[Lyy-iiis]

PR2843's T124 allocation leaves one rank-9 cold-start residue. This PR interleaves two copies and moves one width-7 helper into the other copy's initial rank-16 bin. The complete paid cost changes by H₂ + H₁₆ − 2H₉ < 0, with unchanged native words and all other calls, bridges, owners and costs retained.

The conditional modern outer target is κ = 0.001718658508304930004819 − 10⁻⁴⁶ − 10⁻⁷⁰, exactly 2.30463210106193e−10 above PR2843's full pinned headline. This is an actual allocation improvement, not a decimal refinement. Only research/t248-dirtentry-residue/ is added; the selected frontier is unchanged.

Evidence:

• All 910,656 literal roles, 2,232 calls, 248 source-event projections and complete rank mass verified; five invalid-allocation controls rejected.
• Separately implemented owner and rational-moment auditors import no inherited engines. They confirm a root in (0.0017186585083049300048199, 0.0017186585083049300048200) and agree under optimized execution.
• Four native chains and all 16 retained parent stages freshly replayed. Complete candidate admission explicitly reuses the authenticated cache and binds PR2839's tensor-block norm contract to the actual PR2833 words.
• Clean complete release passed: 119 admission, 12 BIT, 25 actual controller, eight independent-owner and six geometry rejection cases; 63 strict outer inequalities, 17 outer rejections and 19 exact replays. The inherited summary's controller=15 field is preserved and explicitly distinguished from the authoritative 25-entry array. Independent final source/receipt audit passed 29 checks.
• Metadata passes; bounded registered commands were run directly under Python 3.13.7. Python 3.12 launcher execution and a fresh Lean kernel audit are not claimed. Exact versions, commands, immutable source hashes and receipts are included.

The first T248 release correctly rejected an empty cache directory left by local exploratory bytecode quarantine. The failed run is retained as failure evidence. Only that untracked empty directory was removed, all immutable source files were reauthenticated, and the complete clean replay used unchanged predicates.

Reproduction and the concise construction proof are in the package README. Full source is materialized from exact PR2833 Git objects instead of duplicating its 51 MB tree. The new convenience wrapper's CLI/source export/preflight paths were tested; the recorded complete run uses its exact constituent commands.

The inherited all-size compiler, source-access, hosting, precision/tape, recovery, BIT and analytic hypotheses remain. This is neither an unconditional multiplication theorem nor maintainer adoption.

Credit Jeremy Kothe / PR2833 for the dirt-entry words; Marius-Paul Dumitrean / PR2843 for T124 and prior work; Dugongue / PR2839 for the tensor-block replica contract; Tekalign and all retained native/owner/admission contributors. New work prepared for Yiyang (@Lyy-iiis) with substantial OpenAI Codex assistance, including parallel separately implemented audits. No independent human review is claimed. Original licenses/notices are preserved.

Canonical coordination: #2612 (comment)

Task scope is explicitly changed from the original helper-inventory card to this fully priced allocation/composition result. The base was refreshed to current main and priority revision 5; live PR overlap and targeted Discussions were checked, with incomplete broad Discussion pagination disclosed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.