Experience-Densified Completion Pools: Verified Splicing and Parameter Lifting for LLM Agent Serving

Compiled backward completion pools give a domain-specific LLM agent free per-step lower bounds on remaining cost, but that system discards every verified trajectory it produces. We extend it with a harvest-and-serve loop: solved trajectories are harvested into a pool of witnessed executable suffixes keyed by template, bound goal, and state, and replayed as verified splices -- inside a sandbox snapshot, at zero LLM calls, accepted only if the executed diff equals the bound goal. Because acceptance is verified, proposals may be heuristic: sigma-lifting rewrites a stored suffix's values to fit a fresh instance of the same template (a substitution existed for 80.5% of eligible pairs, a shape-level ceiling rather than a transfer guarantee). Over 180 streamed calendar requests, success rose from 49% to 81% while LLM calls fell 3.6-fold. The system matched or dominated request-keyed and goal-keyed plan caches everywhere and was always cheaper; on goals never seen before, only lifting produced hits (5 of 35 strict first encounters, an existence demonstration; every cache variant zero). Seeded offline by a careful solver, an agent that alone solved 2% of project-management requests reached 71%, its cumulative cost crossing below the no-experience curve by request 25 of 40, seeding bill included. An attribution study orders the mechanisms: verified sigma-lifted replay carries the margin, and on the seed-poor domain the lifted cache alone is strongest; the interval machinery contributes a certified trigger and proposal persistence, and a slack-aware trigger preserves the mechanism under loose bounds. Under fault injection, verification rejected 33 of 33 corrupted proposals and restored the sandbox exactly. Verified replay converts an agent's own successes into a safe, growing source of free