PyMC 6 migration: final QA — integration suite, behavioral baselines vs pymc 5, release prep

[cetagostini]

Part of #1038. Final gate before merging the integration branch into main. Depends on all other child issues.

Goal

Prove the migrated stack is behaviorally equivalent (where intended) and release-ready, then merge pymc6_and_pymcmarketing1_migration → main.

Tasks

• Full test suite in the migration env, including integration/slow markers and doctests (CI matrix 3.12 and 3.14).
• Numerical baseline comparison vs pymc 5: for a fixed seed, compare a representative set of outputs (effect summaries, HDI tables, R² scores, counterfactual predictions) between current main (pymc 5 env) and the integration branch. Document any drift beyond MC noise — expected sources: sampler differences, arviz interval policy (must stay 0.94 HDI), draw-wise R² (must be identical).
• pandas 3 sweep: pandas 3.0.x arrives transitively (copy-on-write default, string-dtype changes). The suite is green, but scan causalpy/data/simulate_data.py and the experiments for chained-assignment patterns and dtype assumptions.
• Deprecation cleanup: LevelTrendComponent → LevelTrend (pymc-extras statespace; causalpy/pymc_models.py:2335); note the pymc-extras BSplineBasis.infer_shape fgraph DeprecationWarning resolves only with pymc-extras ≥0.13 (which needs pymc ≥6.1 — blocked by pymc-marketing's <6.1 cap; document, don't force).
• Docs note on the new runtime characteristics: numba is the default pytensor backend (hard dep, noticeable first-compile latency; native crashes replace Python tracebacks in workers), and nutpie becomes pymc's default NUTS sampler only if installed (400 tune-step default vs pymc's 1000) — consider recommending it in the installation docs.
• Changelog/release notes: breaking-change section (Python ≥3.12, pymc ≥6.0.1, arviz ≥1.1; idata objects are now xarray.DataTree). Decide version bump (this is a major/breaking release).
• Open the final PR: pymc6_and_pymcmarketing1_migration → main.

Acceptance criteria

• Everything green (unit, integration, doctests, notebooks, docs build, lint) on the integration branch.
• Baseline-comparison report attached to this issue.
• Final PR to main opened with the breaking-changes summary.

Workflow

Branch from pymc6_and_pymcmarketing1_migration; open PRs with base pymc6_and_pymcmarketing1_migration (not main) — except the final merge PR described above.

[cetagostini]

Coordinator-run PyMC 5 → PyMC 6 baseline evidence: PASS

I ran the pinned harness from #1102 in four independent processes against the only revisions registered by the migration protocol:

• PyMC 5 reference: 79c0a87072fd4653bfaed1eb085f965594c7f03a
• PyMC 6 candidate: 18a524a1a8512aaa21c46e0ccddbc54501c9eb1a
• Explicit 94% HDI, four serialized NUTS chains, 1,000 tune + 1,000 retained draws, seed 1048, target_accept=0.95, max_treedepth=12, explicit tail ESS probabilities (0.05, 0.95)

Result:

Gate	Result
Within-stack exact raw-draw repeatability	PASS (19/19 PyMC 5; 19/19 PyMC 6)
Protocol and harness equality	PASS
Effect-table/coordinate semantics	PASS
Absolute mean-delta gate	PASS
Standardized posterior-drift gate	PASS
Sampling validity	PASS: zero divergences/tree-depth events; max R-hat 1.005158; min bulk ESS 1688.9768; min tail ESS 1742.0680

Cross-stack means are identical to floating-point precision for the registered DiD effect/R²/counterfactuals and Synthetic Control average/cumulative effects/R²/counterfactuals. No migration drift beyond MC noise was found.

Evidence bundle (rendered report, comparison JSON, all four capture JSON artifacts, and exact command log): https://gist.github.com/cetagostini/62d0ebf197c99fd7eef4336fb7de46a1

The first PyMC 6 development attempt exposed that ArviZ 1.x requires an explicit tail-ESS prob; #1102 now registers (0.05, 0.95) on both stacks, rejects unsupported APIs rather than falling back, and has regression coverage. All attached final artifacts were regenerated from clean pinned checkouts after that fix.