Research: reproduce T185 and certify a precision-only kappa refinement

research/t185-precision-replay/.gitattributes

@@ -0,0 +1 @@
+* -text

research/t185-precision-replay/CLAIM.json

@@ -0,0 +1,23 @@
+{
+  "status": "PASS_COMPLETE_PRECISION_ONLY_RELEASE",
+  "baseline_head": "34328b1d5e5bebac62d5cabec993dfbfd7b2123f",
+  "paid_profile": {
+    "1": 4029300,
+    "2": 1068930,
+    "3": 458874,
+    "4": 213120,
+    "48": 66896,
+    "5": 65490,
+    "8": 13320
+  },
+  "kappa": "172010080541688592794400707560633553119999999999999999999998999999999999999999999999/100000000000000000000000000000000000000000000000000000000000000000000000000000000000000",
+  "kappa_decimal": "0.00172010080541688592794400707560633553119999999999999999999998999999999999999999999999",
+  "gain": "707560633553120000009999999999999999000000009999999999999999/100000000000000000000000000000000000000000000000000000000000000000000000000000000000000",
+  "gain_decimal": "7.07560633553120000009999999999999999000000009999999999999999E-27",
+  "complex_lower": "1075063003385553704965004422253959707/625000000000000000000000000000000000000",
+  "conservative_lower": "1075063001545801545851560323691846999/625000000000000000000000000000000000000",
+  "release_receipt_sha256": "6dd5dad8d9899603684e46ee9cd1b915d822aeda7148de4126c2c361b80f3a6b",
+  "native_receipt_sha256": "0bceef521598a678b7687502a84a86b16d3faddf2702939e771eba9a4679baed",
+  "admission_driver_sha256": "8aa91da8b39f212ef786f7788682dce45d6217f399f02a063f62763d096cbfb7",
+  "scope": "Tiny explicit rational refinement of unchanged T185 construction. Same optimal root; no new circuit, runtime gain or full theorem proof. Inherited assumptions retained."
+}

research/t185-precision-replay/DEPENDENCIES.json

@@ -0,0 +1,22 @@
+{
+  "baseline_pr": 2863,
+  "baseline_head": "34328b1d5e5bebac62d5cabec993dfbfd7b2123f",
+  "package_prefix": "research/shared-kernel-packing",
+  "original_manifest_sha256": "e1276c24fba1eb4c83ae040da63addad4a73f83120775ab40a79136c6080d59b",
+  "subset_only": true,
+  "subset_sha256": {
+    "CLAIM.json": "eebaf5fc6d4721ba99418bdd0bbcbb933c5bf4f351025065f39a11120680f0f9",
+    "NOTICE": "1493711f66e5cf8d53dd7d4f67edfe2eab8d17dd232b86e08c2d35cc9398cf39",
+    "CREDITS.md": "5144a2d0774dbbb41cded421f00a05cf7c106a9278bbb2869c8dbe31ed71bbe5",
+    "LICENSE": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4",
+    "MANIFEST.json": "e1276c24fba1eb4c83ae040da63addad4a73f83120775ab40a79136c6080d59b",
+    "review/PROOF.md": "71a45d195eb9ea46e0b0027d7d6f750aef109b1e853b9ff6b90ab864039bae6f",
+    "tools/admit2857.py": "1d4b8b9df01e42f5f2450b5837132861c5aa8cd3a011e6a6fa35e53463417346",
+    "data/PROFILE.json": "ef36eda129e3463ce7fb76dedd953a16035924c201e38ee46dc79d56d4229672",
+    "data/PROVIDERS.json": "b1e31d1b3af2c8db9b8dd8f1337fee7fc2570ee58c1d541c43963c640efa1aa2"
+  },
+  "full_source_head": "e9307b48d7d1b690939935ecef0f8e524e6d92f4",
+  "full_source_prefix": "research/gmx-t74-borrow43-handoff",
+  "full_source_manifest_sha256": "c8ec4d1b7be08d2ad30730a07392ad9cb6d0222f2c35bef43c9a80a39b16e41f",
+  "new_native_overlay_manifest_sha256": "54dc42f25c1d8b85adcf593c4bf6c0f83f7351388af857812790b8a0b4c3d869"
+}

research/t185-precision-replay/EXPERIMENTS.md

@@ -0,0 +1,40 @@
+# Experiment ledger
+
+Deterministic runs, no random seed and no paid compute. This continuation began
+2026-10-16 05:18 UTC with a roughly 60-minute budget. Four delegated agents
+handled frontier/reproduction, independent dual arithmetic, structural changes,
+and outer precision. Native chain concurrency never exceeded two.
+
+At the pinned endpoint, approximate moment contributions are rank1:33.54%,
+rank48:26.56%, rank2:17.78%, rank3:11.44%, rank4:7.08%, rank5:2.72%,
+and rank8:0.88% (floating screening only; exact acceptance uses rational
+intervals). Rank-one native work and rank-48 remainders dominate. The attained
+fixed-inventory dual removes repacking as a degree of freedom; changed native
+work/interfaces or a sharper explicit endpoint remain the tested possibilities.
+
+| Hypothesis | Decision and evidence |
+| --- | --- |
+| Repacking the same inventory can lower the moment | Rejected. Independent enumeration checks all 1,163 patterns; the attained dual closes this fixed model. `audit/check_optimum.py`, `evidence/INDEPENDENT-OPTIMUM.json`. |
+| Add a fifth independent cancellation direction by adjoining one nonzero column | Rejected for the stated domain: 216 columns/provider, none lie in the required span. Redistribution of existing kernel directions by basis changes is not excluded. `experiments/RAW-BASIS.json`. Joint rank was already in PR2857. |
+| Lift one zero-compensation helper entry frame | Rejected: 47 cases/provider, all94 fail fixed centre/warp/endpoint constraints. `experiments/ENTRY-LIFT.json`. No complete-cost candidate reached native admission. |
+| Refine rational endpoints while retaining the whole construction | Exact native/conservative signs and all outer rows pass as a screen. The 40-digit candidate is frozen; complete release passed. `evidence/INDEPENDENT-PRECISION.json`. |
+| Published baseline reproduces unchanged | Failed: exactly two manifest-listed files absent. Separate exact recovery then exposed gzip OS-header inequality. No original check or manifest changed. `evidence/run-logs/` and `notes/PORTABILITY.md`. |
+| Portable finite replay and cold full baseline agree | Passed. Decoded regenerated payload equality plus immutable container hashes; five finite mathematical stages, mutation controls, six fresh native chains, retained-parent admission and strict complete baseline release. `evidence/BASELINE-RESULT.json`. |
+
+Exact commands, working directories, exit codes, timings and log hashes are in
+`evidence/run-logs/*.json`, the native/admission stage receipts and the structural
+`experiments/RUN.json`. Historical absolute paths identify the actual local
+runs; README.md gives portable reproduction commands. Early agent notes remain
+labelled as screens, even when later complete evidence is available.
+
+Only one bounded structural pivot was undertaken. A broader changed-program
+claim by Mason Hale (ROUND-4, posted05:22UTC) was already active in
+[Discussion2612](https://github.com/CrocSwap/integer-mult-bounds/discussions/2612)
+at the final census, so these negative tests carry no exclusive priority claim.
+More packing seeds or repetition of the same47-case entry subset are not
+justified. Other helpers and entry lines remain unsearched.
+
+The combined `reproduce_all.py` orchestrator has source-export, syntax and help
+preflight, but has not been run as a second redundant cold job. Its constituent
+full stages are run separately in this session. That distinction is retained
+in ENVIRONMENT.json and the final handoff.

research/t185-precision-replay/HANDOFF.md

@@ -0,0 +1,49 @@
+# Completed research handoff
+
+Baseline: PR2863 at34328b1d5e5bebac62d5cabec993dfbfd7b2123f.
+Original source: PR2833 ate9307b48d7d1b690939935ecef0f8e524e6d92f4.
+Exact manifests and source subset pins are in DEPENDENCIES.json.
+Research base:1cdd4bf6507eeb2d045335e5f6216152ac28aaad.
+Prepublication main:4b82082f901945f8d0314f143576c7ad31e021fc;
+relevant guidance, selected record and review paths are unchanged.
+
+Full cold baseline passed in1364.56seconds. The separate frozen40-digit
+candidate complete admission passed in926.09seconds, including562.79seconds
+of strict release. Exact κ and gain are in CLAIM.json; this is only a tiny
+explicit rational refinement of the same T185 construction and same root.
+The construction and parameter-refinement method remain predecessor work.
+No new circuit, measured speedup, full theorem proof or maintainer acceptance.
+
+The unchanged published fast wrapper failed two missing files and, after exact
+separate recovery, a gzip-header comparison. The portable finite adapter passes
+with exact decoded regenerated payloads and immutable input hashes. Original
+checks/manifests were not weakened. Full admissions pass separately.
+Both requested main gates failed on the clean base; original logs remain.
+
+Four agents produced frontier/reproduction, exact dual, structural search and
+outer-precision evidence. The1,163-pattern fixed-model optimum is confirmed.
+The216-column/provider screen excludes a fifth independent cancellation
+relation from adjoining one tested column; it does not exclude redistributing
+existing directions. All94 selected entry/frame cases failed, but other helper
+choices and entry lines remain unsearched. Do not repeat these same domains.
+See EXPERIMENTS.md and experiments/README.md for commands, controls and scope.
+
+The native/source/cache and generated full supplier inventories were checked
+again after release. The bounded registered verifier also passes independent
+rational/dual/driver audits, actual supplier manifest bindings,63 outer rows,
+17 outer controls and two receipt/source mutation controls. Full command logs,
+exit codes, runtime and hashes are under evidence/. README.md gives fresh-run
+commands. The combined cold orchestrator was syntax/help/export-preflighted;
+its constituent complete stages were executed separately, not as an additional
+redundant combined cold run. This distinction is intentional and disclosed.
+
+Next work should first price an explicitly changed native/consumer interface,
+or investigate a relation using at least two outside helpers. Coordinate with
+Mason Hale's Discussion2612 structural investigation. Further decimal digits
+provide negligible gains. A separate valuable direction is proving one exact
+inherited all-size/compiler or precision/tape implication; finite tests alone
+cannot settle it. All inherited conditional assumptions remain.
+
+Research and heavy verification ended within the approximately60-minute run;
+subsequent work is packaging/publication. Publication must use digitalis-p's
+fork and must not merge. The earlier PR2866 is distinct and superseded.