Add delete-d option to the observational relate influence gate

[kgdunn]

Summary

• The leave-one-out influence gate on marginal associations has a known blind spot at two supporting observations: a correlation carried by exactly two points survives, because deleting either one leaves the other holding it up, so the jackknife interval never spans zero.
• Add an influence_deletions parameter (default 1) to analyze_descriptive / relate_observational that sets how many observations the marginal jackknife removes together.
  • d = 1: unchanged - the shipped leave-one-out jackknife-t decision (byte-identical; all prior tests pass).
  • d >= 2: the averaged jackknife variance would dilute the single subset that removes both supporting points, so the decision switches to a breakdown criterion - the correlation must stay significant, with the same sign, after removing every subset of d observations. A two-support spike is then demoted while a driver supported by more than d observations is kept.
• The reported jackknife_se remains the ordinary leave-one-out standard error; only the influence_robust decision generalizes. The setting is echoed in result.config["influence_deletions"]. No threshold is introduced - the rule reuses the existing alpha, the number of observations, and d.

Test plan

• uv run pytest tests/test_sensory.py tests/test_sensory_end_to_end.py --no-cov (49 passed). New tests: a two-support spike survives d = 1 but is demoted at d = 2, while a genuine broad driver survives both; the pipeline-level influence_deletions=2 demotes a two-product spike the default gate keeps; max_deletions=0 raises.
• Verified on synthetic and real sparse data: at d = 2 every association resting on one or two products (and the three-support ones) is demoted, while genuine multi-product drivers survive; d = 1 results are unchanged.
• uv run ruff check . passes.
• uv run mypy src/process_improve passes.

Checklist

• Version bumped in pyproject.toml (MINOR: 1.56.0 -> 1.57.0, new opt-in parameter; CITATION.cff synced)
• Tests added or updated where relevant
• ruff check . passes
• CHANGELOG.md updated

🤖 Generated with Claude Code

https://claude.ai/code/session_018M9u5EpSjHa3mjNaYDJ7qq

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!