psiloceyeben

/

oracle7-slim

Like 1

This follows the 4.9.1 repairs incorporated into 10.0. It is based on current main 0e0e858238cb028b751a85a23f4690a0b0d48483; the affected runtime files are unchanged from the reviewed 4.9 usage paths, so these repairs apply without replacing the new 10.0 organs or routing layers.

Two failures reproduced again against the current 10.0 source:

• _base_parse('Convert binary 11111111 to decimal.') binds frm=10, n=11111111, and returns 11111111; it should bind base 2 and return 255.
• GeometryOrganV4().answer('A rectangle has length 3 meters and width 4 centimeters. What is its area?') returns bare 12; the physical result is 0.12 m².

The repairs keep complete inputs and the delivered answer connected:

• Honor explicit source bases; reject malformed numeric input; decline nonfinite mathematics in all three imported math-organ versions.
• Bind battery capacity, inverter rating, load, losses and requested output separately. Keep trailing physical conditions with the calculation they qualify.
• Preserve date/title/relation constraints when selecting source records, and refuse conflicting covering passage sentences instead of selecting one polarity.
• Reconcile arithmetic corrections and author follow-ups with the final delivered turn; preserve private/other-conversation refusals. Grammatical “it” in a self-contained duration/cost question no longer becomes an old topic.
• Bind geometry dimensions completely, retain displayed units and distinguish diagram “in” from inches. The live specialist/family route converts supported mixed-unit rectangles and refuses unsupported mixed physical dimensions instead of multiplying raw numbers.

Validation on this 10.0 checkout: 158 focused unittest cases passed, including the actual geometry specialist → Family → w504 consumer, actual w488 formatting consumer, all three math-organ versions, and controlled conversation folds. Reproduction commands are in tests/usage/README.md; tests need no large stores or full Engine. The geometry specialist's earlier unchanged published scorer replay was 96 dev and 89 heldout valid correct plus four correct refusals in each split, with zero wrong/false-yes results; its scalar scorer alone did not detect the dimensional bug.

Earlier full-Engine 4.9-based candidate checks observed the repaired mixed-unit and diagram answers and preserved conversation controls. The latest 57-request local replay also included a separately enabled transfer specialist and other repairs, so its result is not a qualification of this PR or of 10.0. A 10.0 whole-Engine/release gate has not been run here.

This PR contains 23 runtime/helper/manifest files plus focused tests and small synthetic/public-fact fixtures. It excludes Mini/service plumbing, stores, date metadata sidecars, generated kernels, sibling changes and the unqualified new data-transfer specialist. Historical admission verdicts and release hashes are not rewritten. Recognition remains untrusted Python; the changes do not claim new HOL proofs or upstream admission.