# Full source regression after chapel integration

Started tools/run_all_tests.py with Godot4.7.2 and output artifacts/chapel-full-regression-2026-09-29. all-tests.json is incremental and does not establish full completion until the runner exits. Source hashes at launch are in source-start.json. The runner excludes exported-package tests and the kernel module invoked by run_tests.gd. A separate packaged-build acceptance is still required.

Read-only audit identified obsolete21-facility expectations in Runtime integration, Runtime stairs and exported opening. The two source tests were independently run before changing them and each failed only on that old count. Updated expected count to22, including chapel, and added explicit chapel registration assertion. Targeted reruns: Runtime integration21 checks PASS; stairs67 checks PASS. Export expectations updated for the future rebuilt pack; no exported package has been verified by this action.

Only these three test files changed after source-start.json during the sweep; the production source was not changed by this regression repair. The full runner will later execute the corrected source tests. Preserve before/after logs instead of treating the original failure as a gameplay defect. Final source hashes and the complete result must be checked when the runner finishes.

Runner exec session18695 was confirmed live during this turn, processing test_chapel_boundaries.gd. Continue polling the same session; do not restart solely because a poll times out.

## Public catalog lookup correction during regression

Read-only audit found that `FacilityCatalog.by_id` omitted ORIGINAL_FACILITY_PATHS. Chapel was therefore registered in normal Runtime construction but returned null from this common query. Expanded the lookup to include original-facility paths without altering baseline/optional expansion lists. Added a contract test that every normally registered resource resolves to the identical definition through the public lookup. Before: one failure out of310 (chapel only). After:310 PASS, exit0. Evidence: catalog-lookup-repair.json and before/after logs.

The sweep had already run the previous266-check registry test when this correction was made. Production catalog and registry test are two additional files changed since source-start.json, beyond the three count-repair test files. Therefore retain the sweep result and this targeted correction as distinct evidence; do not claim the sweep alone covered one frozen final source revision. No production caller of by_id was found in scenes, scripts or autoload besides the catalog definition; normal Runtime builds its own registry from explicit paths.

## Completed source sweep

Runner session18695 exited0. All127 unique source suites passed with zero failures. Recomputed discovery matches executed paths exactly (no missing or unexpected suite). `completion.json`, `all-tests.json` and `source-end.json` retain the result. Five previously documented files differ from launch hashes: three count repairs plus catalog lookup and its regression test. The latter's final310-check run separately passes. This is combined regression evidence, not a single immutable-source full sweep. Export/PCK and Windows are not included. New paired restart-debug fixture/runner were prepared separately and were not part of this discovery list.

After this successful sweep, started rendered `run_opening_restart_flow.py` into artifacts/opening-restart-debug-2026-09-29. Read its result.json and phase logs for completion; launch alone is not acceptance.