Iteration 03 · 2026-06-09 · driver: native Claude Code /loop (laptop run)

Feature #2 — edge_spread_bps (PR #92)

← Dashboard · PR terrylica/mql5#92 · card 52 · cut off bootstrap #90

What it computes

EDGE effective bid-ask spread (basis points) over the raw 200-bar strictly-trailing OHLC window — the EDGE estimator of Ardia-Guidotti-Kroencke (JFE 2024). Causal: the window is the 200 bars BEFORE bar i (current bar excluded). None on warmup (<200 bars) / degenerate (<2 price-change periods) window. Substrate distinction from #1: edge consumes raw open/high/low/close DIRECTLY — NO log-return, NO 8 h gap-NULL; edge's internal cross-bar terms span weekend/Ouroboros gaps unmodified, exactly as the discovery probe's ohlc_win substrate.

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

3-way FOSS finding (it earned its keep)

bidask.edge_rolling is a genuinely INDEPENDENT implementation of the SAME estimator — it does not call edge() (verified); it uses a rolling moment-matrix algebra vs edge()'s per-window de-meaning. Alignment empirically pinned: edge_rolling(df,window=200).iloc[i-1] == edge(O[i-200:i]). The two agree to 3.54e-14 over 11200 windows — the summation floor, with NO convention gap (contrast feature #1's statsmodels ddof difference). No second mainstream Python EDGE port exists (frds 2.4.1 lacks it; only cross-language R bidask::EDGE otherwise). Triangulates with Rust==edge() at 9.78e-11.

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 (next to populate_pure_bar_ranks). Residual: a live bigblack producer-vs-oracle smoke on real ticks (no CH/MT5 on laptop — deferred to operator bigblack validation).
• C2 independent-leg integrity — held (edge_rolling verified to not call edge; different algebra; 3.54e-14 agreement = summation floor).
• C3 pairwise summation — held: kernel reproduces NumPy's pairwise sum exactly; residual 9.78e-11 is the libm::log-vs-system-log ~1 ulp floor amplified by EDGE's E[x²]−E[x]² cancellation, ~10× under the locked 1e-9 (documented, not a defect).
• C4 tolerance — not relaxed (kernel 1e-9, measured 9.78e-11; the 1e-6 FOSS bound is a labeled cross-implementation bound, orders larger than the 3.54e-14 measured agreement).

Gates / scope

• mise run check:all-gates green on macOS except (a) the worktree-context check:git-hook-installed (literal .git/hooks path; passes from the main checkout) and (b) ONE foreign env-drift — see handoff. feature-metadata:verify-offline PASS.
• Operator-approved carve-out: the new CompletedBar field required a 1-line Default::default() in 9 explicit literals (2 blocklisted test files) — mechanical, behavior-identical. Same pattern as #91.
• BarOrthogonalExtV1 field-union: cut off bootstrap, this branch creates the struct fresh with edge_spread_bps; it field-unions with #91's anderson_darling_a2 at operator merge (mechanical 3-way merge on data_types/writer/schema). Merge order: #90 → #91 → #92.
• Hygiene debt (NEW, foreign): check:research-orthogonality-null-utils-tests 33-vs-34 — its substrate-gated smoke test SKIPs when the /tmp substrate parquet is absent; owned by the rust-enhancement loop; scripts/research/ untouched by this campaign → operator-follow-up, gate NOT weakened.

Next

Feature #3 — hvg_clustering (card 54, networkx.average_clustering, EXACT+ε). Same chain: oracle → bit-exact kernel → producer column (now BarOrthogonalExtV1 exists on this branch's pattern, still re-introduced per branch off bootstrap) → parity → 3-way FOSS → PR.

Presentation only — never the SSoT.