MIP-0081 tasks

Ordered delivery of MIP-0081, one PR per row. §n are the MIP's. Issues: marola-dev/marola-ml#26–#30 (rows 1–5), marola-dev/marola-site#94 (row 6), #706 (row 7).

#	slug	delivers	tests (must exist before the PR)	depends on
1	eval-seam	marola-ml — §5.1: scripts/eval/backend.py, local.py over a pinned lm_eval, just eval <preset> <task>	backend.py --self-test (a fake backend round-trips a ModelRef); one real arc_challenge run on tiny, its file committed	–
2	results-store	marola-ml — §5.2, §5.3: results.schema.json, render.py, check.py, just eval-render, just eval-check	check.py --self-test with eval_check_refuses_out_of_stderr and eval_check_accepts_within_stderr; render.py --self-test	1
3	card-from-results	marola-ml — §5.5: publish_hf.py --results	publish_hf.py --self-test renders a model-index block from row 1's file	2
4	domain-task	marola-ml — §5.4: eval/tasks/marola_ocean/ (YAML, utils.py) from the pinned resources' question set	the task's scorer matches OceanBenchmark's coverage on three fixed answers	1
5	first-provider	marola-ml — §5.7: one provider module, eval-external.yml (dispatch only), the runners job extended to it; the provider's free run checked by just eval-check	actionlint; ci.yml's runners job fails on a test branch that adds a pull_request trigger	2, a provider agreement
6	partners-page	marola-site — the partners page §5.8 promises	its own issue's screenshots	5
7	mip-implemented	marola — MIP-0081 flipped to Implemented with every PR	mip_graph.py --check	1–6