v0.438.0 — THE KILL TEST SHIP

Stages 1 and 2 of the Navier-Stokes kill test, one band (prepared
2026-09-13 to 2026-09-15; everything since v0.437.0).

B282 (PAPER_2277): THE DEEP TAIL SAMPLE - stage 1. One million
gradient tensors under the personal JHTDB token (value recorded nowhere;
SHIP GUARD v11), inside the database rules: forty sequential 25,000-point
requests, 500k each on isotropic8192 and isotropic32768, zero errors.
- CAP HOLDS on every chunk - worst 1182 ratio 0.0200, forty-fold under
  17/20; no Reynolds trend at 25k depth (B278 envelope flag retired).
- New statistic: the octave stretching efficiency, the scale-resolved
  shadow of the cap - FALLS with intensity on both rungs (~|w|^-0.6),
  agreeing octave by octave to 0.01. Cells above 16x mean (0.6 pct)
  carry 29-34 pct of the stretching. The most intense cell of the
  million is compressed along omega.
- deep_tail_grade() live; stage-1 CSV + provenance in jhtdb_grade/.

B283 (PAPER_2278): THE LOCAL MAXIMUM - stage 2, the kill test taken to
the true local maxima of the
most intense events on the two largest DNS in existence - and the data
hole that manufactured two of them.

- Stage 2 never needed SciServer. The same REST service and the same
  personal token return fourth-order finite-difference gradients on the
  TRUE grid (sint=fd4noint) at every node asked, so full-resolution
  cubes could be pulled where the maxima live: six 64^3 and two 128^3,
  2.6 million nodes, 132 sequential requests, zero errors, around the
  most intense events the deep sample (B282) found.
- Each cube is graded with its OWN maximum. Eight clean cubes on two
  Reynolds rungs, every one between 0.009 and 0.019 - worst 0.0194,
  forty-four-fold under 17/20. The 128^3 cubes hold the same peak node
  as their 64^3 cores and grade lower, so 64^3 is the conservative
  figure.
- The most intense region of the record (isotropic32768): cube mean 75x
  the global mean, peak 11,763x - eleven times what the sample saw at
  that spot - grading 0.012 / 0.011 with a peak-cell efficiency of
  0.074. Inside every cube the stretching efficiency falls with
  intensity (eff(32x)/eff(mean) 0.14-0.45), as it did across the
  million-point sample.
- The Rule-7 finding: the isotropic32768 store has HOLES - whole
  blocks that return zero gradients, deterministic on re-pull - and the
  finite-difference stencil straddling a hole edge manufactures
  |omega| ~ 6,000 spikes two nodes in. Two of the five most intense
  stage-1 "events" on that rung were never turbulence. They are
  rejected with the mechanism named; the 166 zero points of B282 were
  the same hole; two stage-1 chunks are flagged biased low (the chunk
  envelope stands); isotropic8192 is clean in 3.1 million nodes; every
  accepted cube has a zero-free interior and a zero-free halo. Stage 1
  is corrected in place.
- Not claimed: the cap proved; front 1 closed for the whole field -
  eight regions chosen by a million-point sample, in a field of 3.5e13
  nodes. The honest form: the cap holds at the true local maximum of
  every intense event the deep sample could find, with the efficiency
  trend running away from it.

Gate 6,109/0. Dispatch 2,329 keys. Registry 7,125 rows. Acceptance
109/109 (simulator v1.90.0 untouched).
