Update Penalty (New) scenario to accepted TIP-004 (per-algorithm scope)

[m4r1m0]

Summary

Aligns the Penalty (New) scenario with the accepted TIP-RFC-MT-0004 (Last Modified 2026-08-19, Status: Accepted). The TIP no longer groups RxM/RxT into a shared RandomX penalty class — each of the four algorithms is now its own penalty scope.

Changes

• js/config.js — removed PENALTY_CLASS_MAP; reworded the PENALTY_CAP comment (cap applies to per-algorithm runs).
• js/simulation.js — runCompetitionSW/computeAlgoRatesSW now track lastWinner/consecutiveCount per algorithm instead of per class. Only the algorithm that mined the previous block receives the min(2^n, 32) modifier; an RxT block following an RxM block now resets the run and pays no penalty (previously continued the RandomX run at 2x).
• js/simulation.js (2nd commit) — fixed a pre-existing off-by-one in the New path's run tracking: the counter now holds the run length ending at the previous block, so a block extending a run at position r pays m = 2^(r-1) exactly per the TIP's cap table (previously the 2nd consecutive block paid 1x, the 3rd 2x, etc.).
• js/lwma_sw.js — comment-only update (engine is scope-agnostic).
• README.md — class wording replaced throughout; source date → 2026-08-19; config table documents per-algorithm scope; "Why no penalty" critique updated to the TIP's corrected figures (E[m]≈1.16, ~1.7% mean block-time inflation) instead of the old draft's class-based numbers.

Baseline, Penalty (Original), and WTEMA scenarios are untouched, preserving the comparison.

Verification

Test suite run headless on the VPS (Node v22.23.1, vm-context harness over the real block data, blocks 294400–296521):

• T1: LwmaWindowSW with m=1 is numerically identical to LwmaWindow (499 targets, 0 mismatches).
• T2: deterministic modifier checks — only the leading algorithm is penalized, 2^k for k preceding same-algo blocks, cap at 32 for k=6.
• T3: end-to-end runCompetitionSW, 1522 simulated blocks: every block's recorded multiplier matches min(2^(r-1), 32) for its run position; 190 real-data RxM↔RxT switches all reset to multiplier 1 (the class-based behavior would have continued the run at 2x).
• T4: Original and WTEMA paths unchanged (same result counts, no exceptions).
• T5: scenario generation — 22 scenarios, lwma45pnew with penaltyMode: 'new' present.

All tests passed. Note: "Penalty (New)" results will shift relative to the previous class-based version (lower effective penalty, less block-time inflation) — expected, per the accepted TIP.