Minor cleanups from the #4 parallel-scheduler review

[qinhaihong-red]

Parent

#4

What to build

A batch of small, individually-below-the-bar follow-ups from the #4 branch review. Each should be resolved or explicitly declined in the PR.

• Adjacency duplication: the scheduler rebuilds dependents/in-degree from the workflow, duplicating the graph walk the ordering layer already performs (and calls), and its raw len(needs) omits the ordering layer's known-reference filter. Extract one shared adjacency helper so the kernel's graph topology has a single definition.
• record_node_skipped robustness: it is a bare INSERT with no conflict handling; skip-safety lives entirely in the scheduler's seen-sets. A future resume/retry caller skipping a node that already has a row would raise an unhandled integrity error. Make the store path defensive (or document the precondition).
• Test cleanups: the peak-concurrency test hand-rolls an O_EXCL mkdir mutex + counter files with a sleep window (timing/portability flake risk) where the barrier-based tests already prove the contract both ways; and two concurrency tests inline raw workflow dicts because the shared shell_workflow helper cannot set concurrency — extend the helper and dedupe.
• definition_checksum includes concurrency: decide whether a pure scheduling knob belongs in the checksummed definition-of-work (same family as the earlier "checksum is version-internal" finding; no current consumer reads it back). Record the decision.

Acceptance criteria

• One shared adjacency/in-degree helper used by both the ordering layer and the scheduler
• record_node_skipped no longer raises on a pre-existing row (or its precondition is documented and enforced by the caller contract) — done in PR feat: failure semantics — retries, timeouts, cancellation, and resume #69 (commit 487de89): the resume re-skip path UPDATEs an existing row guarded by _started, with a regression test.
• The peak-concurrency test is simplified to the barrier pattern; the concurrency test workflows use the shared helper
• The concurrency-in-checksum decision is recorded (kept with rationale, or moved out)

Blocked by

None - can start immediately

[qinhaihong-red]

PR #69 (the #6 failure-semantics work) resolves the record_node_skipped robustness item in this batch. The predicted scenario materialized exactly: on resume, re-skipping a node that already has a node row from the prior run raised sqlite3.IntegrityError: UNIQUE constraint failed: node.run_id, node.node_id.

Fixed by enforcing the caller-side precondition rather than a defensive store: _skip_transitive_dependents now flips an existing row with an UPDATE (guarded by the scheduler's _started set, mirroring how the launch path already chooses UPDATE vs INSERT) instead of a bare INSERT (commit 487de89), with a regression test covering the resume-re-skip path.

The other three items in this issue (shared adjacency helper, peak-concurrency test simplification, concurrency-in-checksum decision) remain open.