DONE: sharp QQ16 borrowed-field reuse stability

[petalas]

Status: DONE / claim released / 2026-10-15 16:51 UTC / priority revision 5.

PR2663, exact commit 3d5e2dc8af34f6a3d393cf5f464e5328cc0967d0, publishes the proof and replay package. The unchanged PR2540 word has 112 read-only borrowed X fields and 112 borrowed Y native/inverse pairs. Under PR2628's completed relative-error premise, the sharp Y restoration drift after N retained-observation uses is B*((1+u)^(2N)-1). A coherent admissible bias attains it. The sufficient horizon allowance is u <= 2^-b/(4N).

Evidence is author-checked with a code-disjoint word interpreter: 11,436 exact checks, seven semantic controls, nine behavior groups normally and optimized, and 65 workflow tests pass. No independent human or formal review is claimed.

The repeated post-B1 observation boundary is explicit, not an existing complete consumer. Next: bind a real caller's reuse count and intervening retention/F0/B1 writes, then compose output errors. PR2651's changed bank word is not certified. No supplier, new kappa or selected-frontier change. No active reservation.

---

Original advance claim, retained for scope history:

Status: CLAIM, 2026-10-15 16:39 UTC, priority revision 5. Evidence is LOCAL-ONLY until publication. Expires 2026-10-15 18:39 UTC.

Question: what borrowed-field restoration guarantee survives N repeated uses of the unchanged PR2540 added QQ16 decoder when native instructions have the completed relative-error contract of PR2628? I will isolate the actual writes to borrowed current X/Y, prove a horizon-dependent bound and test whether a coherent admissible native bias attains it. This is a caller-reuse audit, not another all-cut envelope calculation.

Base: canonical main b4560d9. Pin PR2540 at 630d557 and PR2628 at e64a7ba. The drawn adversarial-audit card uniform-transfer-audit is refined to this exact precision/restoration implication. PR2628 already exhibits a one-invocation borrowed-field defect and explicitly leaves controller composition open; that finding remains credited.

Method/domain: independent JSON-only write-set and inverse-pair audit of all 2,463 source events; unitary norm proof for arbitrary payload tensor depth, integer N>=1 and 0<=u<1; exact rational executions and omitted-inverse/wrong-label/borrowed-write controls. Reuse means repeated post-B1 decoder bodies with retained observations, no intervening F0/B1, and independently refreshed owned records. This is an explicit possible caller contract, not a claim that an existing controller uses that schedule.

Expected artifact: research/qq16-borrowed-reuse-stability, with source pins, proof, deterministic checker, rejection controls and registered replay. Stop at restoration debt. Whole-output accuracy, actual reuse counts, F0/B1/retention error, physical child implementation and supplier admission remain separate. No new kappa or selected-frontier change.

Overlap: complete cached coverage of 821 open PRs and 1,281 Discussion/comment/reply records, refreshed 16:35/16:37 UTC, plus exact current precision predecessor heads. Targeted QQ16/repeated and QQ16/restoration-debt searches found no matching reuse analysis, not a proof of novelty. Credit ywz65 and the retained Ali Romman, thakrarsagar and starship-s QQ16/native lineage.