feat(semantic): productize the Core tabular evidence lifecycle

[onlyxItachi]

Scope and stack

Additive v1.1 tabular product layer on PR #93. Foundation/base:
4d9a864640ae0c1174f536070209244c625f38aa.
Current head: 8219c3a486ea50198968c7557849a136cb5d6926.
Current GitHub test-merge commit: e55cbdca344a0dce5ea3c2682dda9562e99bfcda.
This draft targets the foundation branch, not the isolated v1.0 stabilization
line. Related: #72, #73; neither issue is complete. No merge or release is authorized.

Implemented boundary

• gafime.semantic: native-backed declaration namespace, immutable snapshots,
  row-keyed optional labels/graphs, opaque candidate and accepted-set handles,
  explicit proposal/evidence/policy, bounded repeated discovery and unlabeled inference.
• One canonical association specification combines Pearson, Spearman or fixed
  corrected NMI with reference, paired-view or optional-label contexts; weighted
  graph energy remains a separate contextual primitive. No pseudo-target,
  universal quality score or generic significance claim.
• Accepted-batch unions authorize later reuse without allowing raw candidate
  handles to manufacture acceptance. Context rebinding preserves program identity.
• Rust owns validation, semantics, selection, resource admission and lifecycle.
  Core reuses native ranking, histogram, Pearson, pointwise SIMD and Rayon machinery.
  Python contains no candidate/row/sample execution loop.
• A backend-tagged host/resident ownership seam supports a separate native-lowering
  child PR. This PR executes Core only; explicit unsupported backends fail
  closed and auto reports Core honestly.
• Current reference, coverage ledger, docstrings, architecture/package checks,
  installed-wheel smoke and adversarial public tests document this boundary.

Adversarial corrections

Normal thread-affinity exceptions instead of PyO3 assertion panics; bounded
indexed extraction instead of unbounded iterators; validated metadata before
owned conversion; existing Core aliases; single-row accepted inference;
inference contexts cannot mint discovery acceptance; accepted-set unions;
terminal closed-state checks before declaration callbacks; source-only imports
and the legacy wildcard-export surface remain compatible.

Exact-head evidence

• Rust 1.89.0 MSRV and 1.97.1 release compiler: workspace check, strict Clippy,
  fmt and full workspace tests pass (444 passed / 3 intentionally ignored on each).
• Isolated CPython 3.14 installed wheel: 414 passed / 13 skipped; installed
  smoke and Core wheel composition pass. Wheel SHA-256:
  1c444b8492afb1a851136e55b1f307045548121724e25e0f44b95c01c0e23441.
• Exact v1 supervised snapshot across three profiles, four metrics, two seeds,
  ranking and significance: one-worker and default-worker outputs are byte-identical
  to the saved main reference. SHA-256:
  80d280d6725bfd540197e0ab636096735102bd75c4552241fcc0942d9e93e051.
• Notebook generator parity, public API/documentation tests, architecture gate,
  release-facing policy and source-tree composition pass.
• Bounded Core lifecycle diagnostic: 8,192 rows / 78 candidates, fp32/mixed/fp64,
  1/4/24 workers, five samples per mode. Cold evaluation materializes 180 nodes;
  accepted-resident evaluation materializes 91 with 78 retained hits. Remaining
  work is the separate paired view plus the reference, not recomputation of the
  retained same-frame candidates. All value-equality assertions pass. These are
  lifecycle/counter observations, not throughput or universal speed claims.
• V1 Contract Validation: success.
• Native Platform Validation: success.
• Build and Validate Wheels: success; all jobs completed.
• Independent exact-head AI Review Record: PASS, no unresolved P0/P1/P2; review threads: zero.

Evidence hygiene: an initial shared Cargo output directory reused child-branch
test binaries. That release-compiler run and its potentially affected local wheel
were discarded as final evidence. The release-compiler suite and wheel above
were rebuilt/retested in this PR's isolated output directory without changing source.

The informational Core production benchmark run
34068178825
stopped at its explicit main-base guard (PR no longer targets main), before
building or timing. This is a stacked PR; no benchmark success or speed claim is
made and no comparative campaign was rerun to work around that guard.

Deliberate limits

The built-in program vocabulary is sources, absolute difference, softsign and
ordered products with explicitly frozen means. Legacy supervised families remain
compatible specialized lowerings, not a generic-DAG rewrite. Accelerator semantic
execution is separate child work; this PR makes no physical GPU claim.

No full compiler/IR/JIT, contributor frontend, learned encoder, CV/NLP dialect,
serialization ABI, RT promotion, Polars 2 or multi-device scheduling is included.
No package/version topology change and no merge/release.

Design and mathematical limits: docs/v1.1-tabular-semantic-product.md.

[onlyxItachi]

AI Review Record

Model: gpt-5.6-terra (ultra)
Role: independent reviewer; authored none of the PR changes
Exact reviewed head: 8219c3a486ea50198968c7557849a136cb5d6926
Base: 4d9a864640ae0c1174f536070209244c625f38aa
Verdict: PASS — no unresolved P0/P1/P2 static-review finding

Reviewed the complete 38-file diff: native Python declarations/input/output/lifecycle, canonical candidate/evidence/context/selection ownership, Core statistical/native reuse and precision, bounds, immutable/accepted reuse, generic residency boundary, old public/ABI compatibility, documentation/coverage and tests. The separate GPU child implementation was not included and no physical GPU claim is approved here.

Resolved findings: terminal getters and method-body extraction now reject closed sessions; explicit semantic module discovery preserves legacy source-only wildcard imports. Regression tests cover these corrections. No additional correctness, safety, numerical, ownership or unsupported-claim blocker was found.

This record is an independent static review, not a substitute for exact-head validation. The final installed-wheel suite has separately passed 414 tests with 13 skips, along with notebook, architecture and composition checks. Required hosted checks and final Rust qualification must still be verified independently before any merge. PR remains draft; no merge authorization is inferred.