Skip to content

session-26-handoff.md

Latest commit

 

History

History
71 lines (60 loc) · 4.79 KB

session-26-handoff.md

File metadata and controls

71 lines (60 loc) · 4.79 KB
Raw

Session 26 handoff — RDP-only cyclic parity audit

Version: 0.26.0-session-26 Project schema: org.rdp-web.project/v1alpha19

This checkpoint starts from the Session 25 source archive and compares the browser core against ten simulated, RDP-only saved analyses. Inputs were passed opaquely to the scanner; comparison code read only WebRDP result JSON and the numeric saved-analysis tail after the RDP5 Recombination Data marker. The supplied desktop source remained read-only and was not compiled. No alternate RDP implementation was consulted.

Source semantics restored

  • RDP candidate breakpoints now follow the supplied CentreBP midpoint rule. The detector retains the first/last informative coordinates for scoring, then stores the midpoint between adjacent informative sites. After cyclic erasure, a midpoint that lands on missing data is shifted toward the detected tract exactly as in the supplied routine.
  • The durable BestXOList-style action cache now covers every committed event, not only restored or manually fixed events. Ordinary cyclic reuse is deliberately narrow: the same original triplet and more than 80% tract overlap. This prevents the same acted-on signal from creating fragments indefinitely without hiding neighbouring events that merely share a pair.
  • The supplied AddjustCXO/MakePairsP scheduling rule is active. Current XOverList signals that touch RList(WinPP) enable their three original-sequence pairs; those permissions are propagated to every selected RList member, and a dirty next-round triplet runs only when all three pairs are enabled. A sparse pair-key set replaces the native dense DoPairs matrix without changing the decision. Unchanged signals remain in the carried XOverList shortlist.
  • ModSeqNumY erasure uses the selected role's final ConsensusOK/FinalTrim distance list. The diagnostic two-of-three evidence union remains visible but is no longer erased as though it were RList(WinPP).
  • The active MakeConsensusC subset quantizes PhPr-family and distance statistics to the supplied precisions before tie decisions, applies the exact post-rounding ±1 family sentinels and supplied tree-distance cap. The TreePhPr second-place contribution is the supplied 14 points, not a generic half of its 18-point win.
  • Automated cyclic scanning has a fixed 64-round safety ceiling. Reaching it preserves all completed events, records cycle-limit-reached, and proceeds to Review/export instead of iterating without a diagnostic bound.

Validation result and boundary

The midpoint rule is strongly confirmed: whenever an ordered event role matched in the focused opening rounds, stored boundaries frequently became exact numeric matches, and the probability did not change because centering is post-score. The saved source membership arrays also confirmed that the selected role's final distance list—not the two-of-three union—is the erasure group for every same-role event examined.

The ten-case audit is not full cyclic parity. Original RDP reports 336 events. Before the source pair gate, this checkpoint reported 509 events, matched 82 ordered-role events and 152 unordered triplets, reached the 64-round ceiling in five cases, and took 125.5 seconds. With the source DoPairs gate it reports 453 events, matches 95 roles and 171 triplets, reaches the ceiling in one case, and takes 72.7 seconds. Heavy triplet-kernel entries fall from 1,399,142 to 307,911 (78%), while total audit time falls 42% because late event grouping/tree work remains substantial. The gate improves all three aggregate parity measures, so it is retained as a general source invariant rather than a fixture-specific adjustment. Exact MakeConsensusC role selection and source event consolidation remain the next causal deficits.

An experimental partial TestMoveInTreeAlt acceptance gate was rejected from active behavior. It reduced total events but worsened early ordered parity because source RDP's mutable PermValid/PermDiffs, vQuickDist, and final tree stages must be ported as one state machine. The helper trace remains non-authoritative and is not called by event selection.

Recommended next checkpoint

  1. Port the remaining active MakeConsensusC inputs (dMax, RCompat families, list correlations, outlier checks, and combined prizes) with the exact source role order and rounding.
  2. Complete source XOverList/BestXOList event consolidation after the now-active MakePairsP scheduler, including how alternate retained records are redirected into existing events.
  3. Port TestMoveInTreeAlt only with its mutable pair totals, UFDist, vQuickDist, and final Clearcut tree decisions together; then rerun all ten opaque comparisons.
  4. Keep the 64-round cap during this work so a parity regression terminates with inspectable results.

Full native parity remains unclaimed.