Use pathmc for causal-discovery tooling

[cetagostini]

Warning

Draft — do not merge yet. This depends on:

1. pathmc shipping the causal-discovery front end (Port causal-discovery tools from PyMC-Marketing into pathmc pathmc#307, PR Add causal-discovery front end (TBFPC, CPDAG, DAG→model) pathmc#349), and
2. pymc-marketing migrating to PyMC 6 — pathmc requires pymc>=6, while pymc-marketing currently pins pymc<6, so the two cannot be installed together until then.

Once both land, bump the pathmc floor in the dag extra to the release that includes the discovery tooling and flip this out of draft.

Description

Re-points the optional causal-discovery tooling at the standalone pathmc library, so the discovery algorithm and DAG→model builder are maintained in one place (pathmc now owns PyMC-Labs' path-analysis / structural-causal-modeling stack). Behavior is unchanged — the names resolve to the same implementations — but they now come from pathmc.

• mmm/causal.py — TBFPC, BuildModelFromDAG, and TestResult are re-exported lazily from pathmc via a module-level __getattr__ (PEP 562). Importing the module no longer requires pathmc (so the core MMM import is unaffected); only accessing those names triggers the import, with an actionable error pointing at pip install pymc-marketing[dag]. CausalGraphModel (dowhy-based backdoor adjustment) stays here — it is imported by the core MMM and pathmc deliberately provides its own native identification rather than wrapping dowhy.
• causal_utils.py — same_markov_equivalence_class_CPdag becomes a thin wrapper over pathmc.same_markov_equivalence_class (drops the pydot dependency; pathmc uses a dependency-free DOT reader).
• pyproject.toml — adds pathmc to the [dag] extra.
• tests — the discovery behavior is now covered exhaustively in pathmc (159 tests). The duplicated TBFPC / BuildModelFromDAG / Markov-equivalence tests are removed here; what remains is the CausalGraphModel suite (unchanged) plus wiring tests that verify the re-exports resolve to pathmc (skipped when pathmc is not installed, i.e. on PyMC 5 CI).

Related Issue

• Related to Port causal-discovery tools from PyMC-Marketing into pathmc pathmc#307 (discovery port) and the structure-uncertainty epic [Epic] Causal discovery → structure uncertainty (learn-the-DAG, model-average the effects) pathmc#306
• Pairs with pathmc PR Add causal-discovery front end (TBFPC, CPDAG, DAG→model) pathmc#349

Checklist

• Pre-commit linting/style checks (ruff/ruff-format) pass on the changed files
• Tests included (CausalGraphModel suite retained; pathmc re-export wiring tests added; discovery behavior tested upstream in pathmc)
• Documentation updated (module docstrings explain the pathmc move and the [dag] extra)

Note: the pathmc-backed tests skip in current CI because pathmc cannot be installed alongside PyMC 5. End-to-end validation on this side runs once pymc-marketing is on PyMC 6; the behavior itself is validated by pathmc's own suite.

🤖 Generated with Claude Code

---

📚 Documentation preview 📚: https://pymc-marketing--2642.org.readthedocs.build/en/2642/

[codecov]

Codecov Report

❌ Patch coverage is 57.14286% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 93.05%. Comparing base (649547e) to head (6e80605).
⚠️ Report is 5 commits behind head on main.

Files with missing lines	Patch %	Lines
pymc_marketing/causal_utils.py	40.00%	3 Missing ⚠️
pymc_marketing/mmm/causal.py	66.66%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2642      +/-   ##
==========================================
- Coverage   94.19%   93.05%   -1.15%     
==========================================
  Files         100      100              
  Lines       15384    14835     -549     
==========================================
- Hits        14491    13804     -687     
- Misses        893     1031     +138     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[juanitorduz]

As this is a breaking change, I just added it to the v1.0 milestone