The gate reads K from the coverage rule, and every SESOI reading from the registered metric¶

# research(2026-10). Closes the two items left "still to do" above.

K. scripts/k_coverage.py applies the K coverage rule (partitions.coverage_runs, the grid partitions.K_GRID) to an organization's likelihood pilot and its simulations, and writes k-coverage-<org>-likelihood.json, which partition_pilot.py --sizing takes as the K source of a test read, and the only one it takes: partition_pilot.py and require_test_read refuse a pilot's own sizing there. OpenStack K = 10 (the reproducibility K, its null within nominal there), Wikimedia K = 20 (reproducibility K 10, coverage K 20). The rule refuses a grid point of another organization, metric, pilot calibration, projected size or null-study count, a K off the grid or repeated, and a gap below the first K that holds; K values above it are recorded and decide nothing. A test reproduces each committed artifact from the committed pilot and grid.

The registered metric. cells.REGISTERED_METRIC is logprob_per_token, and a test read is refused unless the read, its K source and its bounds were all made on it. h1_test_gate, the replication read and decomposition_gate now take their SESOI from metric_sesoi (0.0183 nats per token on the registered metric), so within_sesoi, meaningful and below_sesoi no longer compare likelihood intervals with exact match's 0.01. metric_sesoi moved beside SESOI in decomposition.py, reads likelihood-sesoi.json from the repository root rather than the working directory, and raises when that artifact is missing instead of reporting the metric unregistered.