Cross-swap refuted and reversed → the range confound → the Zarr→NodeRow bake (plan + crate + W0/W1)

.claude/board/AGENT_LOG.md

@@ -1,3 +1,55 @@
+## 2026-08-13 — weather-poc W0/W1 fleet (4 Sonnet workers, disjoint files; orchestrator-consolidated)
+
+4/4 completed, 0 errors. Plan `.claude/plans/weather-soa-bake-v1.md`; the first
+grindwork wave on the crate `weather-substrate-poc-v2` named and never created.
+Files were disjoint by construction (`key.rs` / `floor.rs` /
+`manifest.rs`+`data/*.tsv` / `probes/weather-p1/era5_variable_census.*`); the
+orchestrator owned `lib.rs`, both manifests, the root exclude entry and every
+board file. Guardrails §1 pasted verbatim into all four briefs; no worker ran
+cargo (rule 7) and all three Rust workers reported "not compiled, not run"
+rather than claiming green.
+
+**D-WXS-1a census (Python, the one worker allowed to run).** Reads the store's
+`.zmetadata`, emits the JSON, `--selftest` asserts all 10 constants the plan
+carries. PASS, and re-run independently by the orchestrator. Guard
+disable-verified twice — the worker broke a constant in a throwaway copy, and
+the orchestrator repeated it against the committed file (exit 1, correct
+"the fetched answer is authoritative" message). 0 arrays failed to classify.
+
+**D-WXS-2 key codec.** 5/5. Exhaustive 1,038,240-cell round-trip, collision-
+checked. Bar B1 disable-verified by the ORCHESTRATOR, not asserted by the
+worker: zeroing the HIP lat byte kills 3 tests (collision, ragged,
+out-of-range); removing the seam split kills the wrap twin while the non-wrap
+twin stays green — the pair discriminates in both directions.
+
+**D-WXS-3 floor.** 7/7. Bar B2 disable-verified: widening the "narrow" control
+floor kills only the control, the stay-silent twin holds. The worker chose a
+hand-rolled FNV-1a over `DefaultHasher` *because the latter is randomly seeded
+per process* — a real trap, and one the suite does not cover, so the
+orchestrator verified it separately: `floor_version` is byte-identical across
+three separate process invocations. ε = 0.03 derived from the percentile
+construction, not tuned post-hoc.
+
+**D-WXS-1 manifest.** 13/13, TSV 22 rows (F0 10 / F1 6 / F2 6, reserved slots
+emit NO row — exactly plan §2.5). Collision guard disable-verified: removed →
+only `colliding_entries_are_rejected` fails while BOTH stay-silent twins
+(distinct lo/hi on one pair; row reordering) stay green. Bar B0's end-to-end
+half is DEFERRED and said so in the module doc — the bake does not exist yet.
+
+**The wave's real find is a plan defect, not a code defect.** The key worker
+noticed during implementation that §1.2 assigns one WHOLE byte per axis while
+OGAR's cascade doctrine — the passage §1.1 leans on — specifies the axis bytes
+nibble-interleaved (Morton). The plan deviated from the canon it cites and did
+not say so. Recorded as plan §1.3a + board `D-WXS-2a` with a pre-registered
+comparison, NOT silently rewritten: the prefix-scan claim holds under both
+layouts, and rewriting a spec the moment a worker surfaces a consequence is how
+a correction starts building.
+
+Gates (orchestrator, one shared `target/`): `cargo test --manifest-path
+crates/weather-poc/Cargo.toml` **25/25**; `cargo clippy --all-targets -D
+warnings` clean; `cargo fmt` clean. Commits `010da851` (scaffold), `a8a501bc`
+(census + key), this one (floor + manifest).
+
 ## 2026-08-11 — Workflow `wf_99d677e6-b45` — eval-plan verify/attack (13 agents; orchestrator-consolidated)
 
 13/13 completed, 0 errors, 0 empty; ~4.42 M subagent tokens, 165 tool calls,

.claude/board/EPIPHANIES.md

@@ -1,3 +1,151 @@
+## 2026-08-12 — E-THE-REGIME-LADDER-MEASURED-RANGE-NOT-TURBULENCE-1
+
+**Status:** FINDING `[G]` — measured, same run, found by an operator
+question rather than by any gate. Qualifies
+`E-THE-HYPOTHESIS-REFUTED-CLEANLY-AND-REVERSED-1`, written one commit
+earlier, whose *stronger* half it withdraws.
+
+**The operator asked: do we have a storm-modelling problem, since we do
+not actively model vortices?** Checked, and the answer is yes, twice over.
+
+**First, the plain fact.** `substrate_comfort_d_cz_2_7.py` loads **only**
+`mean_sea_level_pressure`. No wind, no vorticity, no rotation enters any
+bar. D-CZ-0/1 fetched 10 m winds solely to *report* `spd_sigma`; nothing
+consumed them. R4 "STORM" is a **scalar pressure-gradient** regime, and
+the hypothesis it was built to test was about *"high velocity differences
+/ turbulence"* — a dynamical property that appears nowhere in the
+measurement.
+
+**Second, the confound that follows.** Chasing the question produced two
+measurements that qualify the previous entry:
+
+| relationship | statistic |
+|---|---|
+| `L` vs the cell's `saturation` | Pearson **+0.917** (n=8) |
+| `L̄[T]` vs the regime's own value range | **Spearman +1.000 — perfect** |
+
+`L` is essentially *how much of the target falls outside the donor's
+codebook range*. A regime's own range determines how hard it is to cover
+— R4's is ~**18×** R1's. So C3's monotone rise **restates the width
+ordering**, which the `|∇p|` ladder itself produced (deeper low ⇒ steeper
+gradient ⇒ wider range). Width is not the whole story — `R3 → R2` has a
+*wider* donor yet 0.949 saturation, because the boxes sit at different
+absolute pressure levels — so the real driver is **coverage** (width AND
+offset), which `saturation` captures directly.
+
+C4 is inflated the same way: `Δ` was amended to RMSE **in Pa**, and RMSE
+scales with range, so an absolute-Pa margin cannot be compared across
+regimes differing 18×. Normalised as a ratio it reads **3.96 / 1.85 /
+1.14 / 2.35** — not monotone, and R1 is the extreme, not R4.
+
+**What survives / what is withdrawn:**
+
+- **SURVIVES:** the hypothesis is **NOT SUPPORTED**. No sign flip in
+  either measure, normalised or not; `CAL-ABS` wins its own diagonal in
+  all four regimes. Unaffected by the confound.
+- **WITHDRAWN:** *"cleanly reversed, monotonic"* and *"storms are LESS
+  forgiving of bad calibration."* As measured that says **wide-range
+  boxes are harder to cover with a foreign codebook** — arithmetic.
+
+**Why no gate caught this.** C1c *did* pass, and it was the right gate to
+have — but it measured the structure of the **`|∇p|` field** (decay
+length, Gini, tail ratio), not rotational structure, and the confound is
+between the ladder's own discriminator and the codebook's coverage, not
+between two regimes. A gate cannot catch a confound that lives *inside*
+the variable it was told to trust. **The operator's question was the
+instrument here; nothing in the apparatus was positioned to ask it.**
+
+**The transferable rule.** *When a regime axis is built from a scalar
+derived from the same field the codec quantises, the axis and the codec's
+difficulty are not independent.* Before reading a cross-regime encoding
+result as physics, check whether the regime discriminator predicts the
+encoding's own failure mode — here, `|∇p|` predicts range, and range
+predicts saturation, and saturation IS `L`. A dynamical hypothesis needs
+a **dynamical** discriminator (ζ = ∂v/∂x − ∂u/∂y, Okubo–Weiss) and a
+**range-normalised** transfer metric, or coverage will masquerade as the
+finding every time.
+
+**Filed as D-CZ-8**, the pre-condition for re-asking C3/C4 as a turbulence
+question rather than a width question.
+
+## 2026-08-12 — E-THE-HYPOTHESIS-REFUTED-CLEANLY-AND-REVERSED-1
+
+**Status:** FINDING `[G]` — measured, both pre-registered bars, one run.
+`probes/weather-p1/substrate_comfort_d_cz_2_7.json`, plan §7.
+
+**The hypothesis this whole plan existed to test is refuted, and refuted
+in the SHARPEST available form: not "no effect", but a clean monotonic
+reversal.** The operator's hypothesis — *a badly-calibrated substrate
+that maps dynamically preserves MORE structure in strong storms than a
+well-calibrated absolute one* — was tested two ways, pre-registered, on
+an equal-budget cross-swap over four regimes (calm → ocean → active →
+storm):
+
+- **C3 (transfer loss)** required `L̄[R4] < L̄[R1]`. Measured:
+  `L̄` = **0.011 → 0.309 → 0.671 → 0.690**, R1→R4 — a **62× increase**,
+  monotonic across all four tiers. Storms are *less* forgiving of a
+  foreign absolute calibration, not more.
+- **C4 (the crossover)** required a sign flip against the diagonal —
+  absolute winning in calm, dynamic winning in storm. Measured: absolute
+  wins its own diagonal in **all four** regimes, and its margin **grows**
+  from ~1 Pa (R1/R2) to **10.78 Pa** (R4) — the opposite of a crossover.
+
+Both measures point the same direction. This is not two noisy nulls; it
+is one coherent, reversed relationship measured twice by independent
+instruments (`ρ` on rank vs RMSE on the diagonal).
+
+**A real construction bug was found and fixed en route, via the gate
+designed to catch exactly this class of defect.** C0 — *"both controls
+must be WORSE than every real arm in every regime; if either matches a
+real arm anywhere, that cell measures nothing"* — failed on its first run
+in R1 and R4: `GEO-DEGENERATE` was built from `truth[:len//64]`, the
+first slice of an array already shuffled by `rng.choice` for equal
+budget. A random subsample of a flat array is not reliably narrower in
+range than the whole array; that is a materially different construction
+from D-CZ-1's correct one (`p[si,sj][:n_i,:n_i]`, a genuine 2-D corner).
+Fixed by carrying each regime's full 2-D box alongside the flat
+evaluation sample and building the degenerate donor from a real spatial
+corner. **Disable-verified**: reverting to the flat-slice construction
+reproduces the exact original failure; the fix reproduces the identical
+real-arm numbers (C2–C6 unchanged) while making C0 pass cleanly in all
+four regimes. The bug lived entirely in the control; nothing about the
+real arms was ever wrong. Recorded because it is this arc's C0 gate doing
+precisely the job it was built for — catching a defect in the apparatus
+before it could contaminate a verdict — on the very first probe that
+tried to use it for something other than a smoke test.
+
+**And it contradicts an earlier exploratory hint, on purpose, stated
+plainly rather than averaged away.** §6.6 measured, *within R4 only*, a
+single arm's saturation correlating with storm `\|∇p\|` at ρ = +0.444
+(p = 0.058, explicitly labelled not significant, not a result). That
+correlation's *direction* matched the hypothesis. The properly-powered,
+pre-registered, cross-regime test says the opposite. Where an
+unpre-registered n=19 single-regime correlation and a pre-registered
+cross-swap disagree, **the pre-registered result governs** — and this
+entry exists so a later read of the plan does not quietly split the
+difference between "p=0.058, direction matches" and "reversed,
+monotonic, two independent measures" into a false middle.
+
+**C1c passed and licensed the whole exercise** — the regimes measurably
+differ in correlation structure (decay length, Gini, tail ratio of
+`|∇p|`), not merely in gradient magnitude, so the refutation above is a
+finding about calibration under real structural variation, not an
+artifact of four copies of one condition.
+
+**What is NOT concluded.** One box size (16°, 4225 cells), one variable
+(MSLP), three timesteps plus 19 storms. C5 — the *other* half of "good
+geometry vs badly calibrated" — was never run: the golden index floor
+(N ≥ 2,550,409) is three orders of magnitude above what a box this size
+can hold, so `GEO-GOLDEN-HI` has no admissible construction at this
+scale. The geometry axis of the operator's original framing remains
+completely untested.
+
+**Cross-ref:** `E-A-HORSE-RACE-IS-NOT-A-CROSS-SWAP-1` (the instrument
+this run finally exercised); `E-A-CONTROL-THAT-CANNOT-LOSE-IS-NO-CONTROL-1`
+(the exact defect class C0 caught here); `E-A-FIGURE-CITED-TWICE-IS-NOT-
+CONFIRMED-ONCE-1` (why §6.6's direction is not treated as confirmed by
+this section merely because it was written down once already).
+
 ## 2026-08-12 — E-A-FIGURE-CITED-TWICE-IS-NOT-CONFIRMED-ONCE-1
 
 **Status:** FINDING `[G]` — measured. Probe

.claude/board/INTEGRATION_PLANS.md

@@ -1,3 +1,82 @@
+## 2026-08-13 — weather-soa-bake-v1 (PLAN; the missing Zarr→NodeRow bake)
+
+Plan: `.claude/plans/weather-soa-bake-v1.md`. **Does not supersede
+`weather-substrate-poc-v2.md`** — §9 grades it part by part. The gap it closes:
+poc-v2 names `crates/weather-poc` and declares "New repositories required:
+ZERO", but the crate does not exist and the plan describes **no bake step at
+all** — which is why the entire weather arc ran as Python probes over a Zarr
+file, one variable, four hand-picked 16° boxes, three hand-typed timesteps.
+
+**The bake.** One cell = one node, one timestep = one Lance version. The key:
+HEEL = `(lat>>6, lon>>6)` = the 16°×16° tile, HIP = position within it, TWIG
+dormant-reserved — so a lat/lon grid is the literal-x/y binding OGAR's
+256×256-centroid-tile cascade sanctions, and **the arc's hand-picked 16° boxes
+become HEEL-prefix range scans**. The grid needs 5–6 of the 12 available
+quaternary levels per axis (4⁵ ≥ 721, 4⁶ ≥ 1440), leaving room for 0.05°
+without touching the stride. Three wrinkles stated, not designed around: both
+axes tile **raggedly** (721 = 11×64+17; 1440/64 = 22.5), longitude is cyclic so
+a seam-crossing box is **two** prefix ranges, and `MailboxId ≠ NiblePath` in
+code so nothing leans on it.
+
+**The value.** L4 `6×(8:8)` palette256², one byte per ERA5 field, pairs being
+physical couples (`(u:v)`, `(T:q)`) so a wind-vector distance is ONE 256×256 LUT
+read. Argued against the whole le-contract §3 catalogue: L1–L3 are relational,
+L5 kept only as the 3-D-wind alternative, L6 is the Odoo factoring, **L7 is
+location and location is the KEY**, L8 is retired for this data (§12.12 RETIRED
+the `Pair48` mint), G1–G3 are forbidden to new classes. **Stated deviation:**
+L4's canonical Fisher-z read is replaced by the linear `helix::RollingFloor` —
+licensed by le-contract §3's own monotone-bounded-continuous-field demarcation
+and by §12.1's measurement (Fisher-z burns 228/256 buckets on ERA5 temperature).
+Circular variables never ride an L1 byte (359°/1° at max L1 distance); wind
+rides as `(u,v)`, and `wind_speed` is baked separately because §6.5 measured the
+Jensen gap **is** gustiness (mean ratio 1.115, max gap 14.37 m/s).
+
+**Capacity, corrected.** "32 facets/cell" overstates the usable budget: 2 slots
+are key+edges and `VALUE_TENANTS` carves the slab contiguously through offset
+220 (committed assertion: **188 of 480 B**), so the free budget is **292 B = 18
+facets = 216 payload bytes** (discriminant 15 reserved for `BoardAggregates`;
+weather takes 16, offset DERIVED). All 122 fields still fit at 1 B/field —
+**one cell is one node** — but at 2 B/field they do NOT fit as 4+12 facets
+(336 > 292). The 14 statics get their own classid and ONE version (baking them
+per timestep would rewrite ~1.3 PB of constants). W1 bakes a minimal justified
+set, not all 122, so the layout does not presuppose the fidelity gate's outcome.
+
+**Time.** `E-MARKOV-TEMPORAL-STREAM-1` applied literally: a timestep is a Lance
+version, a time series is a version-range read (`QueryReference::at(v, rung)` +
+`deinterlace`), the 19-storm set is a version set. Zero versioning code — the
+`LanceCycleWriter` / `VersionedGraph` surface is consumed. Open risk,
+pre-registered: 92,044 versions on one dataset is unmeasured (`D-WXS-6`, with a
+KILL).
+
+**Re-homing.** `D-WXA-5` moves onto the substrate AND is re-specified: poc-v2's
+ρ ≥ 0.98 is at serious risk of being vacuous because D-CZ-1 §6.4 measured
+real-arm ρ spread at 3e-6…4.7e-5 — so the bar becomes (a) ρ ≥ 0.9996 (where a
+real pair genuinely failed at 0.999556), (b) the shuffled control must fall
+below 0.98, and (c) a 16/64/256-level ladder must be monotone BEFORE any verdict
+(the can-it-DIFFER half). `D-CZ-8` moves on as ζ from a neighbour-key stencil —
+with a falsifier the arc never had: **differencing amplifies quantisation
+error**, so quantised-ζ vs true-ζ is reported per ζ-magnitude decile with a
+laminar stay-silent twin. On §7.9: the full grid **does NOT dissolve the range
+confound** (`L`~saturation +0.917, `L̄`~range +1.000 is arithmetic that 1.04 M
+cells do not repeal); what it buys is (1) coverage-matched donor selection by
+construction, (2) "donor" meaning a ζ-regime rather than a range once the shared
+floor makes variables commensurable, (3) a regime sample that is not n = 4.
+
+**Splits.** Judgment (layout, key, every bar, the `D-WXA-5` re-spec, the
+confound treatment, the mint) stays on Opus/main. Grindwork is
+one-file-per-Sonnet-worker (`key.rs` / `floor.rs` / `lane.rs` / `bake.rs` /
+`statics.rs` / `stencil.rs` / `metrics.rs`, one test file per bar), guardrails §1
+pasted verbatim, tag-files only. Orchestrator-only: `lib.rs` mod lines, the crate
+manifest, `canonical_node.rs`, every board file. Root `Cargo.toml` untouched —
+`weather-poc` is workspace-EXCLUDED. SIMD only via `ndarray::simd::*`; ndarray
+optional, feature-gated, GIT-sourced (never an optional `path` dep). No
+lance-family bump. Doc-only.
+
+**Also recorded:** `weather-substrate-poc-v2.md`'s `D-WXA-*`/`D-WXB-*`/`D-WXC-*`
+ladder has **no rows on `STATUS_BOARD.md`** (`grep -c WXA` = 0, verified) — its
+board hygiene was never discharged, which is a plausible mechanism for how a
+whole arc ran past an unmet gate.
+
 ## 2026-08-12 — substrate-comfort-zones-v1 REVISED (§2 rebuilt: horse race → cross-swap)
 
 Same file, `.claude/plans/substrate-comfort-zones-v1.md`, still **ACTIVE**,