#476: the two halves of a run share a lifecycle

[Masterplanner25]

Summary

Fixes #476 — options (1) and (2) of its fix direction, both directions of each.

Reproduced on main first: the issue's repro (rm -rf .nodus/workflow_framework/runs — the cleanup CLAUDE.md described as safe) made a live waiting run permanently unresumable behind Workflow run '…' not found, with its graph state and checkpoint sitting on disk.

Symmetric cleanup (option 1), both directions:

• nodus workflow cleanup now deletes the run record with the graph state, reported as run_records_removed in its output. Guard: if the record says the run is still live (the administrative mark_waiting shape — a waiting record over a completed snapshot), it is not cleanup's to take without --force.
• The store's opt-in max_terminal_runs cap now prunes the graph state and checkpoint with the record it deletes — a host that caps records wants the run gone, not half of it.
• Stores gained delete_run(run_id) -> bool — a concrete no-op default on the WorkflowStore ABC (an abstract method would break every host store implementation to gain a capability only cleanup uses), overridden by LocalWorkflowStore and SQLiteWorkflowStore. The SQLite version goes through a new _exec_rowcount helper: the existing #516 source-assertion (conn.execute only inside the cursor-owning helpers) caught my first cut calling conn.execute directly, which is that test doing exactly its job.

Honest diagnosis (option 2): a resume whose record is missing while load_graph_state finds state now says so — "has graph state under .nodus/graphs but no run record in the workflow store — the two halves of the run were cleaned independently" — with category: "run_record_missing", plus the recovery hint (workflow cleanup --force for the orphaned state). An id with neither half still says not found, pinned by a falsifiability control.

Option (3) (one owner) not taken — recorded on the issue as the long-term shape, with #380 the natural companion.

CLAUDE.md's "safe to rm -rf" note is corrected; it was this issue's repro.

Test plan

• tests/test_run_store_lifecycle.py (# closes: #476): cleanup removes both halves and reports both lists; the issue's rm-the-records repro now yields the honest diagnosis instead of not found; unknown ids still say not found; the record cap prunes graph state with the record (timestamp-tie made deterministic); SQLite delete_run round-trip.
• 338 store/cleanup/workflow/resume/runner/framework/wait tests pass (including the SQLiteWorkflowStore leaks cursors: the store depends on CPython refcounting to commit #516 cursor-discipline source assertion, now green); ruff + mypy clean; closed-issues gate passes for all eight [Unreleased] issues.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QBC6mwW25PMi72GqL9Q9vX