Match the staged-width conditional κ frontier with a smaller T360 allocation

[zxv]

The staged-width suppliers admit a better complete allocation after exchanging the adjacent final T37 call and initial T8 call. Recomputed source windows and an exact S8 allocation give T360, 1,321,920 literal helper roles and 323,069 physical families. The full paid profile crosses between 0.001730096363528623534230 and 0.001730096363528623534231; the intended conditional released endpoint is the lower endpoint minus 10^-46+10^-70. Exact normalized pricing beats #3626 by about 2.44988228437e-8, as well as #3624, and ties the newer concurrent #3633 endpoint. Relative to that T488 construction, this uses 114,688 fewer physical families and 470,016 fewer helper roles. The independent 2,551-pattern dual replay closes further packing improvements on the same fixed-model 24-place release grid. No priority claim or minimal replica count is claimed.

The package retains the exact predecessor source bytes and credit, an integer certificate, deterministic literal builder, separate audit using actual native frames, full replay entry and authenticated execution evidence. It also records the two other assigned investigations as appendices: one constant-bridge dictionary rewrite is excluded, and unchanged native five-plane consumers expose deferred translation debt. Both broader redesigned mechanisms remain inconclusive.

Validation and independent Astra review status are recorded in research/three-kappa-leads/VALIDATION.json, STATUS.json and reviews/REVIEW-DECISIONS.json. The completed full release used both changed native suppliers freshly checked in the same session, authenticates retained parent evidence, and checks the complete physical/precision/outer release. The complete cold one-command CLI is separate; do not infer it ran from stored receipt verification. The first environment attempt failed because isolated subprocesses could not import user-site NumPy; a dedicated Python3.12 environment resolves that dependency boundary.

All inherited all-size/compiler, common-label/mutable-source, uniform-cost, hosting, bootstrap-time, precision/recovery/tape, analytic and bit-interface premises remain. This is a conditional finite construction, not an unconditional multiplication theorem or a selected-frontier update. Credit Ross Richardson/#3605, hexapode/#3494/#3624/#3626 and all retained source authors. Substantial OpenAI Codex assistance and scoped same-session Astra AI review are disclosed; no human or formal-kernel approval is claimed.

Submission status: complete release passed; all three required fresh Astra medium reviews conclude YES for their stated scopes. Bounded verification also passes in a committed Git-free export. Strict source-history ancestry validation remains unrun in a full-history clone because the shared checkout is shallow.

Reproduction: python3 -B scripts/research_package.py --package research/three-kappa-leads (Python 3.12). Complete fresh replay: python -B research/three-kappa-leads/full_replay.py --work /fresh/release with requirements-full.txt. The latter cold command is supplied but has not been executed end-to-end; the completed release used the authenticated same-session reuse documented above.