iteration 05 · 2026-07-14

DONE — PR #147 OPEN (stacked on #146) card 80substrate: ret_spread (first two-input)locked-ε margin 0.16 worst

#80 gini_correlation_asymmetry_signedreturn_spread

Branch feat/odb-ext/gini_correlation_asymmetry_signedreturn_spread → terrylica/mql5 PR #147 (base = iter-04 branch; loop never merges). First-run green: parity 5/5 · robustness 4/4 · rust:check 0 · check:all-gates 0 · pre-push ✓ · zero fix cycles.

H5 results row (ledger detail)

Substrate	Tier/TOL	Parity batch / stream (max-abs-dev, SPLIT)	non-vacuity	pinned witness	C&H
ret_spread (return leg RG8 gap-masked; spread leg unmasked — mask WITNESSED in the parity gate)	ε 1e-9 rel / 1e-12 atol (LOCKED, np.isclose semantics)	2.407e-13 / 0.0 — 11,200/11,200 within gate; worst locked-budget margin 0.16 (verified 0 failures)	11,200 compared; 4/4 cells; H3 golden 12,000; H1 ×11,200	EURUSD@5+0 → −0.28071837007811623 (RG8 gap window, within ε)	A 7/8 + 1 RESOLVED · D 10/10 · 0 conceded

H2: SWAP-SIGN fork 60/60 with bit-exact negation identity asserted (max-abs in the 1.403 band) · ORDINAL fork 80/80 with probe-identical tie targeting (max-abs 2.285e-03 band). Oracle sha bda6d242… (replay byte-identical); golden sha abd3ebd3….

Decisions + findings (recorded LOUD)

A number that LOOKED like a breach, investigated not rationalized: the attacker's pure-rel metric read 1.416e-09 > 1e-9 — first-hand verification showed the LOCKED tier (rtol·max + atol) holds on all 11,200 pairs with the worst at 16% of budget (small-magnitude value ~1.28e-4 — the atol term's exact purpose). Checklist ruler corrected for iter 06+; artifact untouched.
RG2 by design: np.mean's pairwise summation ⇒ byte-parity out of scope for #80; bit-unequal 10,998/11,200 is the expected signature, NOT drift.
Un-forkable sign convention: the suite proves swap = bit-exact negation on every window.
Clean-room: GiniDistance (GPL) never read; SY-1987 formula-only; net-new (no crypto twin).

In plain terms

Measures the return-vs-spread relationship from both directions (value of one against the rank of the other) and reports the difference between the two readings. A nonzero gap means direction and spread are linked asymmetrically — the liquidity-cost response to moves isn't exchangeable.

What landed (lockstep, one feature commit — 19 files)

kernel gini_correlation_asymmetry_signedreturn_spread.rs (clean-room average-tie ranks) + TWO-RING rolling state (kendall card-94 plumbing) + mod.rs
BarOrthogonalExtV1 field · BarBuilder 4-arg wiring incl. bar.spread_close · writer.rs BarRow (last field)
schema.sql DDL AFTER sign_markov…flux · feature_metadata.yaml · regenerated docs/feature-catalog.html
oracle + oracle/golden jsonl + .sha256 sidecars (dual-dir)
parity suite (5 gates incl. the leg-mask witness) + robustness suite (4 gates incl. the negation identity) + C&H record

Next: iter 06 (#78 hoeffding_phi_squared_midreturn_duration — the FIRST duration substrate; the run's final today-feature) stacks on this branch.