[Clear1nsight]

Change and mathematical scope

Retain the completed native Fraction score-dispatch probe as a focused, scoped negative finding. Its fixed B,C,C,B micro passes 33 prefix comparisons and ten functional/source controls, and both paired differences improve. The smaller pair illustrates only 0.187221 wall / 0.187228 CPU seconds when scaled to the frozen score-addition count. This falls below the parent complete candidate's observed ranges of 0.460568 wall / 0.45 CPU seconds, even with zero added validation cost. The prescribed conservative screen stops expansion.

This is an engineering screen for this probe. It establishes no complete saving, universal obstruction, new construction or kappa gain. No full control union, actual audit, complete caller or complete candidate was built. The validated performance contribution in PR #5864 remains frozen and unchanged. Archive retention does not imply scientific approval or admission.

Task and overlap

Task replay-runtime, refined scope PR5864/native-score-dispatch-micro-only; canonical main de5ec2e, priority revision 8. Parent PR #5864 head: 3f15c94. The original fork artifact was 781440d, already released in this seven-field finding. The donor now requests upstream retention.

This publication isolates the 52-file package from its parent implementation. All 50 scientific, probe, evidence, notice and other package files are byte-identical to that original artifact. Only TASK.json's base and SOURCES.json's publication pins change. The complete immutable 163-file parent export remains a vendored input, authenticated at its exact head and SHA256. The new five-input ancestor is dec843d; the publication head is 3ed05e3. Historical source pins and raw receipts remain intact.

Nearest work: PR #5864 supplies the measured native cache/loop parent; PR #5852 and PR #5853 retain earlier Fraction table and zero/unit feasibility findings; PR #5848 supplies the frozen qualifier. This adds the later, explicitly rejected score-dispatch probe. Fresh targeted upstream searches for its original branch and package path found no matching PR; coverage is bounded and unpublished overlap is not excluded. Live canonical main is unchanged at the pin above.

The branch starts from 01014cb, retaining Ali Romman's exact test-only PR #5761 repair, original ba8a9e6. Its authorship is preserved; it is an inherited dependency, not new work here. Relative to that repair, this PR adds only the 52 package files and does not duplicate PR #5864's implementation.

Validation

Historical evidence retains the one cold native B,C,C,B probe, every sample, 33 prefix comparisons, ten controls and the subsequent conservative stop. Counter asymmetry is documented: the direct arm pays per-add dispatch counters while the generic arm reports its aggregate at exit. The illustrative scaling is not a bound on actual saving. No timing permutation or sample omission followed the stop.

Registered reproduction:

python3 -B scripts/research_package.py --package research/fraction-native-score-dispatch-feasibility

Registered replay freshly compiles the probe and checks complete prefix values/types/branches, generic custom/subtype fallback, nonexact increments, direct exceptions without retry, partial update order, callback mutation, invalid packets, binding drift and runtime restoration. Linux CPython 3.12, existing cc and Python development headers are required; there is no install or silent native skip. Registered replay does not repeat timings.

Publication preparation passes metadata/source pins, all 52 committed bytes and modes, and export coverage for all five declared inputs. Fresh finite replay passes normally, under optimization and from a Git-free read-only export, each within one second on CPython 3.12.3 / GCC 13.3. Commands were serial, bounded by 180 seconds / 1 GiB sampled group RSS / 600 seconds initial aggregate / no automatic retry, with no stderr and unchanged source bytes/modes. No full benchmark or historical actual replay was repeated. Full-history source preflight remains unavailable in the shallow checkout; no such pass is claimed. Hosted exact-head CI will be reported separately.

The original source-authentication harness failure is retained: compiling an isolated function altered imported-module call opcodes; complete-module authentication preserves context. The failed draft and receipts remain. The original publication setup failure is retained in the raw evidence; this isolated publication also encountered and resolved a sparse-index staging setup rejection before committing, with no scientific rerun or data change. These are packaging/harness events, not failed mathematical differentials silently repaired.

The parent 964 comparisons / 118 controls and complete qualification were not repeated for this rejected extension. Full-control integration, the planned prior dispatch controls, new Cache/Event lifetime integration, actual audit, complete performance qualification, independent human/proof-kernel acceptance and physical supplier admission remain unclaimed.

Credit

Original Fraction/error arithmetic and independent integer backend: Ali Romman. Preserve Adam Gall, ywz65 / PR #2878, Mason Hale / PR #5811, PR #5831, PR #5826, PR #5819, and all nested P104/native/compiler/precision notices. Tooling lineage includes PR #5839, PR #5816, PR #5818, PR #5832, PR #5848, PR #5852, PR #5853, PR #5864 and Ali Romman's exact PR #5761 repair.

New probe and recorded execution: Chris R / Clear1nsight with substantial OpenAI Codex assistance. GPT-6-astra xhigh performed static nomination/stop review only, with no compiler or numerical work. This publication audit uses a donor-requested agent. Standard dispatch/C API techniques; Apache-2.0 and original notices retained.

Open next questions

None justified for further measured continuation in this round. A separately proposed mechanism needs its own source, ownership, coverage and complete-cost argument before a new fixed feasibility gate. Constructor-elision and descriptor-metadata ideas retain unpriced obligations; this stop does not exclude them. See the package README and TASK.json. This archival publication dispatches no successor experiment. Kappa research, Work 1/2, feeds and automations remain paused.

[Clear1nsight]

FINDING / DONE / RELEASED — the previously fork-only scoped native score-dispatch stop is now retained in focused upstream PR #5875. No additional complete performance or kappa gain is claimed.

1. Research discussion / PR / artifact: PR #5875, following the original released finding on PR #5864. The donor requested upstream retention; this is an archival submission, open and unmerged.
2. Exact commit: publication 3ed05e3; five-input ancestor dec843d; original fork artifact 781440d; parent PR #5864 3f15c94. Canonical main is de5ec2e, priority 8. All 50 nonmetadata package files are byte-identical to the original; only TASK base and SOURCES publication pins change. The full immutable parent export is vendored, so no parent implementation is duplicated. Ali Romman's exact PR #5761 test-only repair 01014cb retains original authorship.
3. What is established: all 33 prefix comparisons and ten native/binding/fallback/error controls pass. The fixed one-trial B,C,C,B evidence and all samples remain unchanged. Smaller-pair illustrative scaling is 0.187221 wall / 0.187228 CPU seconds against parent complete ranges 0.460568 / 0.45 seconds. Even zero added validation cost fails the prescribed conservative screen. STOP before full-control integration, actual audit or complete construction. Fresh publication finite normal, optimized and Git-free read-only replays pass; metadata/pins, 52 committed bytes/modes, export and all five input declarations pass.
4. What remains uncertain: illustrative synthetic scaling is not an actual saving bound or universal obstruction. Counter asymmetry and all original failed drafts/source-authentication/setup receipts remain visible. The full 118-control parent chain, prior dispatch controls, Cache/Event integration, complete caller and complete performance gate were not repeated or built for this rejected extension. Full-history preflight is unavailable in this shallow checkout. No independent human/proof-kernel acceptance or physical supplier admission is claimed. Hosted exact-head CI is tracked in run 40490266757 and will be reported separately when terminal.
5. Smallest useful next task: none justified for measured continuation this round. Future distinct mechanisms need fresh source, ownership, coverage and complete-cost justification before another fixed feasibility gate. Keep PR #5864 frozen. Kappa research, Work 1/2, feeds and automations remain paused; no successor work is dispatched.
6. Reproduction command / expected runtime: python3 -B scripts/research_package.py --package research/fraction-native-score-dispatch-feasibility. Fresh native finite checks take about half a second locally, require existing Linux CPython 3.12 / cc / development headers, and repeat no timings. New publication normal/O/sealed commands were serial, each below one second, under the retained 180-second / 1 GiB sampled process-group / 600-second initial aggregate / no automatic retry guard, with unchanged sources and no stderr.
7. Predecessor credit: Ali Romman and all retained P104/Fraction/integer/native/compiler/precision notices; Adam Gall, ywz65, Mason Hale; PR #2878, PR #5811, PR #5819, PR #5826, PR #5831, PR #5816, PR #5818, PR #5832, PR #5839, PR #5848, PR #5852, PR #5853, PR #5864, and Ali Romman's PR #5761. New probe/execution: Chris R / Clear1nsight with substantial OpenAI Codex assistance; GPT-6-astra xhigh static nomination/stop review only, no compiler/numerical run. Publication audit is donor-requested agent work. Standard dispatch/C API techniques; Apache-2.0. No merge or selected-frontier change.

[Clear1nsight]

Exact-head hosted CI is now green at 3ed05e3: run 40490266757. Planning passed 513 tests with 22 platform-specific skips and selected only research/fraction-native-score-dispatch-feasibility. On Linux CPython 3.12.15 / GCC 13.3, the fresh native replay passed all 33 prefix comparisons and ten controls, reported PASS_REGISTERED_RESEARCH_PACKAGE outcome=negative-result, and the committed-input diff check passed. No timing or full benchmark was replayed. The PR remains open/unmerged; this confirms the declared finite/source replay, not a complete performance gain, scientific acceptance or kappa admission. Canonical seven-field handoff.