Verify conditional κ = 0.001602080856660739 with bundled macOS/Linux cold replay - #586

This bundles and independently replays κ = 0.001602080856660739 on macOS and Linux. PR584 appeared during verification with the same construction and rate; this PR matches that frontier and adds a self-contained source archive and three-provider portability controls.

PR582's 66/5/1 dispatch and literal owner map are retained. PR583's six frame coalescences are applied to both kernel providers, replacing 12 rank-1 payments with six rank-2 payments per call. Scalars, entrances, widths and physical owners are unchanged. The conservative rate is 0.00160208085385131.

The frozen package bundles 259 original Git-blob-pinned vendor files, normalizes descriptive word metadata, and rebuilds the full PR558 baseline and all provider construction, native and phase chains. It then reconstructs the physical owner word and fresh invoice and runs both exact rational models. Verification uses only the Python standard library and has no source override, proof reuse or discovery mode.

Validation:

• Full cold replay: macOS/Python 3.14.8, 41/41 stages, 261.03 s; Linux/Python 3.12.3, 41/41, 303.31 s.
• All 21 provider proof-content identities, every numerical/semantic invoice field, and every result field except runtime agree.
• Both runs passed 24 new frame controls, 33 local controls, 31 owner controls, 147 assembly controls, three timing pairs, three actual recompressions, 27 proof-content mutations and 27 reference/metadata mutations, plus inherited construction controls.
• The unchanged repository baseline retains explicitly identified PR572 18-group evidence on the same base; it is not claimed as a fresh aggregate baseline run.
• GitHub CI is pending and monitored separately.

Run:

python3 -B research/promoted-kernel-coalescing/verify.py --work /tmp/promoted-kernel-cold --jobs 3

Evidence and the pinned PR584 source comparison are in research/promoted-kernel-coalescing-validation/. The owner bytes agree with PR584; its package was source-compared, not independently replayed here. No packing optimality is claimed: our bounded warm-started allocation search returned the existing 66/5/1 profile.

Credit: Jeremy Kothe's PR582/576/568 construction and dispatch, Dominik Scholz's PR571/569 prefix and promotions, Jonathan Capalbo's PR575/567 prefix and content identities, our PR583 frame coalescences and portable controls, and the retained upstream contributors and OpenAI #109 framework. Original notices and assistance disclosures are preserved; this derivative used substantial OpenAI Codex assistance for research, implementation, verification and writing.

The inherited all-size compiler, mutable-source/common-label realization, whole-call hosting, precision/tape/recovery, analytic transfer and bit H1–H5 assumptions remain. This is conditional finite evidence; κ = 1 remains unproved.