pi-canon: A Durable Fit

On the final capture of this programme, every result consulted in the treated arm came with a choice: let it leave the context window after one reply, or mark it durable and keep it in view. Across those 24 treated sessions on exact-value contract tasks, gpt-5.6-luna kept all 139 search and read results. That unanimity closed a six-study arc that ran the registered ladder of context-withdrawal designs for pi-canon, a canonical project memory for coding agents, to its end. A transient guidance cue could be withdrawn after its one consuming reply: in a 480-session capture, 277 of the fired guidance lines were withdrawn and the behavioural endpoint stayed inside its registered margin, with the point estimate favouring the withdrawal. Task evidence could not: a fixed working-set window imposed on read results drove sessions into rotating re-read loops that never wrote a word, and the capture was stopped by its registered cost ceiling. Returning control to the model removed the hazard and also the economy: offered an opt-in, 2 of 24 sessions used it; offered a disclosed ephemeral-by-default with a per-call opt-out, the model marked every consult durable, with task quality perfect in both arms. The programme closes with a measured boundary rather than a new feature. A cue whose effect is already incorporated can leave the window; evidence still bound by an exact-output contract could not safely be evicted below the task's working set by the fixed window tested; and when the model decides, retention was unanimous under the disclosed default, with opt-in withdrawal appearing in two sessions of twenty-four and nowhere else. The shipped durable-by-default design of pi-canon 0.3.0, unchanged through all of this, is validated for this use case by the exhaustion of its registered alternatives. All six studies began from frozen registrations and hidden-oracle instruments; five produced graded records, the sixth closed ungraded at its registered cost ceiling by its own design. The supplement beside this PDF is the frozen record: the six designs and their staged amendments, the results and halt documents, the graded reports, capture manifests and per-session manifests, the contracts and instruments as run, the fixtures with their hidden oracles, the harnesses with their mutation-tested gates, the probe records, the paper sources with the figure pipeline, and the external editorial reviews of the paper's plan and drafts, verbatim across three rounds. A facts gate in the supplement recomputes every figure datum and pinned prose number from those records alone, 40 checks, the same gate that ran in the paper's build. Raw session transcripts and store snapshots are excluded for size; the record's notes state the exclusion. This is the fourth paper of the pi-canon line and the first with no planned next step: a validation of the shipped design, and a natural bookend. pi-canon is MIT licensed and on npm at 0.3.0, unchanged by this programme. This record is a preprint and has not been peer reviewed.