied behavior that should be reconsidered only in a future intentionally non-faithful cleanup.
• Ported the supplied RDP CentreBP midpoint rule, including shifting a post-erasure midpoint off missing data toward the detected tract. Scoring continues to use the informative-site bounds, so this changes stored coordinates without recalculating the discovery p-value.
• Generalized the BestXOList-style action cache to all committed events using the same original triplet plus greater than 80% tract overlap; manual/restored event invalidation remains broader.
• Restored the supplied AddjustCXO/MakePairsP DoPairs gate. Signals touching the selected RList enable their three original-sequence pairs, permissions propagate across the selected group, and dirty triplets lacking all three pairs bypass every numerical kernel.
• Corrected cyclic erasure to use the selected role's final ConsensusOK/FinalTrim distance list (RList(WinPP)) instead of the diagnostic two-of-three evidence union.
• Matched the active MakeConsensusC subset's method-specific numeric quantization, post-rounding exact-±1 sentinel guards and tree-distance cap, and restored the TreePhPr second-place contribution of 14 points rather than a generic half of 18.
• Added a graceful 64-round ceiling. It preserves completed events, records cycle-limit-reached, and continues into Review/export with an explicit UI notice.
• Audited ten simulated RDP-only saved analyses without inspecting alignment content. The audit strongly validates breakpoint centering and final-RList membership. The DoPairs gate reduced 509 events to 453, improved matched roles from 82 to 95 and matched triplets from 152 to 171, reduced heavy triplet-kernel entries by 78%, reduced capped cases from five to one, and cut native audit time from 125.5 to 72.7 seconds. Full cyclic parity is not yet established; exact role selection and event consolidation remain.
• Kept the partial TestMoveInTreeAlt experiment out of active event selection after its ten-case A/B worsened early parity. Its mutable pair totals, quick distance, and final tree stages must be ported together before enabling that pruning heuristic.

0.25.0-session-25

• Restored the supplied internal daughter/minor-parent/major-parent representative order in BURT and late role/group analysis while retaining the public recombinant/major/minor contract.
• Added TraceSub-equivalent working-row aliases from exact cached signal triplets. Strongest support establishes aliases for all original identities and the anchor triplet takes precedence; older restored records retain their fragment-event fallback.
• Changed selected-role outlier ranks from original rows to the supplied fragment-expanded NextNo population. Symmetric distances are requested once per unordered pair and representative totals reuse the completed rows, roughly halving pair-cache calls without changing summation order.
• Kept fragment evidence off the primary hot path. When needed, its analytical alignment view copies state bytes only and avoids the alignment's O(N^2) pair-similarity table and string metadata.
• Generalized the Pages pthread artifact check to accept any emitted .worker.js/.worker.mjs basename that is actually referenced by the threaded Emscripten loader.
• Preserved the unrelated cyclic selected-result digest and all host-linked BootScan, SISCAN, event-tree, PHYLPRO, cyclic-pruning, and deterministic multicore regressions. Full native cyclic parity remains unclaimed.

0.24.0-session-24

• Added monotonic whole-run timing using performance.now(), with separate setup, primary-screen, cyclic-rescan, reconciliation, current-round, and completed-round measurements. Timing and CPU execution provenance are included in live progress, final results, and project checkpoints.
• Restored responsive progress rendering at a maximum of once per 100 ms. A Win95 segmented marquee remains visible while a new cyclic round is at zero processed triplets, so rapid round transitions no longer appear to lose the progress meter.
• Added deterministic method-level multicore execution. A persistent pool can run GENECONV, BootScan, MaxChi, CHIMAERA, SISCAN, and 3SEQ kernels independently while triplet scheduling, cyclic shortlist/cache decisions, and method-major signal merging remain serial and stable.
• Added hardware-aware CPU selection (navigator.hardwareConcurrency), a headroom-preserving default capped at six heavy-method lanes, and a Settings control that can reduce the analysis to one. The linked regression requires byte-identical progress and result JSON at one and four CPUs.
• Production builds now emit both pthread and single-worker modules. GitHub Pages uses a small same-origin service-worker response shim to apply COOP/COEP after one automatic first-use reload; browsers without isolation transparently use the compatibility module.
• Added pthread/static-isolation artifact checks and the multicore host gate to GitHub Actions. The project schema remains org.rdp-web.project/v1alpha19: timing/execution metadata are additive and the analytical state format is unchanged.

0.23.0-session-23

• Implemented Darren's permanent clean-triplet rule: when an exact working triplet produces no signal on its first complete screen, later cyclic passes skip every enabled discovery kernel. The one-time post-erasure 3SEQ split refresh is restricted to previously signal-bearing unchanged triplets.
• Ported the supplied DoRDP/DropSeqs fragment lifecycle. A fragment created by one event gets one complete follow-up screen; if it participates in no detected signal, it is deleted before another event adds rows. Swap-with-last compaction remaps signal and shortlist working indices.
• Excluded invalid same-origin fragment combinations from the worker batch budget while retaining deterministic schedule traversal, progress accounting, and cancellation boundaries.
• Made erasure maintain valid-site summaries and state fingerprints incrementally, changed active row refresh from O(NL) to O(N), delayed dense fragment allocation until it passes retention gates, and added collision-safe fingerprint prefiltering to exact duplicate checks.
• Replaced repeated prior-event tract marking in breakpoint polishing and breakpoint-context work with inclusive difference-array unions, reducing those stages from O(EL) to O(E + L) without changing coordinates or missing-data masks.
• Throttled routine progress JSON/WASM/worker/React traffic to at most once per 500 ms. Added adaptive batches targeting 40 ms worker slices plus scheduler.yield() with a timer fallback.
• Enabled safe release -O3, LTO, and WebAssembly SIMD; deliberately did not enable fast-math.
• Added a linked cyclic-pruning regression to local checks and GitHub Pages Actions. Its fixture exercises clean-negative reuse, same-origin bypass, and fragment drop/remap while retaining the exact Session 22 selected-result digest.
• Advanced engine/package version to 0.23.0-session-23. Project schema remains org.rdp-web.project/v1alpha19 because these changes add transient telemetry and scheduling shortcuts, not new saved analytical state.

0.22.0-session-22

• Added the supplied PHYLPRO selected-event review path from FindSubSeqPP, PXoverD, MakePDstMat, UpdatePDstMat, and PPRegression: polymorphic-column mapping after the chosen gap policy, paired rolling half-windows, Hamming-distance vectors, optional self observation, and source Pearson/zero-variance behavior.
• Kept PHYLPRO lazy and diagnostic-only. The supplied RDP5 workflow has no implemented PHYLPRO significance calculation, so this route emits no p-value, discovery signal, breakpoint change, or cyclic event-order effect.
• Replaced the supplied all-pairs working matrix with only the recombinant, major-parent, and minor-parent target rows consumed by the ordinary plot. Coefficients are unchanged while rolling work falls from O(LN²) to O(LN) and scratch distance memory from O(N²) to O(N).
• Added complete-window handling for linear analyses and response thinning after full evaluation, always retaining profile ends, all three minima, and breakpoint-nearest points.
• Confirmed the supplied Match(66/68/72/85) assignments are the intended per-column resets for RDP's encoded A/C/G/T values; the browser's fresh state set is equivalent. Explicitly repaired the separate reversed RevSeq compact/original lookup when disabled rows are not a trailing block. That repair and the manual/source column-map divergence are visible in docs/UI.
• Added a typed C/WASM/worker/client endpoint and Windows 95 PHYLPRO inspector with adjustable window, gap policy, and self inclusion, plus role curves, event highlighting, minima, context, and kernel telemetry.
• Added a deterministic brute-force PHYLPRO regression across circular/linear topology, both gap modes, self policy, disabled context, planted mosaic boundaries, and oversized windows. The linked public API and production Pages WASM smoke tests now exercise the lazy endpoint too.
• Added PHYLPRO to source/ABI/Actions gates and advanced engine/package version to 0.22.0-session-22. Project schema remains org.rdp-web.project/v1alpha19 because the transient profile is reproducible from already-saved alignment/event state and request options.

0.21.0-session-21

• Added the supplied ordinary SISCAN path: SSXoverC, nearest-outlier GetSSOL, source-shaped WPGMA/cophenetic context, Get3Score/GetPScores2 categories, Microsoft-CRT MakeVRand flat prefix, DoPerms3, population-variance MakeZValue2, DoSums, FindMaxZ, and ShrinkRegionC boundary construction.
• Preserved the default gap-stripped one/two/three-variable mode and the supplied QuickCheckB missing-braces fast-screen quirk. Primary discovery defaults off; fixed-region representative and finalized-list confirmation defaults on, matching the ordinary RDP5 workflow.
• Added a round-wide cached direct/WPGMA context and reusable seeded random prefix. Event erasure invalidates only state-dependent context, while unchanged exact triplets continue to use the XOverList/BestXOList-style shortlist.
• Pruned disabled rows before direct-pair scoring and WPGMA merging, and regression-tested both disabled and same-origin-fragment fourth-sequence rejection.
• Added SISCAN to cyclic priority after CHIMAERA and before 3SEQ, plus source roles, review plots, progress/work counters, anchor/representative/final-list review evidence, project restore, and expanded CSV diagnostics. Corrected CSV method classification so SISCAN support cannot fall through as RDP support.
• Made the compact three-pair SISCAN review plot preserve signed extrema by selecting the greatest-absolute eligible partition/summed Z at each pair/window; full 15-plus-9 manual plotting remains an explicit boundary.
• Added a focused deterministic SISCAN regression for tract, outlier, roles, probability chain, exact seed-3 Microsoft-CRT bytes, cached replay, fixed-region confirmation, plotting, cyclic invalidation, and same-context restart. New analyses reset BootScan/SISCAN cache telemetry while retaining the reusable SISCAN random prefix. The gate runs in GitHub Pages Actions before the production Emscripten smoke tests.
• Completed v1alpha19 round-tripping for the SISCAN WPGMA build/comparison/merge and random-value counters instead of resetting those four audit metrics on project import.
• Advanced the engine/package version to 0.21.0-session-21 and project schema to org.rdp-web.project/v1alpha19; imports remain compatible through v1alpha1 and restore SISCAN disabled for older schemas.

0.20.0-session-20

• Traced the automatic six-region event-tree workflow through the supplied TestMoveInTree, MakeNJTreesP, bundled Clearcut, SEQBOOT2, FastBootDistIP6, TreeRepsP, CollapseNodes, Tree2ArrayP, and MakeTreeArrayXP2 paths.
• Replaced the generic double-precision NJ/bootstrap helper with source-shaped single-precision Clearcut NJ, including its ordered strict ties, first-round terminal-r2 behavior, packed collapse order, final-clade branches, negative-sign suppression, and five-decimal serialization.
• Reproduced the Microsoft CRT bootstrap stream, both discarded draws, retained unresampled base column, configured shared seed, base-tree support pseudocount, VB6 integer rounding, and strict below-50% collapse rule.
• Replaced raw branch-length analytical matrices with the supplied TreeMidP/UltraTreeDistP midpoint-ultrametric, four-decimal ascending rank / 1000 topology distances. Weak-node collapse promotes to existing parent ranks without recompressing the scale before TreePhPr, TreeSubPhPr, TreeSubDist, TrpScore, phylogenetic grouping, and later role consensus.
• Deliberately repaired the supplied Newick writer/parser root-child defect: every actual edge is handed to the four-digit clamp/midpoint pipeline instead of scanning into an unrelated token in the opposite subtree. This numerical fix is explicit in provenance and remains golden-testable.
• Kept the manual's advertised six-tree workflow with four independently bounded 20-informative-variable-site flanks. The late VB source's two bootstrapped whole-region trees plus 60-site direct-DMatS flank divergence is documented and remains an explicit golden boundary.
• Exposed event-tree kernel, seed, rank-level, normalized-negative-branch, and bootstrap provenance through typed results, project snapshots, the Windows 95 review UI, and lazy graphical endpoint.
• Added a deterministic linked event-tree regression and required it in GitHub Pages Actions beside the existing full-core BootScan regression.
• Advanced the engine/package version to 0.20.0-session-20 and project schema to org.rdp-web.project/v1alpha18; imports remain compatible through v1alpha1.

0.19.0-session-19

• Added source-shaped automated distance-mode BootScan as a primary discovery method in the supplied RDP → GENECONV → BootScan → MaxChi → CHIMAERA → 3SEQ cyclic order.
• Ported the supplied BSXoverR, seeded SEQBOOT2, FastBootDist, strict GetPltVal, ScanBSPlots/FindBeginBS/FindEndBS, MakeBSEvent, BSSubSeq, MakeScoresBS, and ProbCalc behavior used by that route, including SEQBOOT2's generated-but-discarded tail replicate draw that advances the retained Microsoft-CRT random stream.
• Added a bounded 64 MiB shared pair/window/bootstrap profile cache. It reuses pair work across triplets within one mutable-alignment round, reports hit/miss/eviction/byte telemetry, uses shared ownership for safe FIFO eviction, and invalidates at every erasure/re-entry boundary.
• Kept that pair cache distinct from the existing XOverList/BestXOList-style whole-triplet cyclic shortlist, which continues to replay unchanged signal summaries and skip stable method kernels.
• Corrected BootScan scoring to preserve the supplied XPosDiff invariant-boundary indexing, 169-site scale/exponent route, and raw/project-corrected p-value separation. BURT/BenHMM still polishes event coordinates later and does not recalculate the detection p-value.
• Added BootScan discovery evidence, plots, progress/cache diagnostics, settings, review cards, JSON/CSV export, hardened restore, source contracts, and production-WASM smoke-test coverage.
• Advanced the engine/package version to 0.19.0-session-19 and project schema to org.rdp-web.project/v1alpha17; imports remain compatible through v1alpha1.

0.18.0-session-18

• Re-skinned the complete application as a Windows 95 workstation without changing analytical behavior: teal desktop, navy active title bar, classic menu and status strips, gray system face, square raised/recessed bevels, Tahoma/MS Sans Serif fallbacks, inset tables and plots, and segmented navy progress blocks.
• Applied the same visual system to workflow navigation, upload/curation controls, settings, notices, scan phases, review lists, breakpoint/tree/alignment inspectors, evidence cards, and export panels, including a compact raised drawer on narrow screens. No external visual asset or network dependency was added.
• Fixed stop handling during cyclic discovery. At the next 512-triplet boundary, the engine now discards only transient signals from the unfinished round, retains every event committed by prior complete rounds, clears incomplete shortlist summaries, records user-stopped, and runs normal final reconciliation so the UI moves to Review with usable results.
• Added an explicit stopped-cleanly notice and changed the control label to “Stop and review completed events.” Stopping the opening round safely returns a zero-event result when no event has yet reached a complete round boundary.
• Added a production-WASM graceful-stop regression that enters a second cyclic round, evaluates a partial batch, stops, reconciles, and verifies the exact completed-event prefix with no orphaned signals. The existing FASTA-upload and exact cyclic-shortlist digest regressions remain active.
• Advanced the engine/package version to 0.18.0-session-18; project schema remains org.rdp-web.project/v1alpha16 because the saved analysis shape did not change.

0.17.0-session-17

• Traced the supplied GENECONV probability lifecycle through GCCalcPValP2, GCXoverD, MakeMCCorrection, and later BURT polishing. WebRDP's raw KA probability is calculated from the detected fragment before BURT; RDP5's stored XOverList.Probability corresponds to WebRDP's project-corrected value. Review UI now states both scopes and the exact initial opportunity multiplier explicitly.
• Fixed cyclic multiple-comparison fidelity by freezing the source MakeMCCorrection factor from the initial exploratory or query/reference scan plan. Later fragment-expanded schedules still report their real triplet workload, but no longer silently change stored probability scope.
• Added source-shaped cyclic shortlist reuse based on XOverList/XOverDefine, BestXOList, Worthwhilescan, and StoreLPV: significant summaries for unchanged exact working triplets are replayed, empty unchanged summaries skip their method kernels, and only triplets touching erased rows or new fragments run fresh full discovery.
• Preserved 3SEQ's semantic boundary by refreshing it once on every unchanged triplet when the scan first enters the supplied post-erasure CheckSplit3Seq mode; later rounds can safely reuse it. Manual event repair/rejection rebuilds clear the shortlist and start with a full pass.
• Added live counters for triplet kernel evaluations, summaries/signals reused, and method scans skipped. A deterministic two-event production-WASM Pages regression asserts fixed correction, cached signal replay, skipped work, and fewer kernel evaluations than scheduled triplets.
• Compared the optimized and temporary full-rescan builds on an all-five-family cyclic fixture; event roles, breakpoints, rounds, support order, signal triplets, and p-values produced the same digest across 22 events/422 signals while kernel evaluations fell from 131,488 to 32,911.
• Recorded two deliberate GENECONV safety fixes: a bounded root-bracketing fallback is used only when the supplied Newton iteration becomes unstable (and is counted), and the apparent NDiff[3] == 1 typo is treated as the intended assignment to avoid degenerate division by zero.
• Completed WASM and web production builds plus FASTA-upload, cyclic-shortlist, source-contract, TypeScript, and GitHub Pages artifact verification.

0.16.1-session-16

• Fixed FASTA and project uploads under Emscripten 5.0.1 by exporting the HEAPU8 view used by the worker to copy file bytes into WASM memory.
• Added a GitHub Pages artifact smoke test that instantiates the production WASM module, copies a three-sequence FASTA through that heap view, loads it, and validates the alignment summary.
• Completed a clean Node 20 / Emscripten 5.0.1 production build and Pages artifact verification.

0.15.0-session-15

• Ported the supplied late TSXOver(1) 3SEQ Findall shape for the representative event triplet and every finalized nonrepresentative distance-list row. It scans all three source target rotations, preserves the initial lower-P/low-information/corrected gate, and then evaluates both walk orientations rather than allowing the ordinary strictly-better-orientation shortcut.
• Retained forced post-erasure CheckSplit3Seq/SubPVal evaluation for both Findall orientations, the distinct post-SwapRound correction branch, and the extra inverse-parent/inverse-interval XOverList copy for each accepted orientation. Recheck evidence records profile/evaluation counts, qualifying orientations/list entries, best target/direction, bounds, excursions, probability route, split state, raw/corrected P, and cutoff result without moving event coordinates.
• Added event and post-group 3SEQ recheck JSON, review cards/badges/table cells, detailed CSV columns, and late-consensus diagnostics. The late evidence status now names five active families while retaining nativeThreeSeqFullRecheckComplete=false until golden and full catalogue parity exist.
• Ported the supplied later-round 3SEQ FindSubSeqTS2 → CheckSplit3Seq/SubPVal path. The compact upper_bound mapping preserves XPosDiff[x] as the retained-site count at or before an alignment coordinate without allocating a length-sized integer map per target.
• Preserved the two-stage TSXOver gate: only an initially significant call enters split handling; beginning- and ending-side missing runs are tested, the reverse orientation is retried only when its unsplit probability beats the trimmed selection, strict lower-P swapping is retained, and project correction is recalculated before emission.
• Reused the fused triplet missing/erasure byte map and compact walk, so later-round split handling adds bounded coordinate scans and probability evaluations but no alignment-byte pass. Split state, revised bounds, raw/corrected probability, route, and excursion survive JSON, CSV, review, and hardened restore.
• Changed the compact exact DP state from extended precision to supplied Single precision. This narrows every accumulation like the desktop lookup and materially lowers bounded WASM memory and arithmetic cost.
• Corrected the scaled GetTSPVal fallback to use the supplied post-truncation onM/nM or onN/nN exponent and the 1e-300 positive-underflow floor. Also separated the source's pre-wrap probability excursion from the post-CheckwrapC boundary excursion throughout the contract.
• Aligned the visible and serialized kernel order with the actual source call sequence: FindSubSeqTS → Seq3PVals/GetTSPVal → CheckwrapC → TSXOver.
• Advanced the engine to 0.15.0-session-15 and project schema to v1alpha15. Imports accept v1alpha1–v1alpha15; v14 retains ordinary 3SEQ support but cannot claim v15 split evidence.
• Kept manual 3SEQ permutation envelopes and full late event-catalogue reconstruction explicit rather than claiming the new non-coordinate-changing corroboration is complete dispatcher parity.

0.14.0-session-14

• Ported ordinary automated 3SEQ directly from the supplied manual and active FindSubSeqTS / FindSubSeqTS2, Seq3PVals/Get3SeqPvalC, GetTSPVal, CheckwrapC, SiegmundDiscrete, SwapRound, and TSXOver routes, without consulting alternate implementations.
• Added the three source-order candidate-recombinant rotations and corrected target-to-parent equality-slot mappings {0,2,1} / {1,0,2}. Parent-equal, all-different, and missing sites are excluded, while the DLL's effective four-information-rich-site minimum is retained.
• Preserved maximum descent/ascent discovery, the source's probability-before-CheckwrapC call order, origin extension and beginning advance, linear interval conversion, strict lower-P orientation choice, provisional parent swap, and both supplied low-information exits. JSON, restore, CSV, and review retain pre-wrap probability and post-wrap boundary excursions separately.
• Replaced the desktop four-dimensional exact lookup with a compact finite hypergeometric maximum-drawdown dynamic program. Results narrow to source Single, cache by (m,n,k), and obey an eight-million-transition / 8192-entry browser bound.
• Added the supplied large-profile SiegmundDiscrete/ApproxNu route and bounded scaled-exact fallback, including its post-truncation exponent ratio and positive-underflow floor, with exact and approximate evaluations retained separately. Literal zero-P emission, the p > 10^-15 Dunn–Šidák branch, and the smaller-tail product route are preserved.
• Reused the existing fused RDP/MaxChi/CHIMAERA/GENECONV equality preparation so 3SEQ adds no full alignment-byte pass. Method code 4 joins the source-major RDP → GENECONV → MaxChi → CHIMAERA → 3SEQ strongest-first cyclic scheduler and the existing grouping, erasure, fragment, consensus, BURT, ordered review/repair, and export workflow.
• Added enable/default state, correction explanation, target/exact/approximate/candidate progress, complete JSON/restore/CSV anchor evidence, a signed three-target random-walk plot, an anchor evidence card, and visible approximation disclosure.
• Added rdp_restore_threeseq_discovery, four authoritative restore counters, CMake source/export coverage, and hardened import checks for counts, excursion bounds, role rotations, pair IDs, low-information calls, post-erasure split claims, and unknown method labels.
• Advanced the engine to 0.14.0-session-14 and project schema to v1alpha14. Imports accept v1alpha1–v1alpha14; all pre-v14 projects restore 3SEQ disabled so saved discovery semantics cannot silently change.
• Added the supplied-source 3SEQ trace, Session 14 handoff, status/workflow/architecture/validation updates, and static contracts for pair mappings, source gates, bounded exact/fallback work, ABI, restore, UI, plots, exports, version, schema, and Pages deployment.
• Kept the ambiguous FindSubSeqTS2 alias, CheckSplit3Seq/SubPVal post-erasure re-probability, manual permutation envelopes, late-list corroboration, authorized native golden comparison, and runtime validation explicit as unresolved boundaries. No compiler, bundler, preview server, dependency install, or project/runtime test was invoked.

0.13.0-session-13

• Ported ordinary automated triplet GENECONV directly from the supplied manual and active FindSubSeqGCAP6, GetFragsP, GetMaxFragScoreP, CalcKMaxP, GCCalcPValP2, lowest-P/overlap helpers, and Module31.bas::GCXoverD, without consulting alternate implementations.
• Added the three inner pair-match and three outer discordant-sequence signed tracks with the supplied six-way provisional recombinant/minor/major role mapping, ignored-indel default, skew rejection, inactive automated minimum-fragment predicates, and configurable one-overlap default.
• Reused the existing fused RDP/MaxChi/CHIMAERA variable-site/equality preparation, so enabling GENECONV does not add another complete triplet alignment-byte scan.
• Preserved source mismatch penalties, Newton initialization, lambda/K calculation, strict critical-score gate, single-precision narrowing points, and Karlin–Altschul low/high-score probability branches. Added a bounded bisection fallback only when the native unbounded Newton iteration becomes numerically unsafe, with an exported diagnostic counter.
• Restored literal critical/tail rounding and native zero-underflow handling, preventing an exact zero KA probability from being promoted to an artificial positive display-floor value.
• Replaced the direct quadratic positive-fragment extension with signed prefix scores, next-strictly-lower bounds, and rightmost range maxima. This preserves the native first-negative stop and latest-ending tie while reducing each track to O(R log R).
• Corrected circular construction to preserve GetFragsP's initial category run plus terminal origin-wrapped copy without repeating the complete signed list. One category pass now builds all six tracks and coalesces only the positive outer fragments coalesced by the supplied helper.
• Preserved stable track/fragment lowest-raw-P ordering and implemented the overlap allowance with a lazy range-add/range-maximum tree in O(F log V).
• Joined independently labelled GENECONV signals to the strongest-first cyclic scheduler, shared cross-method grouping, three-set reconciliation, BURT polishing, event correction/rebuild, fragment re-entry, and final alignment exports. Exact corrected/local-P ties use the supplied RDP → GENECONV → MaxChi → CHIMAERA method-major order.
• Added non-coordinate-changing ordinary-kernel GENECONV corroboration for event representatives and every finalized distance-list row. It reuses MaxChi's prepared triplet workspace and retains best-track/tract/role, raw/corrected probability, skew, overlap, workload, and fallback evidence in project JSON, the review matrix, and CSV.
• Added GENECONV settings, validation, live counters, method-aware scan state, complete per-signal score/KA evidence, three-colour inner/outer -log10(raw P) review plot, anchor card, event method badges, CSV diagnostics, and interpretation boundaries.
• Added ABI/worker restore for method code 3, rdp_restore_geneconv_discovery, three scan settings, five authoritative progress counters, CMake source/export coverage, and cross-layer parameter-arity checks.
• Advanced the engine to 0.13.0-session-13 and project schema to v1alpha13. Imports accept v1alpha1–v1alpha13; pre-v13 projects restore GENECONV disabled so saved discovery semantics cannot silently change.
• Added the supplied-source GENECONV trace, Session 13 handoff, architecture/workflow/fidelity/ validation documentation, and static contracts across kernel, fused preparation, ABI, restore, types, settings, progress, review, plots, exports, version, schema, and Pages deployment.
• Kept permutation probabilities, manual pair scans, alternative indel modes, full late native event reconstruction, and all native-versus-WASM golden validation explicit as unresolved boundaries. No compiler, bundler, preview server, dependency install, or project/runtime test was invoked.

0.12.0-session-12

• Ported ordinary automated CHIMAERA discovery directly from the supplied manual §8.5.1, FindSubSeqDP3/6, FastRecCheckChim, AlistChi, and active VB CXoverA dispatch without consulting alternate implementations.
• Added all three candidate-recombinant target rotations and their target-specific information-rich binary strings: monomorphic and all-different sites are excluded, target/parent-one matches encode 1, and target/parent-two matches encode 0.
• Reused the combined RDP/MaxChi alignment-byte pass for CHIMAERA input. Target profiles are filtered from cached equality tracks, avoiding three additional complete triplet-row reads while preserving independently capacity-retained workspaces.
• Added fixed/fallback CHIMAERA window selection with supplied default 60, missing/earlier-erasure and linear-end bans, rolling raw χ² construction, source-shaped critical screen, and a per-target lazy peak heap with strict raw-score/coordinate ordering.
• Added source-shaped window growth, far-side tract choice, optimized second breakpoint, alignment- coordinate mapping, completed/rejected smoothed-basin destruction, three-consecutive-wasted exit, and the supplied 100-attempt bound independently for every target rotation.
• Preserved raw, within-triplet position/rotation-corrected, and optional project-corrected probabilities separately, plus initial/grown windows, critical difference, flank χ², filter flags, and inside/outside parent-one match rates.
• Joined method-labelled CHIMAERA signals to the strongest-first RDP/MaxChi cyclic scheduler. The selected target and provisional parent order remain auditable before shared late role consensus, BURT confidence, tract erasure, fragment re-entry, and fresh complete passes.
• Added CHIMAERA enable/window settings, target-profile/peak/candidate/limit progress counters, method-aware null and retry-limit notices, a dedicated evidence card, and a single-trace target/ parent-one χ² plot rather than misrepresenting the profile as MaxChi's three pair traces.
• Expanded result/project JSON and event CSV with CHIMAERA methods and complete anchor diagnostics. Added ABI/worker restore for method code 2, all discovery evidence, four authoritative workload counters, and the exported rdp_restore_chimaera_discovery symbol.
• Added secondary FastRecCheckChim representative/finalized-list corroboration across all three target rotations. It reuses MaxChi's prepared triplet bytes, retains raw/within/project-corrected strongest-target evidence in project JSON/review/CSV, and never moves an event's coordinates.
• Persisted a review and CSV warning when MaxChi and CHIMAERA are the only supporting methods, since the manual treats those methods as closely related rather than independent confirmation.
• Advanced the engine to 0.12.0-session-12 and project schema to v1alpha12. Imports accept v1alpha1–v1alpha12; v10/v11 projects retain MaxChi but restore CHIMAERA disabled, while projects through v9 retain their RDP-only discovery semantics.
• Added filter-wide reference assignment, first-appearance group compaction, exact dataset eligibility feedback, per-signal query/reference role metadata, live cyclic working-role counts, and honest zero-work terminal progress retained from the opening audit of this session.
• Added a supplied-source CHIMAERA trace, Session 12 handoff, validation cases, and static contracts across the kernel, fused preparation, C ABI/CMake/worker bridge, restore gates, types, settings, progress, review, plots, exports, version, and schema.
• Kept CHIMAERA permutations, full late-list event reconstruction, native lookup rounding, and all native-versus-WASM golden comparison explicit as unresolved boundaries. No compiler, bundler, preview server, dependency install, or project/runtime test was invoked.

0.11.0-session-11

• Ported the supplied automated query-vs-reference analysis-list workflow from manual §4.2 and Module3.bas::MakeAnalysisListQvR, without consulting alternate implementations. Each scheduled primary triplet contains one query and two references from different reference groups; later role inference remains unconstrained so a reference can still be identified as recombinant.
• Added dataset-level query/reference assignment with positive numeric group IDs, one-click parsing of documented REF-A<name>-style names, an all-queries reset, per-row status, filter-aware multi-row assignment beyond the 500-row rendering cap, first-appearance group compaction, and validation that at least one eligible query plus references from two groups exist before the scan plan proceeds.
• Added fully exploratory/query-vs-reference scheme selection, exact initial workload counts, scheme-aware scan/review/export language, live round-specific correction counts, result summaries, CSV role/assignment columns, and full reference-group persistence in project checkpoints.
• Made cyclic progress report the current active working-row, query-record, reference-record, and represented-group counts after every erasure/re-entry round; legitimate terminal zero-work rounds are no longer visually replaced by the initial workload.
• Added the manual's visible reference-as-recombinant distinction: event navigation and detail use an amber flag, all three current role cards/edit choices expose their query/reference input role, event JSON stores the selected recombinant's input role/group, and export summaries count these calls without forcing a query-based interpretation.
• Replaced the desktop routine's materialized triplet array with an O(N) active query/reference catalogue and a lazy cross-group reference-pair/query cursor. Source pair/query order is retained; bounded worker batches, cancellation, cyclic tract erasure, and fragment re-entry use the same scheduler without allocating O(T) triplet records.
• Kept the exact scheduled reference-record workload distinct from the supplied choose(reference groups, 2) × query origins multiple-testing rule, retained the native correction cap, added saturating 64-bit arithmetic, and exposed the distinction in settings and saved results.
• Made query/reference identity follow original-sequence provenance through working fragments, excluded same-origin working copies, recomputed groups/correction/workload after every erasure or repair, and added a dedicated no-eligible-constrained-triplets terminal state.
• Fixed schedule rebuilding for corrections at event zero and all-rejected saved prefixes by unconditionally refreshing active sequence, role, cursor, total, and correction state after replay.
• Hardened v11 restore by rejecting repeated or constraint-breaking saved triplets, rebuilding a missing grouped correction from the scheme rather than raw workload, filling legacy per-signal factors consistently, and validating every C ABI role-buffer length instead of silently defaulting.
• Completed method-correct MaxChi plot reconstruction by feeding the selected triplet's supplied ten-character input-missing map into banned-window generation, while retaining forced breakpoint, peak, and pair-maximum samples and the later-round reconstruction disclosure.
• Advanced the engine to 0.11.0-session-11 and projects to v1alpha11. Imports accept v1alpha1–v1alpha11; older projects remain exploratory, and projects through v1alpha9 retain MaxChi discovery disabled so saved detection semantics do not change.
• Extended the source-only contract checker across the constrained scheduler, C ABI/worker pointer bridge, restore compatibility boundary, options/results/types, dataset/settings/scan/review/export UI, version/schema markers, and delimiter balance. No compiler, bundler, preview, or project runtime was invoked.

0.10.0-session-10

• Ported the supplied ordinary-triplet MaxChi MCXoverF discovery path as an independently labelled method stream beside RDP, without consulting alternate implementations.
• Added all three variable-site match profiles, fixed/fallback window selection, native critical-difference screening, MissingData/prior-erasure and linear-edge exclusions, and the source's strict raw χ² peak ordering across boundary first and pair second.
• Added source-shaped peak growth, FindSide left/right tract selection, OptLeftBPMC and OptRightBPMC breakpoint refinement, missing-boundary adjustment, alignment-coordinate mapping, and preliminary roles that remain subject to the shared late role consensus.
• Reproduced the literal SmoothChiValsP eleven-divisor/twelve-term off-by-one behavior and padded terminal cell used by the supplied destruction routines.
• Ported accepted-region DestroyPeaks and rejected-peak DestroyPeakP behavior with bounded circular traversal, the source's numerical peak-inclusion adjustment, accepted-hit WasteOfTime reset, three-consecutive-failure exit, and 100-attempt cap.
• Replaced repeated full-profile maximum scans with a linear-time heap build and lazy invalidation of destroyed raw peaks. Rolling left/right match totals still make all three initial χ² profiles linear in variable sites rather than window width. The combined path prepares RDP categories, MaxChi matches/variable prefixes, and the triplet missing/erasure map in one alignment-byte pass.
• Joined RDP and enabled MaxChi signals in the strongest-first cyclic scheduler. Either method can anchor an event; cross-method supporting signals share the existing two-identities/strict-30%- overlap grouping, late role/group evidence, tract erasure, bounded fragment re-entry, and fresh complete-pass lifecycle.
• Kept FastRecCheckMC2 representative/finalized-list MaxChi confirmation separate from exploratory discovery, so its evidence cannot silently replace detected or manually edited event coordinates.
• Added method-aware progress counters, signal/event badges, RDP pair-identity versus MaxChi χ² plots, and a responsive MaxChi anchor card exposing attempt, side, pair, peak, windows, flank χ², raw/within/project probabilities, and missing/linear filters. Scan and review now warn when any triplet exhausts the supplied 100-peak retry bound with positive raw peaks still present.
• Moved the native global-maximum significance gate ahead of smoothing and retry accounting, so rejected triplets avoid unnecessary SmoothChiValsP work and report zero attempted peaks just as the supplied MCXoverF path does.
• Made plot thinning retain both inferred breakpoints, the selected MaxChi peak, and each pair profile's maximum. Later-round/fragment-assisted plots are now visibly labelled as original- alignment reconstructions instead of being mistaken for the historical erased working profile.
• Expanded CSV with anchor/detection methods and full MaxChi discovery diagnostics; expanded result and project JSON with discovery method, method set, settings, counters, and the per-signal trace.
• Added C ABI and worker restore support for method-labelled signals and MaxChi discovery state. Completed-project reload also retains cumulative rounds and all four MaxChi workload/limit counters, the exact terminal reason, and final-round processed/total triplets without inflating cumulative work while saved events are replayed. Advanced projects to v1alpha10; imports accept v1alpha1–v1alpha10, and older projects restore MaxChi discovery disabled so their saved RDP-only detection semantics do not change.
• Extended the source-only contract checker across the MaxChi kernel, C ABI/CMake/worker bridge, schema/version, settings, result types, review UI, and legacy-project compatibility boundary.
• Added a supplied-source MaxChi discovery trace, a Session 10 handoff, and a golden corpus for smoothing/index aliases, peak ties, growth/side optimization, destruction basins, retry bounds, method-aware plots/exports, and combined cyclic ordering.
• Kept native lookup-table rounding, preliminary role selection, permutation/manual-doublet modes, and all native-versus-WASM golden comparison explicit as unresolved parity boundaries.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.9.0-session-9

• Ported a bounded MaxChi confirmation kernel directly from the supplied DNA5/VB paths, without consulting alternate implementations: three variable-site match profiles, the source NormalZ/ChiPVal2P tail, GetCriticalDiff-shaped match screen, and fixed 70-site default.
• Reproduced MakeBanWinP's distinct zero/length boundary map for native input-missing runs, accumulated prior-event erasures, and linear sequence ends. Missing windows are excluded from the first peak pass and stop the source-shaped growth path at the corresponding outer edge.
• Replaced repeated per-boundary window summation with capacity-light rolling left/right match totals for all three pair tracks. The strongest-peak scan is now O(variable sites) with a small constant; only the selected peak enters the bounded growth pass.
• Preserved MaxChi's three probability scopes in the result contract: raw one-degree-of-freedom tail, within-triplet length / min(initial,grown window) * 3, and the current project-wide correction. A source recheck hit is reported separately from profile availability.
• Added representative-triplet MaxChi evidence to each reconciled event and reran the same kernel for every finalized nonrepresentative distance-list candidate beside the primary-RDP recheck. MaxChi remains corroboration-only and does not change RDP-authored event coordinates or groups.
• Added a responsive MaxChi review card and finalized-list column, including pair, peak position, χ², window sizes, all probability levels, missing/edge filters, and explicit event-discovery boundary. Settings now distinguish full RDP discovery from available MaxChi confirmation.
• Extended JSON, project JSON, CSV, late-consensus capability metadata, TypeScript contracts, and the static source-contract checker for the MaxChi evidence path.
• Advanced engine/source labelling to 0.9.0-session-9 and reloadable projects to v1alpha9, while accepting v1alpha1–v1alpha9.
• Kept MaxChi smoothing, ordered multi-peak destruction/retry, exploratory event construction, and native golden comparison explicit as the next fidelity boundary.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.8.0-session-8

• Revalidated the reattached RDP4/RDP5 DLL and VB source archives for safe paths and traced the manual's dataset → preliminary scan → hypothesis refinement → ordered review/export workflow.
• Completed the active FinalTrim RFF=0 list path after OKSeq 6: retained the final fixed-point nearest-nonrecombinant thresholds, added the correlation-gated first expansion and strict raw-tree parent-bounded second expansion, and preserved ascending candidate admission.
• Ported selected-role list pruning for all three INList positions, retaining swap-last removals, the collapsed/raw matrix branch, correlation free tickets, the source's asymmetric index spelling, and the inherited third-list loop index rather than normalizing it.
• Made final OKSeq 15 membership active and completed ConsensusOK CScore with the exact ordering of the final-membership multiplier, RCorrX, and declared-Long matrix accumulator semantics.
• Ported ConsensusOK's primary list rebuild, score/regional/class thresholds, raw/collapsed topology gates, exact position fallback, six-distance equivalence pass, low-score direct widening, straggler pass, and all-three-list restoration when any rebuilt role is empty.
• Ported the shared selected-role conservative cleanup after the RFF guard: native Single-precision movement sums and representative ranks, raw/direct outlier constraints, swap-last removals, bounded-list admission, and the always-true strict four-matrix inlier branch.
• Ported the primary-RDP branch of the post-group signal recheck: every finalized nonrepresentative candidate is rescanned against its role's two representatives using the supplied LowP * 100000 threshold lift. Emitted, candidate-recombinant, event-overlapping, and ordinarily significant outcomes plus the best tract and probabilities remain separate in JSON, CSV, and review UI.
• Extended erased-tract breakpoint context from immediate contact to the manual's RDP-specific uncertainty rule and the supplied CheckEnds implementation. The port rebuilds the current erased-triplet information map, applies the distinct beginning/ending DLL ranges, source-shaped input MissingData, literal wrap comparison, and linear edge gates. Immediate adjacency, qualifying event IDs, nearest information-rich count, native range, and each warning reason remain separate in result JSON, on-demand alignment JSON, CSV, and review UI.
• Added a line-level supplied-source trace for native RDP breakpoint uncertainty and kept its manual review flag distinct from statistical breakpoint probability/confidence evidence.
• Ported the complete active BURT/PolishBP path as an isolated capacity-reusing unit: supplied three-symbol recoding and circular expansion, DNA5's 20-argument/21-start Viterbi training, pseudocount updates, forward/reverse posteriors, strict 0.995/0.999 bounds, signed nearest- interval matching, breakpoint selection, source-shaped missing-data adjustment, three-usable-site revert, and final gap relocation. A local 15-bit Microsoft-CRT rand() adapter preserves the supplied Windows DLL's seeded emission starts without depending on Emscripten libc.
• Made BURT evidence active before cyclic tract erasure and after representative-only corrections. Statistical 99%/95% ranges, HMM positions, signed containment state, applied/reverted movement, training diagnostics, and unavailable reasons are retained in result/project JSON, on-demand alignment JSON, CSV, and the responsive review UI; manual breakpoint edits remain authoritative.
• Corrected cyclic erasure to the active supplied ModSeqNumY path: both breakpoint coordinates are erased for linear and circular tracts, and the same inclusive definition now drives fragment re-entry, erased-tract review context, recombination-free output, and fragmented FASTA output. Equal endpoints follow the source's wrapped branch and therefore represent the full alignment, including in event-overlap grouping and circular result labelling.
• Reconstructed the current triplet's accumulated native MissingData union before each BURT pass. It combines source-shaped input missing runs with prior inclusive erased tracts that affected any current representative, matching PolishBP's position in the native cycle without copying an alignment-sized mask for every sequence.
• Completed the manual's four common accepted-event recombination-free FASTA variants by adding current-group sequence removal and union-tract column removal beside tract masking and aligned mosaic fragments. All four share ordered-review/downstream-reconciliation gating; empty sequence or column results fail with an explicit explanation. Column removal uses a linear difference mask.
• Added the supplied default-enabled “polish breakpoints” setting across UI, C ABI, worker, project save/restore, and result metadata. Disabling it skips BenHMM work and explicitly preserves the RDP or manually edited coordinates.
• Fed finalized distance membership back into every role's manual two-of-three evidence rule and the automatic co-recombinant group used by cyclic tract erasure. Manual group overrides remain separate.
• Added memoized outside/inside direct, raw-patristic, and collapsed-patristic candidate-pair lookups so the new widening passes do not repeatedly rescan alignment columns.
• Removed heap allocation from the frequently called symmetric tract-overlap test by using fixed two-segment stack storage for linear/wrapped spans, while retaining inclusive and equal-endpoint source semantics.
• Removed temporary-vector allocation and concatenation from every primary triplet candidate pass; the scanner now appends into one capacity-retaining signal buffer without changing call order.
• Split the manual's enabled, masked, and disabled sequence states across the UI, worker/C ABI, core, results, and project replay. Masked rows now correctly re-enter secondary RDP, distance, and phylogenetic evidence; disabled rows remain tree context but cannot enter scans, event roles, co-recombinant groups, or finalized evidence lists.
• Added full, enabled-only, and masked-or-disabled-only FASTA exports around sequence curation. The two complementary partitions preserve original aligned rows, use the current UI state rather than requiring a scanner instance, work directly after alignment loading, bypass unrelated event-decision gating, and reject an empty selection explicitly.
• Restored the manual's large-dataset curation commands as modern controls: reapply the supplied closest-pair auto-mask, enable all, mask all, or disable all. Each operation updates the primary triplet count, keeps masked/disabled sets disjoint, and uses the same unsaved-analysis guard as an individual row edit; this also makes every row controllable when rendering is capped at 500 rows.
• Restored FastRecCheckP's EN != BE detected-event gate so a full circular candidate run cannot create an equal-endpoint primary call; later manual/restored equal-endpoint events retain the active ModSeqNumY full-alignment semantics.
• Reused alignment-sized CheckEnds and BURT symbol/lattice/posterior buffers across events, avoiding repeated capacity growth in both new breakpoint paths; HMM lattice/backpointer/path buffers now follow the DLL's one-clear-per-event lifetime instead of being filled per Viterbi iteration. The compact per-triplet accumulated missing mask retains capacity across events as well.
• Expanded JSON, CSV, and review UI evidence with both expansion stages, selected-role removal, OKSeq 15, completed CScore, ConsensusOK admission stage, final membership, fallback status, and the primary-RDP post-group recheck.
• Advanced engine/source labelling to 0.8.0-session-8 and reloadable projects to v1alpha8, while accepting v1alpha1–v1alpha8.
• Refreshed the Pages workflow to the current Node-24-generation official actions (checkout@v7, setup-node@v7, configure-pages@v6, upload-pages-artifact@v5.0.0, and deploy-pages@v5) and explicitly includes the verified .nojekyll file in the Pages artifact. The application build remains on Node 20 and pinned Emscripten 5.0.1; the Emscripten setup wrapper is the current v16.
• Added a line-level supplied-source trace and updated the parity corpus for order, exact-equality, class-threshold, widening, and fallback cases.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.7.0-session-7

• Added a first-party GitHub Pages Actions workflow. Default-branch pushes and manual dispatches install the locked npm graph under Node 20, provision Emscripten 5.0.1, check TypeScript, build the single-worker WASM/Vite application, upload dist/, and deploy through the github-pages environment with the required least-privilege token permissions.
• Added a Pages artifact verifier that requires the entry point, .nojekyll, Emscripten module, and a binary with the WebAssembly magic header; rejects development and root-relative asset URLs; and rejects symbolic/hard links before upload.
• Added a pre-build source-contract gate that compares every public C header function with its keepalive definition, the CMake export list, and worker calls, then aligns engine/package/lock versions and the emitted/imported project schema.
• Added package-lock.json and npm ci deployment instructions so Actions resolves the same dependency graph on each build. Lock metadata was generated with lifecycle scripts disabled; dependencies were not installed and project code was not executed in this checkpoint.
• Preserved repository-root, repository-subpath, and custom-domain deployment through the existing relative Vite base and document-relative worker WASM URL.
• Versioned the dynamically imported Emscripten loader and its locateFile requests with the same package checkpoint, preventing a newly hashed UI bundle from reusing stale stable-name WASM. The worker also rejects and destroys an engine whose exported version differs from the UI package.
• Kept Pages on the single-worker compatibility module because Pages cannot provide the COOP/COEP headers needed for pthread WASM. The numerical core remains off the DOM thread.
• Added explicit unsaved-project state after scans and every review/repair mutation. The UI now shows dirty/current/saving checkpoint state, warns before tab exit, and confirms destructive dataset, mask/settings, or re-scan replacement until a project checkpoint is downloaded.
• Advanced the engine/source checkpoint to 0.7.0-session-7 without changing project schema v1alpha7 or the documented native-consensus boundary.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.6.0-session-6

• Added a bounded C++/C-ABI/worker endpoint for inspecting the original alignment around both event breakpoints without transferring complete alignment rows to the main thread.
• Added an on-demand graphical breakpoint inspector to ordered review. Current recombinant and parent rows are always first, followed by current/automatic co-recombinant groups, masked traces, and distance/tree supporting evidence; rows are deduplicated and explicitly capped.
• Added 1-based coordinate rulers, circular-origin wrapping, linear-end clipping, selectable ±15/30/60-site context, role/group labels, and parent-aware nucleotide colouring.
• Added persistent per-event warnings when a beginning or ending boundary touches a prior erased tract, plus closest expected major→minor/minor→major informative-state brackets in the inspector and corresponding JSON/CSV fields.
• Corrected the active opening FinalTrim duplicate pass: after a candidate/pair is counted in more than one RList, the diagnostic correlation copy is cleared for every occurrence of that candidate in every role list, including occurrences not independently above the duplicate threshold.
• Ported the source branches that populate FinalTrim OKSeq elements 7, 8, 9, 12, and 13: collapsed/raw whole-event tree position, whole-tract relative JC distance, and both paired breakpoint JC distance scores. Native strict/tied rewards, negative penalties, warning/saturation gates, closest-pair role modifiers, and deliberately asymmetric cases are preserved.
• Added each mapped matrix score, availability/fallback state, and the raw active matrix subtotal to JSON, CSV, and a dedicated review table. These diagnostics do not change membership.
• Ported CalcMatchY OKSeq 17/18 from the supplied VB/DLL sources: exact four-flank reconstruction, VB banker's rounding, signed variable-site map, circular rolling window, regional product, six breakpoint samples, standard grouping thresholds, and source bounds. Scratch vectors are reused.
• Ported the opening ConsensusOK raw-tree topology-consistency rejection for OKSeq 18. JSON, CSV, and review retain both the raw and filtered class, checkpoint values, breakpoint-presence flags, and labelled JC fallback; this evidence remains deliberately non-pruning.
• Corrected every role-indexed comparison to the supplied CompMat ordering, including the role-1 parent order, while retaining the active bare-sequence-index quirk in the positive OKSeq 14 branch.
• Confirmed from the active VB path that OKSeq 10 is explicitly cleared and 11 is never assigned; both are now represented as intentional zero slots rather than missing functionality.
• Ported active FindActualEvents and MakeMatchMatX2P behavior for OKSeq 14: best direct event overlap, inverse-parent rejection, circular/equal-endpoint tract traversal, candidate/reference tract intersection, JC saturation, exact rewards/penalties, and the source index quirk. Prefix counts replace repeated full-alignment rescans without changing interval results.
• Ported CheckPatternX's three-region informative-state shares and the ConsensusOK base inputs OKSeq 0–5 plus RCorrX, including representative sentinel/zero behavior.
• Ported FinalTrim OKSeq 6 nearest-nonrecombinant fixed-point membership with native post- StripDupInv swap order, correlation gates, unfound-event removals, raw/collapsed tree limits, and paired breakpoint-distance veto.
• Added a labelled non-pruning CScore preview through OKSeq 0–6. It preserves the VB NS As Long half-to-even narrowing after each matrix contribution and leaves only pending final membership OKSeq 15 neutral. JSON, CSV, and the review table expose every input and the raw versus source- rounded matrix values.
• Added machine-readable lateConsensus and breakpointInspection capability metadata to result JSON so consumers can distinguish complete early evidence stages from the non-pruning late stack.
• Preserved the exact edge lists, branch lengths, bootstrap support, and working-fragment leaf provenance from all six event NJ trees, then exposed them through a bounded on-demand C ABI and worker request without enlarging ordinary result or project JSON.
• Added a responsive graphical tree reviewer with whole-tract/5′/3′ paired comparisons, arbitrary- root disclosure, current role/group colouring, retained-fragment labels, bootstrap percentages, unavailable-region fallback messaging, and a reversible below-50% branch-collapse view.
• Added machine-readable treeInspection capability metadata alongside the breakpoint inspector.
• Advanced reloadable projects to v1alpha7, accepting v1alpha1–v1alpha7, and advanced the engine/source checkpoint to 0.6.0-session-6.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.5.0-session-5

• Replaced one-shot ordered support rebuilding with strongest-first cyclic discovery: after each event, the current co-recombinant tract is erased and every eligible distinct-origin triplet is screened again from the beginning.
• Added source-shaped fragment re-entry below the supplied 100,000-site cutoff. Working fragments are gap padded, retain original/event provenance, cannot form same-origin triplets, omit short or duplicate copies, and are bounded by an explicit 256-fragment browser cap.
• Made the Bonferroni opportunity count round specific, excluding same-origin synthetic combinations, and retained each signal's correction factor for faithful project replay.
• Replaced linear round-signal duplicate scans with a canonical original-triplet hash index and removed repeated copies of the immutable O(N²) identity matrix during history rebuilds.
• Added fragment-assisted event analysis: saved anchors recover the appropriate working fragment for distance profiles, role metrics, and the bounded event-tree panel while all reported sequence identities remain original alignment records.
• Ported the active MakeACOR topology-affinity inequalities, the exact MakeRList two-correlation override quirk, the distance-triangle warning branch, and active StripDupInv removal of inverse-only candidates. All pass/block reasons remain visible in JSON and review rows.
• Expanded role identification to nine displayed metrics. Eight use the mapped native full/half contribution weights, including raw/collapsed-tree PhylPro, leave-one-role-out, displacement, and weighted triplet-ordering paths; nativeWeightParity remains false while method families are missing.
• Closed the rejected-event workflow hole. Rejecting an event now invalidates the downstream chain, restores its tract during replay, preserves the rejected call for audit, and rediscovers later events in fresh cyclic passes. The UI gates decisions to analysis order.
• Added manual complete co-recombinant group editing with a preserved automatic two-of-three baseline. Current group membership now drives cyclic erasure and accepted-event FASTA variants.
• Enforced event order and rebuild readiness in the C++/WASM boundary, including rejection of final FASTA export while reviews or downstream reconciliation remain incomplete.
• Made pending project snapshots resume from a valid compact prefix: stale later signals/events are omitted on import, retained signal IDs and event anchors are remapped, and the rebuild marker remains.
• Added a one-click project checkpoint directly in ordered review so the manual's save-often loop remains practical during accepted, unreviewed, and pending-rebuild states.
• Ported the active opening FinalTrim duplicate-positive-correlation cleanup as non-pruning, per-pair JSON/UI/CSV diagnostics.
• Advanced reloadable project JSON to v1alpha6, accepting v1alpha1–v1alpha6, and expanded JSON/CSV with round correction factors, fragment provenance, erasure counts, late-correlation gates, and weighted role contributions.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.4.0-session-4

• Added a standalone deterministic phylogenetics core with pairwise Jukes–Cantor distances, neighbour joining, canonical split matching, ten-replicate column bootstrap, and 50% branch collapse.
• Added the six event subalignments and all three paired-tree phylogenetic membership checks, with a 100-sequence tree panel and explicitly labelled Jukes–Cantor fallback for larger active sets.
• Added iterative detectable-set closure and complete at-least-two-of-three co-recombinant groups for each of the three presumed-recombinant hypotheses.
• Ported CalCR's five inverse category relabellings, the dominant-category RCorrWarn gate, and the positive aggregate score/target portion of MakeRList.
• Added an auditable six-metric recombinant-role vote using distance PhPr/SubDist, tree TreePhPr/TreeSubDist, topology change, and three-set support; exact native weighting remains labelled.
• Added explicit application of role recommendations in the ordered review workflow.
• Added accepted-event tract-masked and event-ordered mosaic-fragment FASTA exports, gated in the UI until every event is reviewed and downstream reconciliation is clear.
• Expanded project JSON and CSV with tree regions, bootstrap summaries, all evidence sets, per-sequence phylogenetic diagnostics, correlation relabelling/warning data, role metrics, and final group membership.
• Advanced reloadable projects to schema v1alpha4 while retaining v1alpha1–v1alpha3 import support.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.3.0-session-3

• Added source-shaped five-region extraction around both breakpoints, including the native 60-informative-site boundary walk and boundary inclusion/exclusion behavior.
• Added three paired six-value distance-pattern correlations with exact six-observation Pearson probabilities, native-style overlap eligibility, and uncorrected P < 0.05 membership.
• Added all three presumed-recombinant role hypotheses for every reconciled event.
• Added per-sequence correlation coefficients, p-values, site diagnostics, eligibility, and detectable-support flags to project JSON.
• Added the conservative intersection of detectable-signal and distance-correlation evidence as a two-set consensus subset; it is explicitly not labelled a complete co-recombinant group.
• Expanded the review UI and CSV export with hypothesis counts, detailed correlation evidence, and the pending phylogenetic boundary.
• Advanced reloadable projects to schema v1alpha3 while retaining v1alpha1 and v1alpha2 migration support.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.2.0-session-2

• Added strongest-first reconciliation of repeated RDP triplet hits into detectable-signal event sets.
• Added the supplied greater-than-30% symmetric tract-overlap rule and two-shared-sequence indexing.
• Added relaxed RDP trace-profile checks for masked sequences, with significant and trace-only evidence separated.
• Reworked review around event hypotheses, grouped evidence, and masked trace evidence.
• Added manual role/breakpoint correction and ordered re-identification of later events.
• Added direct project import/resume, session-1 migration, and project schema v1alpha2.
• Changed CSV output from primary-signal rows to reconciled-event rows.
• Kept the checkpoint source-only and deliberately uncompiled per project instruction.

0.1.0-session-1

• Mapped the supplied RDP5 manual workflow into a five-stage browser interface.
• Added local alignment parsing, diagnostics, and the supplied RDP5 auto-mask routine.
• Added the first C++/WASM port of the primary RDP triplet screen.
• Added worker batching, progress, cancellation, provisional-signal review, and exports.
• Added responsive UI styling and static-host-relative asset configuration.
• Documented fidelity boundaries, performance design, and the native parity plan.
• Deliberately left the checkpoint uncompiled per project instruction.