Iteration 02 · 2026-06-09 · driver: native Claude Code /loop (laptop run)

Feature #1 — anderson_darling_a2 (PR #91)

← Dashboard · PR terrylica/mql5#91 · card 110 · stacked on bootstrap #90

What it computes

RAW Anderson-Darling A² of the de-gapped 200-bar strictly-trailing close-logret window vs a fitted Gaussian (sample mean + std ddof=1). Causal — excludes the current bar's own return. None on warm-up (<200 bars) / <8 finite returns / degenerate window. (Card 110's "intra-bar per-tick" prose is superseded by the campaign's bar-close decision — uses inter-bar close log-returns, which forex_bars persists.)

The chain — one SSoT, bit-exact to the column

3-way FOSS finding (it earned its keep)

scipy (ddof=1, SSoT) and statsmodels (ddof=0 MLE) use the identical A² formula (source-verified) and agree to ~2.3e-3 — the difference is ddof convention only (≤5e-3 bound still catches a formula/binning bug). statsmodels underflows to +inf in 72 deep-tail windows (|std-residual| up to ~10.4) where scipy/Rust (log_ndtr) stay finite — positive evidence of the kernel's tail robustness.

Challenge-and-Held (inline; sub-agent dispatch skipped per the tick-3 signal)

• C1 producer integration — held + residual observe() sits at the once-per-emitted-bar finalize point (722→768→780; gap cases drop non-emitted bars). Residual: a live bigblack producer-vs-oracle smoke on real ticks is the final integration check (no CH/MT5 on laptop — deferred to operator bigblack validation).
• C2 FOSS ddof — held (characterized above; bit-exact assurance is Rust==scipy@4.5e-13).
• C3 log_ndtr a≤−14 branch unexercised (fixture max |w| = 10.4) — conceded residual: correctness rests on faithful scipy-port, not parity; robustness for extremes the data doesn't produce.
• C4 tolerance — not relaxed (kernel 1e-9, measured 4.5e-13; the 5e-3 is a labeled cross-convention bound).

Gates / scope

• mise run rust:check GREEN; check:all-gates green on macOS except the worktree-context check:git-hook-installed (literal .git/hooks path; passes from the main checkout). feature-metadata:verify-offline PASS.
• Operator-approved carve-out: a new CompletedBar field required a 1-line Default::default() in 2 blocklisted test literals (mechanical, behavior-identical).
• Hygiene debt: no NEW foreign drift; standing items unchanged (PR #89 merged; signal_ring Linux break = Terry-owned → these anchors macOS-verified, unverified on bigblack pending that fix).

Next

Feature #2 — edge_spread_bps (card 52, bidask.edge SSoT, ε 1e-9). Same chain: oracle → bit-exact kernel → producer column (now that BarOrthogonalExtV1 exists, no carve-out needed) → parity → 3-way FOSS → PR.

Presentation only — never the SSoT.