Green runs (issue-273)

Every command below was run on the final tree, after the self-review rewordings.

Unit — the coupling's gate order and the pending context

$ uv run pytest -q -p no:randomly cli/tests/test_graphlink.py cli/tests/test_graph_drive.py
65 passed

Covers T1: the exempt actions (start, context) and the gated ones (advance, clean); that an exempt action records no graph.skipped and a gated one still does, naming the action; the declared-specDir parity across both; and the pending context — the graph's start node, the status, and that it is not a waiting human gate.

Integration — the reproduction, against a real dispatcher and a real runtime

$ uv run pytest -q -p no:randomly \
    cli/tests/test_graphlink_integration.py \
    cli/tests/test_harness_config_scaffold_integration.py
22 passed

Covers T2, T8 and T10, all Gherkin-documented:

• A work item minted as a plain ticket starts at the human gate — the pointer lands on phase-selection, loop:phase-selection is set, one checklist is posted asking for the-loop execute, and no graph.skipped is recorded.
• Starting the graph is not the same as walking it — an unauthorized the-loop execute on the freshly started graph leaves the pointer where it is.
• The prompt is rendered before the graph is entered — the block names the node, says NOT ENTERED YET, names the human gate, and carries no the-loop graph complete line.
• pending never masks a work item in flight — a started graph reports its real node and a status that is not pending. Green on both sides of the change, by design: it is the control for R2.3.
• The adoption suite's plain-ticket spawn now writes graph-state.json where it used to assert its absence, and records no skip.

Full suite

$ make test
uv run --project cli python -m pytest -q cli
2482 passed, 1 skipped

Eleven tests are this work item's: six new (two in test_graphlink.py, four in test_graphlink_integration.py) and five rewritten in place, where the assertion that existed pinned the behaviour the ticket reports. Eight of the eleven failed before the production change (see red.md); the other three are guards and controls that are green on both sides by design — the advance still-skips test, the started-graph still-reports-its-node test, and the event-log skip record retargeted to advance.

The suite count therefore rises by the six new tests: 2476 passed, 1 skipped before, 2482 passed, 1 skipped after.