v3.2.0 — Exemplar Gold Standard

Compare: v3.1.0...v3.2.0

Exemplar Gold Standard. Every exemplar GNN file renders and executes on
every framework that can represent it and is explicitly flagged unsupported
(not faked, not failed) on the ones that cannot. Continuous-state exemplars
are genuinely continuous (linear-Gaussian) with native JAX / NumPyro /
PyTorch / Stan / RxInfer.jl backends; the Stan renderer is rewritten from a
non-compiling stub into runnable HMM and LGSSM programs with a cmdstanpy
executor; Step 12 summaries merge across input folders; the Julia pre-exec
gate no longer blocks scripts on a toolchain-less launcher.

Verified end to end on macOS with Julia 1.12.7 (RxInfer 5.5.0,
ActiveInference.jl 0.1.2), CmdStan 2.39 and an ephemeral torch: 29 files,
249 renders OK / 0 failed / 12 unsupported; 194 executions OK / 0 failed /
55 dependency skips (torch, bnlearn); PyTorch 29/29 via
uv run --with torch; pytest 3120 passed.

Changed (2026-09-01 — exemplar gold standard: continuous models are continuous, Stan is real)

• Continuous exemplars are now genuinely continuous. The three files under
  input/gnn_files/continuous/ declare only the linear-Gaussian state-space
  model (x, y, optional u, F/H/Q/R, prior_mean/prior_cov, optional
  goal_mean/control_gain) — the discretized POMDP stand-ins introduced on
  2026-08-03 are gone. continuous_navigation is a closed-loop navigator
  (u_t = 0.3 · (goal − μ_t)); the other two are passive. Extractor
  (gnn/pomdp_extractor.py) reads dimensions from F/H and sets
  model_kind = "continuous"; the render processor passes the block through
  verbatim instead of building canonical A/B/C/D.
• unsupported is a first-class render status. Frameworks that cannot
  represent a model kind (PyMDP, ActiveInference.jl, DisCoPy, bnlearn on continuous
  models) return {"unsupported": true, "status": "unsupported"} with the
  reason "continuous-state model: … supports discrete POMDPs only". They are
  excluded from the success denominator, listed under
  unsupported_framework_renderings, and never handed to Step 12.
  framework_registry entries carry supports_continuous.
• Native continuous backends. JAX, NumPyro, PyTorch and Stan render and
  execute the LGSSM (online Kalman filter, Joseph-form update, closed-loop
  control) via the shared generator render/continuous_script.py; NumPyro
  additionally fits the same model with NUTS (mcmc_posterior_means,
  mcmc_r_hat_max). RxInfer.jl's existing continuous strategy is reached
  without the A/B/C/D canonicalisation guard. Verified end to end on this
  machine for all five.
• Stan renderer rewritten (render/stan/stan_renderer.py). The previous
  generator emitted target ~ normal(source, 1.0) per edge with undeclared
  variables and could not compile. Discrete models now render an HMM whose
  latent chain is marginalised with the forward algorithm and whose per-state
  observation distributions carry Dirichlet priors centred on the declared
  A; continuous models render the explicit Kalman marginal likelihood. Each
  render emits <stem>_stan.stan plus a cmdstanpy driver <stem>_stan.py.
  New execute/stan/ runner, stan added to Step 12 discovery and to
  utils.framework_availability (skips with uv sync --extra stan), new
  optional extra stan = ["cmdstanpy>=1.2"]. render_stan() remains as a
  declaration-only sketch.

Fixed (2026-09-01)

• Step 12 summary no longer overwritten per input folder.
  execute/processor.py merges the prior execution_summary.json (details
  keyed by script path; counts, per-framework status and overall status
  recomputed) so the durable summary covers every folder, mirroring Step 11.
• Julia pre-exec gate false positive. A julia launcher without a
  working toolchain made Meta.parseall return a non-zero exit with no
  GNN_PARSE_FAIL marker; the gate treated that as malformed code and blocked
  every Julia script. The probe now degrades to the advisory regex sweep
  unless the parser itself reported a failure.
• INDEX.md excluded from model discovery (gnn/discovery.py), fixing
  four tests that counted 30 exemplars instead of 29.
• CI hygiene. ruff format drift across 45 files cleared; Bandit B108
  (hard-coded /tmp in the CLI models command) replaced with
  tempfile.gettempdir(); the two Julia-live test modules now skip when the
  committed Julia project environments are not instantiated (CI ships julia
  without RxInfer/ActiveInference.jl); the framework-availability test honours
  toolchain probes; repository-terminology audit clean.
• ActiveInference.jl precompile on Julia 1.12. Distributions pinned to
  0.25.100 – 0.25.125 in src/execute/activeinference_jl/Project.toml
  (DistributionsAD 0.6.58's ReverseDiff extension breaks against the
  @check_args change in 0.25.126); Manifest re-resolved.