Run or resume a serial graph with per-node Linux guardians
reflow_artifact_plan_run_bounded.RdValidate all stage resource policies before any writer, execute pending nodes serially and verify cache hits without new supervisors. Scientific definitions and synchronous graph APIs are unchanged.
Usage
reflow_artifact_plan_run_bounded(plan, directory, supervisor_directory,
resources, runtime_files, previous = NULL, cancel = function() FALSE)
reflow_artifact_plan_resume_bounded(plan, directory, supervisor_directory,
resources, runtime_files, cancel = function() FALSE, reconciliations = list())Arguments
- plan
An artifact dependency plan.
- directory
Graph generation directory.
- supervisor_directory
Separate policy/control root with an existing parent.
- resources
Exact named list in canonical stage order; each element has the resource fields documented by
reflow_artifact_run_bounded.- runtime_files
Common named SHA-256 vector covering all writer closures and backend files.
- previous
Optional completed prior generation for verified in-place reuse.
- cancel
Function returning one nonmissing logical cancellation value.
- reconciliations
Explicit per-stage artifact retry reconciliations. Never clears guardian locks.
Details
Control policy and attempt journals are separate from graph inventory. Orphan READY adoption requires a matching successful guardian outcome and fresh locked verification. Missing journals, failed guardian outcomes and changed policies fail closed. Ordinary parent unwind releases its graph/control locks; stale locks after process death require external inspection. Guardian death is unqualified and requires external cleanup. Linux only, trusted descendants, per-process address space, requested threads, and explicit finite deadlines of 1–86400 seconds per node; no graph-wide deadline, parallel scheduler, selector or targets integration.