The test-efficiency decisions

The test suites, and each lab’s own copy of the same tooling, had grown slow and duplicated. These decisions make the work cheaper without weakening any gate: one runner where there were several, reuse where it is proven safe, and a fixed order so a known slowdown is fixed before it can become the baseline.

The user, 2026-10-15: “go as per your recommandations, but further optimize”.

• The web budget floor: the web batch merges with the floor derived by the budget tool. Its known CPU regression is fixed in the web writer’s next rounds, and the per-file CPU budgets (Q-TX2, Q-TX4) are seeded only after those fixes, so the regression is never frozen in.
• Batched chat reads: the product loads chats in one batched request; the performance check is revised on purpose, not loosened.
• A worker’s import of the guard is loaded from its own small module.
• Q-TXL1: one mutation runner in the harness tooling, used by every lab (replacing their seven).
• Q-TXL2: the memory and browser labs reuse unchanged harness results under the reuse rules of Q-TX1.
• Q-TXL3: one owner per confinement property, settled at the sandbox transfer.
• Q-TXL4: sandbox attack fixtures already owned by a hostile test are removed once the mutation matrix proves they kill no unique mutant.
• “further optimize”: every lane implements its efficiency worklist, ordered by saving per hour.

What is missing: not built yet in this repository: there is no shared mutation runner, no cross-lab result reuse, and no per-file budget seeding.