[aunjgr]

Embedded MO input currently schedules GPU work from each available queue snapshot. This coalesces toward the existing 32 MiB expanded target and flushes on EOS, descriptor/row bounds or input-credit pressure. Producers publish durable pressure before waiting; direct claims enforce the same policy as readiness checks. Immutable string/NULL/size summaries avoid rescanning complete frames and remain charged to the existing 64 MiB input window.

Refs MatrixOne #28966, #28968. Public ABI, numeric semantics, GPU execution and existing cancellation/retry ownership are preserved.

Validation: the native control suite passes 91 cases / 750 assertions in normal, ASan/UBSan/leak-check and TSan modes; three concurrency controls pass 100 repetitions each in one TSan process. The linked GPU suite passes 35 cases / 34,640 assertions, including 1/2/4-stream producers. Pinned formatting/lint hooks pass. All 22 canonical SF1 queries match native MO exactly through the public SQL/schema and bounded row comparer.

Matched Q1 development measurements use the same MO base 64b7a43852, unchanged verified MO GPU artifacts, frozen compiler/RAPIDS provider, datasets, two streams and canonical SQL; one warmup excluded, five measured repetitions. Baseline Sirius is 955b6881; candidate is 585536bf.

Scale	Baseline GPU median	Candidate GPU median	Input units, median
SF1	0.775129s	0.185689s	600 → 12
SF10	9.296148s	2.217461s	7264 → 117

Both scales pass the planned ≥20% latency improvement and ≥90% input-unit reduction gates. These are targeted premerge development results. The dependent MO PR will pin the actual merged native commit and run the full guarded release campaign, including the full-suite regression/resource gates. Numeric optimization and legacy cutover remain separate work.