probe #625 main re-run: W4 kill reproduced; D4 bond via imported #653 census

[LightChainr]

Re-run of #625 against main (per the assignment; note #629's probe branch was already merged to main on 2026-09-07). Companion to #622's W4. Closes nothing. Does not enter docs/STATUS.md. Draft, left open, not for merge.

What this is

• Verification run on a clean Mac clone of main @ 8b5bc840. Game B site L=3 (2^9, Mac-OK) and the D2 physical tables were re-run locally — reproduced bit-identically (the unchanged physicalZ.json / gameAB.json are not modified in this PR).
• D4 without 2^18 enumeration. The bond census is imported from PR Repair #632 bond laboratory; wrapping-type census L=3,4 (#640) #653 (repair/632-bond-lab-and-wrapping-l3l4 @ 7109b174, results/probe-invariant-shape/census-exact.json; main's copy is the pre-[DeepSeek / P1 repair] #628 exact laboratory: bond duality, CDF normalization, and probability simplex #632-repair one with dual_fail=118133). scripts/probe/bond_L3_Z_imported.py derives the D4 verdict from the 3-count census arithmetically; no 2^18 enumeration was performed on this machine.
• Tests: tests/test_probe_mhalf_vs_shape.py — 5 fast OK, 1 slow skip (Python 3.13.7).

Verdict (D5)

W4 at L=3,4:  KILLED
evidence:     Game B L=3 exact: on the m-pinned line F_lambda(p)=1/2
              (m = E_lambda[r_b]-1 = 0 EXACTLY for every lambda),
              Z(0.5; anchors 0.1/0.9) sweeps 0.73817 (lambda=1/16)
              -> 0.54242 (1) -> 0.32192 (16); interior-u spread =
              0.416 (a1/a9), 0.271 (a2/a8). L=4 spot (1, 5/2, 2/5):
              spread 0.178. Both ~4e7 x the 1e-8 tolerance.
Z_3 vs Z_4:   ||Z_3-Z_4||_inf on {0.1..0.9} = 0.0161 (anchors 0.2/0.8);
              at p_L^H |delta| = 0.0066, at p=1/2 |delta| = 0.0837.
bond vs site: INCOMPARABLE — bond X-law at p=1/2 is 3-atom
              (0.28786, 0.42429, 0.28786); Z_bond undefined by the
              probe's own rule; jumps recorded. Census imported from
              #653 (dual_fail=0, M(1/2)=0 exact, P(r=0)=P(r=2) exact —
              duality-oddness checked from counts, not assumed).

Files (5 changed: 3 added, 2 modified; no deletions)

• notes/probe-Mhalf-vs-shape-main-rerun-20260908.md (added — D5 + provenance)
• scripts/probe/bond_L3_Z_imported.py (added — D4 from imported census)
• results/probe-Mhalf-vs-shape/bond_census_imported_653.json (added — imported Repair #632 bond laboratory; wrapping-type census L=3,4 (#640) #653 census with provenance)
• results/probe-Mhalf-vs-shape/bond_L3.json (modified — regenerated in imported mode, v2 schema)
• results/probe-Mhalf-vs-shape/latest.json (modified — v2 index for this variant)

Standing

Does not close #606, #619, #622, #608, #625. No production, no N=725, no arXiv, no ninth mechanism, no L^-θ fit. Census numbers are inputs, not deliverables.

[LightChainr]

Verification comment, not a merge request. Ticket stays open.

Matches the 2026-09-08 assignment, not the original issue-body dest branch:

• draft against main
• Game B site L=3 (2^9) Mac-OK
• D4 bond imported from PR Repair #632 bond laboratory; wrapping-type census L=3,4 (#640) #653; no 2^18 on this Mac
• no STATUS, closes nothing

W4 kill is the same qualitative verdict as the already-merged #629 probe: on the m-pinned Game B line, Z still sweeps (spread 0.416 vs 1e-8). Bond vs site stays INCOMPARABLE because the imported 3-atom X-law makes Z_bond undefined by the probe's own rule. Duality-oddness is read from dual_fail=0 counts, not assumed — that is the #653 repair this re-run was for.

Minor: the notes §Deliverables line says "0 modified"; the PR actually modifies results/probe-Mhalf-vs-shape/bond_L3.json and latest.json. Provenance is in the new imported JSON. Not a science issue.

Do not close #625 / #622 / #608.