Research construction: remove known-zero JP8 clears (9,228 primitives per orientation; no new kappa)

[faw01]

Finite complex-supplier circuit improvement; no new κ. The historical #626 JP8 word repeatedly clears coefficient work that is already known to be zero. A conservative specialization of Sagar Thakrar's #696 lowering removes 4,614 such clears per real orientation, saving 9,228 real-component primitives per complex orientation beyond #696. Every original addition still ends with zero work.

Complete forward/backward scalar bills fall from #696's 200,380 / 200,396 to 191,152 / 191,168. The complete local invoice, including the unchanged adapter reserve, falls by 9,228 to 11,334,342,396,608. All source accesses, helpers, frame events, wrappers and recursive children retain their original charges. These are instruction counts, not measured runtime improvements.

The state analysis forgets its knowledge at every non-scalar event and frame change. It changes no scalar macro or clearing boundary. This differs from #722/#739/#755/#787/#793/#811/#830 coefficient sharing and batching; #755's pinned source was inspected and explicitly excludes unit coefficients from sharing. The final refresh also found #837's distinct bidirectional Horner construction within coefficient macros. Their savings are not added here.

Only research/jp8-clean-work/ is added: executable lowering, run-level proof, source pins, deterministic receipt, independent backward census, dyadic/Fraction checkers, negative controls, package tests and local sprint claim.

Reproduce with standard-library Python 3.11+:

python3 -I -B research/jp8-clean-work/verify.py

Expected: PASS_JP8_CLEAN_WORK orientations=2 columns=716+9 extra_saved=9228,9228 controls=15.

Validation passed on Python 3.11.17 and 3.15.0: both complete scalar orientations on 716 native plus nine arbitrary private columns; 574 coefficient identities and 13,728 intermediate-state comparisons; 15 rejection controls; four package tests including optimized minimal relocation, gzip recompression, dependency corruption, forged invoice and false-clean-state rejection. make selected-record-check and staged whitespace pass. Full repository/native/global replay, remote Linux CI and Lean were not run; geometric admission and the all-size/compiler/hosting/precision/tape/analytic assumptions remain inherited. Global artifact bindings are not regenerated.

OS-random track seed: 33 (complex supplier). Exact base: 7ee02b7c96d85f22b38d28bd7ed49bb8c1aeb261. Upstream subsequently advanced to #799 at f567884aeb8851b6e794af1184119a3bfbe1e631; its review was checked and all nine consumed historical inputs remain identical. The selected frontier and upstream/ are untouched.

Credit: Sagar Thakrar (#696), icekylinx (scalar/coefficient lineage), Jeremy Kothe (#622/JP8), Chafik Boukhalfa (#626), DreamingOfClouds, Rohan Arun and the retained predecessors. Contribution by faw01, with substantial OpenAI Codex assistance in research, implementation, proof, verification and writing. Apache-2.0; original notices remain intact.