Correct quantity binding, final-outcome evidence and isolated runtime state
This fixes concrete correctness and integration failures in slim 4.9 at 083e3a6c. Changes are grouped around question/quantity interpretation, evidence propagation, and the state visible to a caller.
- Swarm plans keep named quantities distinct: height cannot satisfy width merely because both use metres. Explicit aliases connect compatible roles. Research and fallback parsing require unambiguous compatible measurements rather than the first digit in prose.
- Formula execution and export use a private snapshot whose admission hash was checked. Cached signatures cannot authorize changed, revoked or expired programs. Local helper/data files follow the same snapshot; this is version binding for trusted Python, not a sandbox.
- w551 replaces blanket
provedattribution with a scopedkernel_checkretaining the whole kernel reply. Binaries are SHA256-bound; request/reply sizes, pipe writes and reads are bounded, with distinct request IDs. The same change is proposed for GitHub 4.0 in https://github.com/psiloceyeben/oracle7/pull/1. - Terminal contradictions survive later fallback folds. The composed fallback requires a unique candidate and actual output movement when checking input dependence; constant outputs and parser-only differences do not count.
- A shared final-turn boundary records the delivered answer once, including replies returned by newer tip folds before calling a parent. Replacement/withholding drops stale proof fields, and a withheld lower answer cannot become a later pronoun referent. Pronoun re-entry carries the prior delivered evidence separately from its rewritten result. Rollback covers named in-memory stores, not arbitrary effects.
- Possession alone is no longer treated as authorship. An exclusive-creator contradiction requires identified entities, the relevant citation and positive conflicting source evidence. Existing WITH/YEAR premise guards still need work: missing relation rows are not proof of contradiction, and an incidental first QID can be mistaken for the requested work.
- Both meteorology generations validate whole numeric spans, units, humidity/temperature roles and unused operands. Fixing v2 alone was insufficient: the actual Family loader could recover a wrong answer from the older organ. Tests cover both orders. Both manifests recognize ISA refusals.
- Explicit ORACLE7_STATE now owns vessel JSON/events/requests and the derived SQLite store. A real Engine smoke found the old code writing these into the source tree despite a separate state directory.
Validation: 102 focused unittest cases across the seven affected surfaces, including the actual fold functions with controlled dependencies, executable pipe/cache fixtures and the real Family loader. Meteorology v1 scores 98/98 and v2 110/110 on their shipped panels, with 59 independent quantity checks and no wrong/false-yes answers in those scoped runs. These are not a new whole-runtime gate or held-out admission.
A separate local appliance test materializes all ten stores with pinned compressed/raw hashes and exercises the real Engine via HTTP, including ordinary facts, weather, pronoun follow-up, exact retry, conflicting retry, graceful shutdown and replay after restart. Full source-tree mutation checking is included. The appliance adapter/packager and new veri7vn Earth binary are maintained separately; this PR contains source repairs and focused tests only.
The existing verdict and RELEASE_HASHES files are intentionally not rewritten to pretend the changed code still has the old qualification. Both weather bundles and the changed runtime need the maintainer's fresh admission/release checks. No models, corpora, credentials, generated binary payloads or private transcripts are included.
Thank you for carrying these repairs into 10.0. I checked current main 0e0e858: 32 of this PR’s 34 changed files are byte-identical there; w565 and swarm/point.py have further upstream changes. The newer quantity/conversation/geometry repairs are now proposed against 10.0 in #2 with 158 focused checks. The separate current-source orbital-year and date-precision counterexamples are recorded in #3 and #4 with immutable patch links and explicit qualification limits. This note does not claim a new whole-Engine 10.0 gate.
- root/fable-content/oracle-m-series-2026-08-15/self_loop/admitted/oracle_release_runtime_w557.py
- root/fable-content/oracle-m-series-2026-08-15/self_loop/admitted/oracle_release_runtime_w565.py
- root/fable-content/oracle-m-series-2026-08-15/self_loop/admitted/runtime_final_outcome.py
- workbench/swarm/point.py