[docxology]

Public source validation could probe long Unicode content as a filesystem path, ignore explicit literal-content mode, or reinterpret a file's contents as another filename. The public facade now delegates once to the guarded formal validator, which explicitly honors is_content=True. Empty/directory/invalid UTF-8 inputs return structured invalidity. A missing required Footer in the compact coordination acceptance fixture is appended without changing its original parameter/equation bytes. All 38 tracked models now pass STANDARD formal validation with identical file/content verdicts.

A tracked GNN model declares five-step planning over 64 pruned policies, but ordinary planning adapters used one-step scores. The shared source-admission guard now refuses unsupported horizons, conflicting aliases, incompatible policy counts and explicit policy lists before writing scripts. PyMDP, ordinary JAX, RxInfer, ActiveInference.jl, PyTorch and NumPyro share the rule; alternative JAX solver/demo entry points refuse declarations their defaults do not consume. Existing destinations survive refusal.

All authored parameter/equation bytes and horizons/policy counts, dependency locks and default runtimes are preserved. Only the missing Footer is appended; the other 43 tracked input files are byte-identical. No retained policy subset or pruning algorithm is invented. Named JAX execution contracts and continuous finite-action cpomdp retain their separate horizon rules. Stan/bnlearn inference, THRML fixed-action smoothing and DisCoPy diagrams remain separate questions. E4-P1 admission is bounded; multi-step source semantics and independent source-to-numerics qualification remain open.

Validation:

• 105 focused public/formal validation controls and 163 planning/renderer/native controls pass with zero skips on each of Python 3.11, 3.12 and 3.14. Real subprocess comparisons against the immutable pre-fix source reproduce and resolve the wrapper defects. All 16 repository source gates pass.
• The full core selection passes 8,322 tests plus 10 subtests, with zero selected skips.
• Fresh 38-model × 12-backend rendering/native execution produced 229 inference programs and 24 categorical diagram programs, with 203 explicit unsupported combinations and zero execution or reported validation failures. All 456 outcomes are unchanged after the public-validation correction; relative to pre-batch main, exactly six affected planning combinations now refuse and all other 450 outcomes remain unchanged. All 767 runtime owners and 44 tracked input files retain their qualified hashes. Execution success is separate from complete prose-semantic equivalence.
• Timestep receipts retain authored and reported values. The explicitly labelled static JAX exception conditions one observation; it is not a full temporal rollout.
• A separate independent source-literal audit of the tracked 50-step HMM matches seven native questions: five filtering outputs, RxInfer full-sequence smoothing, and ActiveInference.jl next-state prediction. Short-prefix path enumeration agrees within 1.7e-15, and seven negative controls distinguish/refuse altered conventions. THRML is diagnostic without an ESS/convergence qualification; Stan's emission-parameter inference answers a distinct question from the source's fixed-emission HMM. This audit does not establish universal extraction equivalence.
• Python 3.15's standard tracing profiler and five unprofiled samples exercise 10,000 admitted one-step, 10,000 refused five-step and 10,000 invalid Boolean declarations per sample. A separate lightweight processing profile covers all 38 exact inputs. These profiles do not establish a speedup, inference/GPU performance or complete scientific-stack compatibility. Missing cp315 wheels remain explicit; the frozen Python 3.14 default is preserved.
• Genuine manuscript rendering at source 48db2919b8ab171e0c2a349e1b850804ad0de928 produced 31 pages. All six contact sheets and the updated count figure were reviewed; all 152 manuscript checks and strict postcommit freshness/hydration/token gates pass. Artifact child: be80431c471a3dd30fc93d690069a588b979ab26.
• A separately committed input/test qualification leaf 05190ae35860678991189b83cb4730165fee58d0 adds only the Footer and the all-model validity assertion. All 771 source-owner bytes, manuscript counts and actual render bytes remain unchanged; strict manuscript checks still pass. Fresh native execution of all 456 combinations at this leaf preserves every program outcome; the one source-digest change is explicitly recorded.

Companion PR #193 records the planning-admission refresh; PR #194 refreshes official custody for the two subsequently corrected validator owners. All 771 GNN and 197 FEP owner hashes, six source checks and both emitter freshness checks match; both FEP mathematical document bodies and all FEP runtime owners are preserved. native_claim_ready remains false. The final sole pin commit points to reviewed companion main 0d980814b8a4cbf167cefbf262143b6498feeb86. Final exact-head and published-main checks remain required before closeout.