Retain the P104 relay floor under all original helper reads

[ywz65]

P104's released unfinished-helper handoff asks whether original G/Q/D source rows outside the root code can evade the monotone relay floor. Read-only access to every original helper at its actual chronological frame still requires 208 new monotone correction-helper lifetimes. The new finite premise is that none of the 72 original helper paths visits any of the 104 target endpoint hyperplanes. A coupled source-row induction proves every actual read supplies only K + U(V) at its current frame, so the predecessor quotient/arrival argument survives.

This extends PR2383 at 41f3d517b521500b114f3491f5f5b68a0550f410, which excludes unfinished original rows outside K. The 208 numerical floor and its original monotone-relay proof belong to Philippe Van Velzen and Hiroki Kumazaki / PR2353, 2704eadf6b5205d4b965e9c6fc6dce28b89783cf; the root-code quotient extension belongs to commdata2338 / PR2383. Those results are independently reconstructed as premises, not relabelled as new findings.

Only research/p104-live-helper-reads/ is added. It contains the written arbitrary-length proof, exact 280-input event replay, second source-decomposition certificate, all 72 helper paths and 104 target/minor witnesses, semantic controls, export tests and complete provenance/replay metadata. No pending predecessor executable is imported. Example: at cut 593, G25 = initial_G25 - 2 X1 is outside K but supported at its real frame span(8,64,416), properly inside Y2's endpoint.

Validated on CPython 3.12, standard library:

python3 -I -B research/p104-live-helper-reads/verify.py
python3 -I -O -B research/p104-live-helper-reads/verify.py
python3 scripts/research_package.py --package research/p104-live-helper-reads
python3 -B scripts/research_baseline.py
git diff --cached --check

The registered run checks all 4,058 original events and 280 inputs, 2,458 source-decomposition states, 434 helper/register-frame pairs, zero original-helper endpoint contacts, eight private lifetimes, all dirty restorations and eight semantic controls. Its four export tests check read-only Git-free replay, changed/missing inputs, receipt drift and duplicate JSON keys. Normal and optimized direct verification passed. A separate AI-assisted reviewer read the mathematical argument and source binding and reran the optimized verifier on native CPython 3.12.7 with no material finding; no independent human or proof-assistant review is claimed.

Limits: actual-frame, read-only access to the unchanged original word. Writable original helpers, new descents, changed paths, target feedback and arbitrary-frame outside-code oracles are excluded. Lifetimes are not simultaneous registers or excess paid ranks. No correction construction, native phase audit, global owners, favorable complete moment, precision/reserve, tape, all-size supplier or new kappa is claimed. The selected frontier and frozen sources are untouched.

Canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4, was fetched for this round and again at 11:38:38 UTC; both predecessor heads remained unchanged. The advance claim names the exact handoff. Selection nonce 691d45d0f00693aa125dd9c581fbf5c9; scope change and stop condition are explicit in TASK.json. Fresh overlap covered 660 open PRs at 11:21 UTC and 1,186 discussion records at 11:25 UTC, complete pagination/no warnings; this excludes arbitrary diffs and unpublished work.

Credit also Rohan Arun / PR1972 for the frozen word, Chafik Boukhalfa / PR1991 for its retained native realization, Adam Gall / PR1940 for the schedule, Josie Elliston / PR1071 for the fixture, Ali Romman / PR1125 and the retained geometry/compiler lineage. Prepared for ywz65 with substantial OpenAI Codex assistance in derivation, implementation, checking and writing. Apache-2.0. Open next questions explicitly change an original path or write before a private descent; repeating current helper reads is no longer the proposed escape.