Add finite cpomdp controls and correct discrete native filtering

[docxology]

Continuous cpomdp sources can declare an ordered finite action set, unique action IDs, units and an asymmetric state-by-control matrix. The renderer preserves equal vectors as distinct choices, records every policy's pragmatic and epistemic components, and refuses incompatible modes, ambiguous aliases and unsupported backends. Total policy-step evaluations are bounded before enumeration. Existing compass-rose and passive contracts remain available.

The full native corpus audit also exposed invalid all-zero posteriors in discrete PyTorch/NumPyro runs that exited zero. Both templates now propagate the previous posterior through the previous action's transition matrix before conditioning on each new observation; t=0 conditions the initial prior directly. Zero/nonfinite predictive evidence and invalid persisted validation fail execution. Ten independent native filtering/EFE and failure controls pass, including NumPyro float32/float64. The four original invalid traces remain retained as regression evidence.

Validation:

• 82 focused checks pass without skips on each supported JAX split: cpomdp 0.4.4 with JAX 0.9.2/Python 3.11 and JAX 0.11.2/Python 3.12. Independent NumPy witnesses compare all posteriors and every policy component; the maximum observed absolute error is 2.7e-15.
• All 38 tracked models were checked through the public renderer: five supported native runs succeeded and 33 were explicitly unsupported. Protected runtime metadata and input bytes are unchanged.
• A separate fresh 38-model/12-backend matrix at qualified revision 3bfa9531e1b4a90906e908aaea3856160126b852 completed 235 inference programs and 24 categorical diagram programs, with 197 explicit unsupported combinations and no native execution or reported native-validation failures. Source/script/result hashes and authored timestep checks are retained; process success is separate from numerical equivalence.
• 36 of 38 tracked models have at least one successful inference backend. The hybrid_discrete_continuous.md and multi_agent_lgssm.md compositions remain unsupported across all 12 backends; their 24 refusals are included in the unsupported total, and broader composition execution remains open.
• Execution coverage is separate from complete semantic fidelity. The deep-planning example declares five-step planning and 64 pruned policies without defining the retained policy set; observed PyMDP/PyTorch/NumPyro runs use one-step evaluation. That source-policy contract remains an open E4 follow-up and is not qualified by the execution matrix.
• The final local core run passed 8,206 tests plus 10 subtests. All 16 repository gates and 68 existing composition/CLI/continuous-control checks passed. Python 3.15 admission timings and a deterministic trace are retained without an inference acceleration or full scientific compatibility claim.
• An isolated nine-package released-wheel Python 3.15 CPU environment also runs all five admitted tracked cpomdp exports at authored timesteps. Eight finite-control executions match the independent NumPy posterior/every-policy reference within 2.7e-15. Five repeated four-step runs have median wall time 0.058909s, with separate first-call timing and a deterministic trace. This does not certify the complete scientific environment or claim a before/after speedup.
• The default managed Mac local-LLM recipe completes an unchanged tracked model and the new finite-control source: 18 complete answers in four fresh request groups, using gemma3:12b with stable identity and 100% GPU residency. This qualifies optional generation and completion, separate from inference correctness.
• The selected parse/validate/export/render/execute/analyze pipeline completed native inference, posterior/EFE analysis and framework reporting. It retains two advisories: the default execution sandbox is off, and the optional cross-model report does not yet include experimental cpomdp.
• All 152 manuscript checks passed. The genuine 31-page render at source 4ed8b330e7af2cfb9e6007664cba4c0939043f45 was visually checked and committed as dc7bcf72da78bd95d21c8204a98aea1ed9be4731; strict postcommit custody passes.

Companion PR #191 refreshes the official source pin and both emitted signatures while preserving all 197 FEP owners and both mathematical document bodies relative to its accepted FEP runtime base. All six bridge checks pass and native_claim_ready remains false. The final commit pins reviewed companion main 184ee53f55b526fdc858e2d962eba16e6066fb19; final exact-head and published-main gates remain required. E3-FA1 is bounded; broader E3/E4 work remains open.