[osipovartem]

Summary

• Make cached BatchSplitStream outputs consume the existing cooperative scheduling budget. They previously bypassed the upstream yield points.
• Share the existing Tokio, Tokio-fallback and per-stream policies without changing the public CooperativeStream API.
• Keep first-slice and unsplit upstream-polled outputs free of an additional budget charge. Preserve row order, zero-copy slices, errors, empty batches and split metrics.

Targets df55-physical-udf-instance, the runtime branch used by Rustice, rather than the diverged upstream-sync branch.

Validation

• cargo fmt --all
• Strict workspace Clippy: --all-targets --all-features -- -D warnings
• Complete ./dev/rust_lint.sh, including rustdoc; repository Rust 1.97 and CI-pinned license/typo tools
• Extended workspace tests: 10,781 passed, zero failed, 10 ignored
• Focused stream/cooperation tests: 11 passed in each of the default, per_stream and tokio_fallback policies
• Independent source review completed; no remaining blocking correctness, API or test findings

Local Cost Screen

Same native MemoryStream -> cooperative -> BatchSplitStream -> drain pipeline, frozen dependency/feature graphs and explicit optimized CI profile (opt-level=1, Rust 1.99). Inputs contain 1M Int32 rows; batch size 1,024 produces 977 batches. Order, buffer sharing and metrics are checked before timing; every timed sample checks row/batch totals.

Eight fixed-CPU processes follow ABBAABBA order, four per revision, with 20 warmups and 1,001 samples per case/process. Medians below are the median of four run medians; tails pool 4,004 samples per revision/case.

Case	Median Before/After (us)	p95 Before/After (us)	p99 Before/After (us)
Unsplit fast path	58.624 / 58.039	70.943 / 69.466	87.378 / 88.435
Cached splitting	68.625 / 66.895	83.263 / 88.737	100.924 / 115.033

The cached-path tail increase is explicitly accepted as an observed cost of restoring cooperation: +5.474 us at p95 and +14.109 us at p99 in this synthetic screen. The existing 128-unit policy is retained; ordinary unsplit outputs receive no new charge and slicing/allocation behavior is unchanged.

The earlier, separately retained short cohort (two warmups, 101 samples per process) showed larger cached-tail increases: +16.92% at p95 and +33.93% at p99. The cohorts are not pooled, and attribution of every tail increment remains incomplete. Independent review accepted this narrow fairness/cancellation tradeoff. These workstation measurements establish neither a production latency bound nor a general query speedup.