Repository navigation
[Post-0.31][Cost] Measure observed per-VM stack high-water marks #36
Description
Follow-up to #35, explicitly deferred by the 0.31 decisions. 0.32 is a planning target, not a delivery commitment. This is outside 0.31 acceptance.
The 0.31 reports provide static single-VM call-chain bounds. Add observed stack use only with a defined measurement contract.
Decisions: define the entry stack baseline, instrumentation boundary, transient outgoing reservations, EXEC replacement, SPAWN child identities and unavailable outcomes. Never add independent VM stacks into one fictitious call chain.
Acceptance:
- Version the metric and record VM/process identity and instrumentation configuration.
- Test nested calls, outgoing arguments, loops, early failures and supported external-process boundaries.
- Demonstrate measurements against hand-calculated fixtures and distinguish static bounds from observed peaks.
- Keep existing cost ceilings and acceptance oracle; run backend and CI gates.
Activity
a19q3 commentedon Oct 4, 2026
Completed within the 0.32 implementation scope. This closes the observed per-VM measurement task, not a release/public-deployment claim.
cellscript-execution-trace-v1contract records each(vm_id, generation)separately. The baseline is SP after argv loading and before the generation's first instruction. Transient outgoing reservations count; independent SPAWN stacks are not summed and EXEC starts a new generation.The published research contract and evidence and gate/provenance record retain their original source identities: clean CI/backend source
6a453d3015784784a27ba1aa0a4b6b1a73db79a8, with the exact dev adjustment documented. Existing cost ceilings were preserved; 6,782 comparable metrics passed with zero regressions.Closure audit on 2026-10-04: recomputed all seven archived uncompressed byte counts/SHA-256 values in
validation.json; all match. Re-rancargo test --locked -p cellscript --test cost_measurement --test policy_artifact_checker -- --test-threads=1on the 0.32 working checkout based on35bf983d: 13 measurement tests and 9 checker tests passed. This focused replay is additional evidence, not a claim that the historical full gates were rerun at the latest client commit.