EscapistsReintroduced/docs/evidence/crash-start-2026-10-03/README.md

Independent crash start — 3 October 2026

Prison remains the default. A saved Prison/Crash selection is frozen before scenario preparation and replicated for the round. Crash replaces only the detached runtime start definition with an independently authored replicated static wreck scene. It retains the ordinary one-life on-foot admission, finite supplies/shared map, clue sites, radio defence and player-piloted extraction rules. Both the admission/cache area and the native-measured wreck footprint are surveyed before any placement. There is no animated crash, artificial injury, pilot AI or new original-Escapists dependency.

The geometry, first completed crash route, strengthened crash route and repeated default prison route are archived separately. Two terrain-takeover files were independently edited by the user’s other Game Master task after the crash run and during the first default run. That run completed but its strict source guard refused qualification; raw unqualified evidence is retained. A repeated default run checks a frozen copy of the resulting source; its archive retains that snapshot root. The first full 307-program run also finished all programs successfully but detected a further peer attribute edit. The complete suite was repeated against 1014 frozen source inputs and two preserved historical telemetry fixtures, with ERR_REPO explicitly pinned so legacy tests cannot read the changing live tree. The crash launch snapshot differs from final source only in those two peer-owned files; this is recorded, not represented as an exact final-source crash run. The strengthened route observes five retained original clue defenders with native AI active at LOD0 after the existing scripted approach, with no activation override. review.json records exact assertion counts and source relationships. The first crash run failed a fixture assertion equating every enemy's current native activation with prison-hold release. The corrected Workbench-only probe reads the actual hold flag and records native active/dormant counts separately. Native LOD can deactivate agents; the exact cause for an individual agent in that first run was not measured. Production AI activation behavior is unchanged.

All 307 registered regression programs pass on unchanged frozen suite inputs. Three subsequent documentation-only updates are hash-bound in review.json; final code, tests and tools remain exact. All intermediate failed attempts are retained, including a snapshot missing two historical telemetry fixtures and an orchestration-wrapper error; neither required a production-code change. The first suite failure is retained: a convoy-test double needed the newly referenced current-start getter/formatter, with its tested assertions unchanged. Existing actual prison callback tests now exercise both initial phases. New actual-method tests cover geometry guards, isolated definition conversion, once-only server choice, prepared-only progression, saved-command validation, save failure and current/next separation; native boundaries remain doubles.

Each successful native run retains 13 engine errors, so clean qualification still fails. Native route fixtures teleport the protected participant and extraction vehicle and use a shortened defence; they do not prove manual play, real piloting, rendered UI or multiplayer. Original 533 reference files, staged content and four sealed development candidates remain unchanged. The candidates predate these changes. No package or external publication is produced here. See the operator guide at docs/mission-starts.md; next work is a separate refreshed development candidate and remaining gameplay/MP/performance acceptance.