[88plug]

The H59, H59d and H59w pre-registration (#197) needs a runner amendment with tests merged before the first H59-family hold. This is that amendment. It also fixes the H57d runner's Od render for the H57w rows that have not run.

H59 runner (EXPERIMENTS.md "H59 amendment R1")

bench/h59/:

• run.py: the arm table, the hold orders, and host profiles for crescent (Xeon 6741P, H59), the desktop (i9-13900T, H59d) and wildct (i7-13700, H59w).
• Each arm is rendered with k3d-gpu agent manifest --model in the hold's cluster. K3D_GPU_CLUSTER is set, revisions are pinned with K3D_GPU_AGENT_REVISION, and inherited agent variables are dropped.
• The render is staged against the campaign's model directory and checked. An arm whose render is not the registered one (images, model paths, flags, the P-core balloon) does not start.
• run.py also writes crescent's load verdict under H54's void rule, for h59.py --load GATE.json, and runs the prefix-reuse phase (H25's shape).
• hold.sh, hold-run.sh and arm.sh run the registered holds:
  • holds 1-3: N=1, N=4, N=16 (crescent only), TTFT, prefix reuse, in the rotated orders (O, L), (L, O), (O, L);
  • hold 4: quality, in (L, O);
  • spares 5 and 6, each repeating a failed hold.
• Each hold runs once, under the host's own lock, start rule, floors and cap.
• The records use the layout bench/h59/h59.py (ledger(H59, H59d, H59w): pre-register engine choice per CPU class #197 S1/S2) scores; the scorer is unchanged.
• prep.sh downloads the pinned files and checks them against the Hub's sha256. It builds the drafter by running the render's own agent: llama.cpp backend the fit picks, and busy Intel cards left out of the device plugin #182 conversion Job once.
• bench/wildct/row.sh gains h59w and h59w-prep.

The amendment also records:

• every flag, read against the engine's source at its pin (llama.cpp b11960, OVMS v2026.4.1);
• the engine rule as merged in agent: llama.cpp backend the fit picks, and busy Intel cards left out of the device plugin #182.

H57d runner fix (EXPERIMENTS.md "H57d amendment R2")

bench/h57d/arm.sh rendered Od without K3D_GPU_CLUSTER. Since #183 that render goes through context k3d-gpu, which a hold's kubeconfig does not hold. So Od lost its P-core label and got 1 CPU, which is H57dp's deviation in #247. Od's render now gets the hold's cluster. The amendment governs the H57w rows on wildct, which have not run.

Reported for the owner, not acted on: the desktop load gate

On the H57d records and the H57dp records (#247), the foreign P-core load read over the N=4 phases moves with the arm, not with the host:

• during OVMS's phases the host's other work leaves the P-cores;
• during the llama.cpp arms' phases (8 decode threads) it stays.

In the arms' 60 s idle windows the arms match: H57d spread 1.46, H57dp 0.21, against the 2.00 tolerance. Over N=4 they do not: 2.61 and 2.70. That is the post-treatment trap H54 amendment R4 fixed on crescent.

Changing the H59d/H59w gate is the owner's decision. No gate is changed here. The runner records both readings.

Tests

• bench/h59/test_run.py (new, in CI, 33 tests, about 3 min):
  • renders both arms through the real launcher against the fit fixtures, with a kubectl stand-in that answers only for the hold's context;
  • runs arm.sh against a local stub server for four desktop holds, then scores them with h59.py end to end (SUPPORTED on a 2x-faster L);
  • covers crescent's N=16 phase and void re-run, the gate branches, the prefix phase, hold.sh's guards and the shim.
  • checks the stub server's request log against the registered counts and settings (warmup, N=1, N=4, TTFT, prefix, GSM8K, the 900 s boot limit);
  • refuses a sibling CPU SKU (the exact model number the engine rule matches), a spare 6 before spare 5 or over a measured spare 5, and a hold that started but recorded no rc.
• bench/h57d/test_h57d.py: renders Od through arm.sh and reads the P-core label and 16 CPUs. It failed before the fix.
• tests/wildct_hold_test.sh: the h59w rows.

Run locally:

• bench/h59/test_run.py, test_h59.py, bench/h57d/test_h57d.py, bench/h54/test_h54.py, bench/wildct/test_wildct.py: OK;
• tests/wildct_hold_test.sh: 41 passed;
• tests/ci_wiring_test.sh: 70 passed;
• tests/claims_check.sh: clean;
• scripts/images-lock check and tests/image_pins_test.sh: pass (the test file holds no tag-only image string);
• tests/ledger_drift_check.py: agrees;
• console/ledger-rules.json: matches;
• shellcheck v0.11.0 at warning: clean.

No prediction, hypothesis or verdict cell changed. Only the method column of H59, H59d, H59w and H57wa gained a link. No host was touched and no hold ran.