[Clear1nsight]

The full recovery query does not need 104 aligned copies simultaneously. This contribution streams each source through one work bank into eight persistent full-cover query sums, then hosts those nine roles in nine existing Z_U output banks whose initial definitions are proven unused and deferred.

The complete 224-output F17 recovery now requires zero additional owned banks. It keeps PR #3985's 224 H1 children, 448 native wrappers and 352 cover maps, while reducing CLEARs from 536 to 416 and scalar allowance from 27,184 to 26,944. All 224 current X/Y fields remain borrowed and unchanged. Only nine output definitions move. Whole-word private scratch peak remains 16; the changed future block has no owner-private body.

The exact source def-use census restores the nine charged CLEAR/X-copy definitions before the first preserved post-query Z update. Nine temporary terminal clears disappear because those same owned banks are immediately reinitialized as outputs. This does not erase borrowed state or assume free allocation. The standalone nine-external-bank stream is retained as a regenerated, independently replayed comparison, including its separate zero/identity return.

PROOF.md supplies the ordered operator proof, streaming induction, precise lifetime and complete invoice. WORD.json is the full 3,813-event hosted word. The verifier authenticates 22 canonical runtime files before inherited execution, regenerates the complete PR3985 comparison from canonical sources, checks the new word on both original and independent-current routed arrays, and checks the standalone reference. It verifies all 73 phase coordinates, both signed native kernels on 114,688 columns and all 224 native occurrences. Eleven semantic controls cover dirty initialization, repeated use, premature clearing, native/maps, private scratch across native calls, illegal aliases and paid output restoration.

The access ledger is computed from the actual words. Under the explicit one-full-cover-pass-per-scalar-instruction protocol, the hosted block including moved output definitions uses 698 scalar passes, 1138N scalar reads and 698N scalar writes, while the prefix loses 18 passes and 18N reads/writes. Native/map/wrapper traffic, tape-head movement, layout, instruction control and bit precision remain unknown. These are scalar-access counts, not a physical movement tariff. No κ improvement, workspace optimality or supplier admission is claimed.

This is the assigned streaming-linear-algebra / pebbling contribution in the canonical investigation. Main 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority revision 6, was refreshed before the one bounded output-hosting pivot. PR #2722/#2794 price different fixed access words. The parallel materialized-copy hosting contribution #4007 has a different private-scratch/access schedule (26,960 scalar allowance versus this stream's 26,944); this contribution independently verifies the nine-output deferral and imports no sibling executable or scientific data.

Credit ywz65 / #2707, Arham Tahir / #2681 and the retained QQ16/native/geometry lineage; #3985 for the rectangular correction; the parallel output-lifetime agent for the deferred-output insight and the coordinating agent for the bounded nine-bank composition. Nick Petalas / #2722 and Arham Tahir / #2794 are distinct access precedents. Substantial OpenAI Codex assistance is disclosed; no independent human or formal-kernel review is claimed.

The next bounded step is an actual caller's ownership and terminal use of the unchanged 224 output banks, followed by a complete layout/movement and numerical-error budget. Fewer temporary banks alone does not establish productive-family coverage or either all-depth supplier cap.

Validation completed on Python 3.12.3: registered isolated optimized/warning-strict primary; normal and optimized complete read-only replays; Git-free read-only export; whole-word and typed-receipt mutation rejection; source authentication before execution; retained-data integrity; overwrite refusal. Separate agent proof/def-use review and coordinator optimized full replay passed on the frozen science hashes. REVIEWS.json states their exact bounded scope and the preliminary block-end error corrected before final replay.

Publication head 0fff81d8997cef646538bcd4f60d7d455cf93b8e was created through the GitHub connector from canonical 20ecf2891d80b5acf18d980b2164fb44ea7faa90. PASS_RESEARCH_COMMIT_BYTES verified all 37 package/runtime-input files and canonical source-history availability against that exact commit; staged and committed diff whitespace checks passed. Hosted CI subsequently passed on this exact head: run39227131409, replay job117977112828, merge effcc1b against canonical20ecf2891d80b5acf18d980b2164fb44ea7faa90, CPython3.12.15. The registered replay and committed-input immutability checks passed. The last coordinated main/priority/selected/overlap checkpoint at08:58:49UTC was unchanged; overlap coverage remains partial.