kerfmaster pin list · Agreed & queued (build after fidelity)
rulingmixedA decision by Aristide or Jordan. True because it was decided; it can be superseded, but it cannot be stale.
Content last changed 2026-10-04 — computed from the item itself, not typed.
Contract — unassigned stage
This contract states a decision, not an implementation.
takes —
makes —
fails if Persistence work started before Jordan's dev tests pass or while the ruled optimization pushes (D2 heal, D8 PDF) are still open; a persisted cache key without a code-version stamp; a persisted road with no kill switch or fallback counter.
Contract last changed 2026-10-04 — computed, not typed. Dated separately from the text above, so neither date can speak for the other.
What it is. The speed engine's heal cache, plan cache and plan snapshots (levers A/C family) are MEMORY-ONLY today: every engine restart or deploy throws them away, so the FIRST click on any part list re-pays the full structural fill. Measured on the speed engine at the 1-65 graduation battery (merged tree bcfc8054, engine a3c4117c): cluster A cold first clicks 12.51-16.44s vs warm 0.81-4.37s, against the ~2s adjust target of perf-target-seconds-everywhere. The persistence lever writes those caches to DISK so a restarted or freshly deployed engine serves its first click at warm level (~1-4s), permanently, for every part list. This is the last structural floor on the part-list adjust path -- no in-memory lever can remove it.
Why it waits (the ruling's two conditions). (1) JORDAN'S DEV TESTS FIRST: the graduated levers (A/B/C/D8/E/F1/F2/F3, speed bcfc8054) are with the dev side for Jordan's tests; persistence changes cache lifetime semantics on exactly the surfaces those tests exercise, so it must not land under them. (2) OPTIMIZING EXHAUSTED: the D2 heal push (44558b-7 at 8.64s: heal ~5.0s + planning ~3.5s left; honest odds stated 2026-10-04 -- good for 2-3s more, poor for ~2s on a part that size without a new lever family) and the ruled D8 PDF push (12.31s warm vs ~10s) run first; persistence is the LAST lever so cache contents are final-shape when they start being persisted.
Pros. The 12-16s first click after any restart/deploy drops to warm level permanently, for every part list -- the one class of wait no other lever touches.
Cons and the mitigations that are PART OF THE SPEC when built. STALE CACHE is the real risk: a persisted entry from an older planner served after an upgrade is silent wrongness that the byte-identity gates never see at click time (they gate commits, not cache reads). Mitigation, mandatory: the cache key CARRIES A CODE-VERSION STAMP so an upgraded planner can never read an old entry -- a stamp miss is a plain cache miss, re-filled fresh. Second: disk growth -- a size cap with eviction. Third: one more thing to reason about on deploy/rollback -- mitigated the same way as every Project Speed lever: kill switch env (persisted road off = today's memory-only behavior), counted fallback on /api/diag, byte-identity battery on the build, golden-corpus gate before the commit.
Provenance. Floor named at speed steps 1-56/1-59 (cache-fill structure, flagged as its own lever class, never built on a bare go); pros/cons given to Aristide 2026-10-04 at 1-65; this ruling same day. Ties: perf-defect-closure-plan, perf-target-seconds-everywhere, speed-aware-dev, gcode-identity-gate.