[ice-black-tea]

Follow-up for #240. This test-only slice removes repeated setup without reducing the supported backend matrix or changing production behavior, test selection, or the workflow.

• Session repository instructions: 64 → 30 real agent turns with all four storage shapes × both checkpoint modes retained. Every combination learns a scope, reattaches a new Runtime, sees the updated rules before its first model request, and learns the second scope. Durable stores still close and reopen for a union readback. A shared SQL + frontier representative keeps same-runtime file overwrite/delete/recreate, unchanged-file reattach, fork, and independent-session isolation.

• Evaluation: the single-lifetime graph-capture authorization test uses real in-memory Persistence; denial, grant, real execution and separate recapture denial remain. Existing SQLite/shared SQL reopen coverage stays.

• Recovery: seed 258 real receipts in one setup transaction, then check committed order and status. The cancel receipt is still after the production page boundary; crash, recovery, cancellation and no-duplicate-effect assertions stay.

• Human scoring: a single in-process decision/rescore test now uses real in-memory Persistence with explicit volatile opt-in. All decision, duplicate-submission and prior-result assertions remain; durable reopen and foreign-coordinator checks are unchanged.

Local scoped evidence on base a642b557:

• Same eight Session nodes: 36.52s → 19.23s, one matched local comparison; both 8 passed. Complete Session repository-instructions module: 36 passed.
• Complete evaluation authorization module: 7 passed; target call 2.73s → 0.43s in one matched comparison.
• Recovery focused test passed on baseline and candidate; measured seed 0.715s → 0.474s. End-to-end node speed was not consistent, so no total recovery gain claimed.
• Human scoring node: 9.02s → 3.47s in one matched local comparison; 3 passed with the opt-in and foreign-coordinator guard cases. This measures the specific test, not CI wall time.
• Independent scoped reviews, selected Ruff checks and diff checks passed. These are local results, not a measured whole-CI speedup. Full CI is the natural PR run; no manual dispatch or rerun.

Prior head d3b8c9d6: Python 3.11 shard 4/5 reported xdist gw2 keyboard-interrupt on an unchanged child-continuation test. Scoped reproductions did not reproduce it, and the log has no causal traceback. Do not treat this independently reviewed scoring change as a confirmed fix; evaluate the latest head's natural checks.

This does not close #240.