Research audit: PR626 owner boundaries and bridge-omission blind spot

[tekuuu]

This scoped audit checks the selected PR626 ownership table independently and demonstrates a precise limitation of timing-only verification. All 10,368 window/call decisions and 35,424 literal helper roles pass. 2,736 roles legally use their idle deadline's block, so replacing strict event-position inequalities with strict block-time bounds would reject valid selected roles.

Deleting any one of the 72 literal bridges leaves the borrowing windows unchanged, but the corresponding normalized scalar controller is wrong. The package supplies exact input/output counterexamples for all nine source bridge types, an invertible-product argument, and a separate eight-vector implementation. For example, deleting source bridge 3 maps (1,0,0,0,0,0,0,0) to (1,0,0,1,0,0,0,0).

Status: research audit / independent finite verification; no new κ. This is not a defect claimed against production admission: the selected audit also binds the full source controller and its scalar obligations. It explains why that binding cannot be replaced by a matching timing table.

Reproduce using only the Python standard library:

python3 -I -O -B research/owner-boundary-audit/verify.py

Expected: PASS_PR626_NEXT_TOUCH_BOUNDARY_AUDIT, matching EXPECTED.json. Validation passed with and without -O, with identical outputs; from a committed-files-only archive in an unrelated working directory; and with make selected-record-check. The 956 rejection controls cover deadline order, both bridge endpoints, window endpoints, whole-family aliasing despite disjoint coordinates, and unknown events. Full supplier/repository-wide replay was not run because this standalone package changes no production or construction files.

Exact base: 66849e1475c930fd7732b1a0067ffee7e1e95f60. Selected source #626: 660de893d27bb7abd1d5b01b1640207528a5c851. At that base κ remains 51110230636731/31250000000000000. SOURCES.json pins all inputs and the actual inherited controller; REVIEW.md explains the proof, source chain and inherited physical/all-size/analytic obligations. upstream/ and the selected frontier are unchanged.

Assignment seed 60, adversarial mathematical audit. No pinned sprint issue was available, so the narrow claim and result are recorded locally. The refreshed queue through #672 contains no duplicate of this comparison. Rohan Arun's #553 already formalizes the historical PR531 windows; this adds the selected 492-helper data, a different next-touch method, and paired timing/scalar omission evidence. Credits to chafreaky (#626), jeremykothe2 (#622/#623), GamingPuzzled (#620), icekylinx and the retained supplier lineage are preserved. Prepared for Tekalign / tekuuu with substantial OpenAI Codex assistance, including a separate agent review; Apache-2.0, not independent human peer review.