docs(server): record that the fold skip was inert on prod and rolled back The page said "not enabled on prod", which is no longer true and would have misled the next reader. Enabled 2026-10-01 and rolled back within the hour. The reading: observed=290, skipped=0, refolded=0 — the observer ran and NO fold decision happened, because prod runs NOETL_STATE_BUILDER=offserver and the drive therefore never calls rebuild_state. Also records why there is no worker-side consumer: on the pod that actually runs offserver, cold_rebuild is the cheapest outcome (0.23ms), everything a certificate could remove is 16.4s over 2d19h, and 87% of state-builder time is the "incomplete" outcome, which a certificate cannot help. Adds the metrics table, because observed-beside-skipped is what made the inert run legible rather than looking like a clean ramp. Refs noetl/ai-meta#366, noetl/ai-meta#367
docs(server): chain-certified fold skip + the single-replica constraint Documents NOETL_CHAIN_CERT on the deployment spec (wiki-maintenance Rule 2a) and adds a page for the constraints that are easy to violate: - the skip is sound only at ONE server replica; a replica increase must unset the flag first, and divergences() is a backstop not a licence - MAX_COMMIT_BATCH must not be raised while it is on (-0.90% at 512 in-gate, -2.66% at 2000 out-of-gate) - why a run of skips is bounded at 16: reconcile only learns the real chain length from a fold, so an unbounded run of skips makes the miss-detector unable to ever fire - why all six INSERT sites observe rather than the nominal chokepoint, which repo history shows is not one (ai-meta#263) Refs noetl/ai-meta#366