Dräckfink (quantum-dirty-borrowing)
Change workspace ownership across phases rather than merely shortening a local register. Read Bona's mechanism and limits.
Read the shared contract before a trial. Start with the five-line post; the shared native-test order, stop rules and promotion gate apply.
Scheduler versus arithmetic synthesis
Bona assumes blocks already satisfy a dirty-borrowing contract. It does not synthesize a clean-to-dirty replacement for a data register. Require the exact gadget and identity-on-lender proof before constructing a scheduling graph. A helper named and_dirty elsewhere is insufficient: show it actually owns the binding allocation and has a legal host.
If a rewrite is necessary, classify it as a separate arithmetic-synthesis hypothesis and compare its entire added T against the row's remaining budget before building. Use the current front and product budgets in the shared benchmark reference; reopen a rejected attachment only on a changed, quantified premise.
Practical lane
- Extract lifetimes from the active emitted circuit, including all phases tied at peak Q. Label each temporary clean-required or dirty-safe; an unused wire is not automatically a clean wire.
- Build an endpoint/dependency graph. For each proven dirty-safe gadget, find an idle host interval with no operand alias or dependency cycle. Start with unchanged-order placements; price necessary rescheduling separately.
- First experiment: remap one actual peak allocation into an eligible host, with a machine-readable old/new allocation trace. Check the gadget's forward operator against the intended operation tensor identity on the borrowed state.
- Remeasure the whole circuit. If another phase still binds, the result is a local mechanism only. Expand only to a bounded set of co-binders with identified compatible hosts.
Structural extensions
Compare shared workspace scheduling against regeneration and against moving a consumer to expose an interval. These are separate hypotheses: regeneration changes T, rescheduling can change depth and dependency pressure. An allocation diagram without emitted-circuit evidence is insufficient.
Falsify if all peak temporaries require clean state, eligible hosts overlap operands, the splice introduces a cycle, or the unaffected floor prevents the claimed width target. Retain a useful local schedule without labeling it a board gain.
Deliver the dependency graph, borrow/restore boundaries, all peak-owner traces, operator check and full Q/T delta. Do not measure or reset arbitrary borrowed data.