H2O as the default miner adapter base

New text jobs pin h2o-lightning-4b-v1.2.3. Existing jobs retain their frozen profiles. This is a training-base change; the shared Jev inference endpoint and ImaJev image profile retain their identities.

Install and qualify

1. Install the subnet source pinned by subnet-ref.txt and complete its H2O miner setup. Keep the Python 3.13 platform environment and install H2O's separate Python 3.12 GPU environment. Set ZILS_H2O_RUNTIME_PYTHON on miners and processors.
2. Rehearse 202610100001_h2o_training_profile.sql with make check-queue-db, then apply it through the existing migration ledger. It adds H2O and updates profile routing; it does not rewrite jobs, releases, leases, or historical evidence.
3. Qualify each H2O worker on its actual hardware, including training and adapter reload at the 2,048-token boundary. Register its exact profile/runtime hashes in zils_worker_profiles with the measured memory guard and evidence. H2O approval is explicit: the migration does not copy JevK5 qualifications.
4. Start processors with the H2O reference plus references needed by existing jobs. Start miners with the same references. Then set ZILS_TRAINING_MODEL=h2o-lightning-4b-v1.2.3 on the training API and processor. Verify /v1/config and a newly created job both expose the H2O profile.

From the subnet checkout, the hashes for operator registration are:

.venv-kev/bin/python -c 'import json; from zils import models; from zils.miner_grading import trainer_identity; print(json.dumps({**models.profile_identity(models.H2O), "trainer_sha256": trainer_identity(models.H2O)}))'

The processor uses the existing platform command with the new primary reference:

export ZILS_H2O_RUNTIME_PYTHON=/absolute/path/to/zils/.venv-h2o/bin/python
.venv-platform/bin/python -m zils_platform.coordinator process \
  --state .private/processor --reference /absolute/path/to/zils/models/h2o-reference \
  --additional-reference /absolute/path/to/zils/models/jevk5-reference \
  --device cuda

Load the existing server environment before running this command. Only include references installed and retained by the operator. Keep the same protected ZILS_COMPUTE_LOCK across all processes sharing a GPU. A runtime upgrade changes trainer fingerprints; graded routing needs fresh H2O benchmark contexts and qualification reports. JevK5 scores cannot qualify H2O jobs.

Add a separate h2o section to the existing workflow configuration. It uses the same fields as the existing text section: hotkey, min_free_mib, releases, runtime_url, token_env, and either registry or register_command. Retain the original section for legacy jobs. The H2O release directory must differ from the legacy directory, and its measured min_free_mib must be at least 12,288. Optional training_services, nvidia_smi, gateway_runtime_url, and release_group retain their existing meanings. Missing H2O configuration leaves its jobs waiting rather than assigning them to a legacy runtime.

Serving accepted adapters

Each adapter runtime loads one pinned base. Keep H2O and JevK5 releases in separate private release directories and route their registry entries to the corresponding runtime. A runtime rejects mixed-base catalogs and never loads a JevK5 adapter into H2O. The existing --shared-model-dir mode remains JevK5-only.

Run H2O adapter serving with the H2O interpreter, the subnet and platform source checkouts on PYTHONPATH, and the platform's requirements/api.txt plus the subnet's requirements/rehearsal.txt installed in that isolated environment. Do not install the platform package itself into the Python 3.12 environment; its control-plane package requires Python 3.13. Keep runtime credentials server-side and preserve the existing loopback HTTP contract. Publish and activate only adapters that passed the independent evaluator.

For example, with a privately provisioned ZILS_H2O_ADAPTER_RUNTIME_TOKEN:

export ZILS_SUBNET_DIR=/absolute/path/to/zils
export ZILS_PLATFORM_DIR=/absolute/path/to/zils-platform
uv pip install --python "$ZILS_SUBNET_DIR/.venv-h2o/bin/python" \
  -r "$ZILS_PLATFORM_DIR/requirements/api.txt" -r "$ZILS_SUBNET_DIR/requirements/rehearsal.txt"
mkdir -p .private/h2o-releases
PYTHONPATH="$ZILS_SUBNET_DIR:$ZILS_PLATFORM_DIR" "$ZILS_SUBNET_DIR/.venv-h2o/bin/python" \
  -m zils_platform.adapter_server --releases .private/h2o-releases \
  --reference "$ZILS_SUBNET_DIR/models/h2o-reference" --port 8932 \
  --token-env ZILS_H2O_ADAPTER_RUNTIME_TOKEN

--reference starts an empty H2O adapter slot before the first release exists. Only verified accepted releases are exposed through the runtime health catalog.

Rollback

Restore ZILS_TRAINING_MODEL=jevk5-4b-v0.3 to change subsequent jobs. Keep H2O references, workers and serving processes available for H2O jobs and releases already created. Do not delete the additive migration or mutate frozen profiles. Retain the previous code/runtime installations until the replacement passes its hardware checks. Source changes alone do not update a running service.