[Arham-Tahir64]

The literal QQ16 F0 common-chart scalar projection loses a unit data input under actual binary64 round-to-nearest arithmetic, while every dirty helper restores exactly. With Y0=1, L224=2^65 and all other inputs zero, exactly two ADDs round; Y0 becomes zero and four other Y outputs become -1/2. Exact arithmetic and wider-precision controls recover the original state.

This closes the concrete scalar-rounding gap explicitly left outside #2995 and #3017. It does not establish that component rounding in the actual native producer realizes these scalar faults. The native maps, B1, complete producer grid/range/cost contract, F17 consumer and supplier admission remain open. No new kappa or selected-frontier change is claimed.

Adds only research/qq16-f0-rounded-scalar/: a written finite witness, pinned certificate and task provenance, complete numerical receipts, an exact nearest-even interpreter, and behavioral tests. Four finite precisions (8/24/53/113) reproduce the witness. Actual separate binary64 multiplication/ADD agrees with the favorable completed-ADD rounding model at all 21,025 scalar instructions. All 480 borrowed helpers restore and eight private outputs are zero. The omitted native events are explicitly inventoried.

Validation on native macOS arm64:

• Registered CPython 3.12.14 replay passes, including five behavioral groups and optimized Git-free export.
• An independently enumerated finite binary system checks 7,740 rounding decisions; signed ties, scaled/sign-flipped inputs, wider/clean controls, real-word gate mutations and source/receipt corruption pass.
• Optimized CPython 3.14.7 behavioral replay, package metadata/source pins, selected baseline bindings and committed-byte preflight pass.
• Ruff is unavailable locally. Full native/selected cold replay and hosted CI success are not claimed.

Reproduce:

python3.12 -Werror -I -B scripts/research_package.py --package research/qq16-f0-rounded-scalar

Scientific/submission base: canonical main d8e0f83f00191df62174b641e9fff055c357f641, priority revision 6. Final completely paginated overlap refresh covers 277 open PRs, 2,440 closed PR bodies and 1,766 discussion records; 13 textual leads were inspected. No competing F0 concrete-rounding result was identified; arbitrary diffs and unpublished work are outside coverage. Canonical investigation. Family decision: narrow and stop at this witness.

Credit Nick Petalas/#2965 for the producer-accuracy handoff, Arham Tahir/#2995 for the earlier scalar relative-fault audit, Tekalign/#3017 for its native binding, Ali Romman/QQ16 and the retained native/scalar lineage. New work by Arham Tahir with substantial OpenAI Codex assistance, Apache-2.0; no independent human/formal-kernel review claimed.