Host QQ16 correction temporaries in existing output banks

[Clear1nsight]

PR #3985's complete recovery word uses 112 extra full-cover temporary banks although 104 eventual output slots have no live output yet. This contribution derives those exact dependencies and gives a literal allocation inside the existing 224 owned output banks. All 224 original outputs still finish in F17, and all 224 borrowed current X/Y arrays remain unchanged.

Only Z_T (eight ports) is read before the future logger. Defer Z_U and Z_A; place the 104 query copies in those slots. Once all eight private queries have captured their inputs, eight query slots can become response slots. Restore Z_A while W_A is still unchanged, consume the responses, then define Z_U and resume the unchanged suffix. The checker records 104 exact query/response/output lifetime intervals and rejects early reuse, live-selected-output overwrite, initialization/frame/alias/return errors.

The complete local bill keeps 224 H1, 448 wrappers, 352 cover maps, 1,920 unit ADDs and 896 quarter ADDs. Additional full-cover temporary banks fall 112→0. Removing the 112 obsolete final temporary clears reduces CLEARs 536→424 and scalar allowance 27,184→26,960. All first-use clears and original output definitions remain paid; arbitrary owned dirt is not assumed zero. Peak owner-private scratch stays 16.

This is concrete reuse under the pinned decoder's existing 224-owned-output contract, not a claim that these banks are free in the original gauge44 physical caller. The first-B1 cut is source-bound; six-tape placement, movement/control, output consumption, precision, family coverage and recursive supplier admission remain open. No kappa or optimality claim. The independently reviewed paired streaming construction uses a different access schedule and achieves 416 CLEARs / 26,944 scalar allowance with nine deferred outputs and zero future-private registers, versus this construction's 424 / 26,960, 104 deferred outputs and eight future-private registers. It is not a runtime dependency; this owner-private schedule is not claimed Pareto optimal.

Source: PR #3985 head 400df7f, retaining only its exact WORD data and regenerating the full word from canonical PR #2707. All transitive canonical inputs are SHA/Git-blob authenticated before executable reuse. Complete new-word replay checks both independent current arrays against the routed inverse and original arrays through F0/B1, all 73 frame coordinates, arbitrary 224 output dirt fields, borrowed return and 114,688 signed native columns with all 224 occurrences bound.

Validation: normal and optimized complete replay, six read-only/export/source/word/typed-receipt/overwrite boundary tests, metadata and input pins passed. Independent coordinating-agent code/proof review and frozen replay passed; exact hashes/scope are recorded. Hosted registered replay has passed; details follow. Final publication head and committed-byte/source-history checks follow below. Package: research/qq16-output-pebbling; reproduce with python3 -B scripts/research_package.py --package research/qq16-output-pebbling.

Canonical main 20ecf28 remains unchanged, priority 6. User-assigned output-bank/register-allocation follow-up; shared claim #3985 (comment). Partial overlap coverage; no novelty guarantee.

Credit #3985; ywz65 #2707/native methods; Arham Tahir #2681 ordered inverse; Ali Romman's QQ16 lineage; commdata2338 #2651 and Discussion 2617 for prior early-record/output-bank coalescing; Nick Petalas #2722 access tariffs and #3339 consumer ledger; retained geometry/native/supplier contributors. This concerns the newer zero-retention query/response lifetimes, not a second claim for #2651's early-record saving. Substantial OpenAI Codex assistance; independent AI review is distinguished from human/formal verification.

Publication head: fb1c0da01c02f2a5fbc557dc16c1ebb90dac4e76. Exact committed bytes (36 package/input files), canonical source-history ancestry, official Git-free export and export-integrity check, and diff whitespace check all passed. The full registered launcher has not been redundantly rerun locally; direct registered commands passed, and hosted registered replay has now passed.

Hosted validation completed successfully for publication head fb1c0da01c02f2a5fbc557dc16c1ebb90dac4e76: Registered research packages run 39226247360. Completed replay-job logs bind merge cc76758de4a12b5c9e66d64712ee499dc6f02d7e to that exact head and canonical base 20ecf2891d80b5acf18d980b2164fb44ea7faa90, using CPython 3.12.15. The complete mathematical replay passed, all six boundary tests passed in 215.245 seconds, and the full registered launcher reported PASS_REGISTERED_RESEARCH_PACKAGE research/qq16-output-pebbling outcome=completed. The committed-input check also passed with git diff --exit-code. This discharges the previously pending hosted registered-launcher check; mathematical and physical scope qualifications above remain unchanged.

The parallel hosted-streaming construction is now published separately as PR #4015, head 0fff81d8997cef646538bcd4f60d7d455cf93b8e. Its different access schedule attains the stronger local scalar/private counts already compared above; physical access tariffs remain to be compared.