Bootstrap — mint the v3 cohort and pin the substrate

✅ Done — PR1 #124 OPEN (base of the stack)

Status: PR1 #124 OPEN · feat/odb-ext/rqw-crex-bootstrap · base=main — first in the squash-merge order.

• v3 cohort minted (not v2-extended) — orthogonal_extension_cohort_v3.rs + forex_orthogonal_extension_v3_ordered_names_list.json — a brand-new sibling registry, kept strictly separate from v2.
• Anchor pinned three ways — Rust const + JSON ordered-names list + .sha256 sidecar. 4 anchor-pin tests recompute the SHA + HARD-HALT on drift. The anchor was re-locked after the #7 witness correction (−0.605 → −0.5838646869163802).
• All slots are None-stubs — so test_schema_contract stays green; no BarRow / COLUMN_NAMES / schema.sql change until the feature PRs land.
• Numpy-pinned SSoT / tolerance doc — the gap-aware substrate recipe is committed verbatim from multislice_orthogonality_probe.py with the pin numpy 2.5.0 / scipy 1.18.0 (uv --python 3.14); tolerance is VALUE-within-1e-9.
• 3 foreign hygiene fixes (all SAFE, logged in HYGIENE_DEBT.md): clippy 1.95, git-hooks worktree-compat, meta-test GIT_DIR.
• Dashboard placeholders — this hub + the two feature iteration pages (so flowchart clicks never 404).
• Gates: all green on this diff.

The load-bearing substrate (in plain terms)

Both features measure the shape of the recent returns, so they must agree on exactly which returns to look at. The recipe is transcribed from multislice_orthogonality_probe.py (the k_crex / k_rqw paths) — not the discovery harness, whose sign-dropped, no-gap-null variant would diverge silently.

1. Take the trailing 200-bar window of closes.
2. Compute logret = diff(log(close)).
3. Null every inter-bar gap > 8h — weekend and Ouroboros weekly-reset jumps are not real returns, so they become NaN.
4. Drop the NaNs.
5. Require at least 8 finite returns; otherwise emit NULL for that bar.

This is the no-silent-drift discipline: we port the pinned convention exactly rather than "improve" it. Parity is verified VALUE-within-1e-9 against a pinned numpy oracle (numpy 2.5.0 / scipy 1.18.0), not a byte-sha of a frozen blob.

The two hardening witnesses (front-loaded)

• #38 — scale-witness. CREx is scale-sensitive: raw vs standardized inputs differ by roughly 1/std. A reviewer who z-scores the returns "to normalize" would silently break parity. The witness asserts that z-scoring FAILS — locking the raw-input convention.
• #7 — sign-witness. RQW carries a sign. The witness pins a concrete value: XAUUSD@25, open_time_us = 1768179397541000 → RQW = −0.5838646869163802 (binding, frozen fixture). Any sign-drop regresses this. Note: the contract's −0.605 was a stale full-production figure that no numpy quantile method reproduces — the anchor was re-locked to the binding value. The kernel reuses the already-shipped RollingPercentile HF7 primitive rather than re-implementing octile interpolation.

Orthogonality KEEP evidence (already satisfied)

Both features cleared the KEEP test on 2026-06-12-xi-batch34-crossed-verdict.md and are not re-run here: #38 ξ 0.160 / 0.395 (close) — independent; #7 ξ 0.135 / 0.260 — independent. This run is production wiring, not discovery.