test(callaway): machine-precision golden tolerances — SE-audit C6 closure

[igerber]

Summary

• Closes the SE-audit C6 remainder ("only dr-tightening remains"). Measurement-first: the no-covariate DR golden agreed with R did at ~6e-11 ATT / ~2e-11 relative SE (no-covariate DR is deterministic algebra — no propensity IRLS enters), yet the test asserted ATT only, within 0.02, and never asserted the SE. Both are now pinned at 1e-8 with the SE assertion added.
• The reg (0.02) and ipw (0.05) ATT bands — both predating the DRDID estimation-effect IF terms that brought the SEs to machine agreement — are tightened to 1e-8 on measured ~3e-11/~4e-11. The ipw SE keeps its documented 1e-6 IRLS-vs-glm platform headroom; the covariate-DR scenario keeps its documented 2e-3 band (measured ~9.5e-4 — the known DR small-sample numerics via the propensity nuisance).
• TODO SE-audit row updated in place: C6 marked closed with the measured values; C7/C8 flagged as needing re-derivation from a fresh loose-tolerance inventory (their definitions lived in the audit session's scratchpad).

Methodology references (required if estimator / math changes)

• Method name(s): Callaway-Sant'Anna ATT(g,t), DRDID dr/reg/ipw panel estimators — test-tolerance changes only, no code
• Paper / source link(s): existing golden fixtures from R did 2.5.1 (benchmarks/data/csdid_golden_values.json, unchanged)
• Any intentional deviations from the source (and why): None

Validation

• Tests added/updated: tests/test_csdid_ported.py (3 golden tests tightened; 1 new SE assertion). Full file passes (72).
• Backtest / simulation / notebook evidence (if applicable): in-session measurement sweep of all four dr/reg/ipw golden scenarios (values quoted above).

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall Assessment

✅ Looks good — no unmitigated P0/P1 findings.

Executive Summary

• PR changes only CHANGELOG.md, TODO.md, and tests/test_csdid_ported.py; no estimator, weighting, variance, SE, assumptions, or defaults changed.
• Affected methods are Callaway-Sant’Anna ATT(g,t) golden tests for no-covariate DR, covariate reg, and covariate ipw.
• Registry and fixture generator support the tightened R did::att_gt() parity checks; no undocumented methodology mismatch found.
• No new inline inference, partial NaN guard, control-group, or parameter-propagation anti-patterns found.
• Targeted tests could not be run locally because pytest is not installed in this environment.

Methodology

• Severity: P3 informational
  Impact: The tightened assertions at tests/test_csdid_ported.py:L950-L1042 are consistent with the CallawaySantAnna registry’s IF-based SE contract and R did::att_gt() parity target (docs/methodology/REGISTRY.md:L474-L545, docs/methodology/REGISTRY.md:L782-L790). The changed scenarios are generated from R did::att_gt(..., bstrap=FALSE, cband=FALSE) in benchmarks/R/generate_csdid_test_values.R:L71-L83 and benchmarks/R/generate_csdid_test_values.R:L95-L140.
  Concrete fix: None required.

Code Quality

• No findings. The modified tests do not introduce inference computation, new public parameters, or changed estimator control flow.

Performance

• No findings. The PR only tightens small golden-value assertions over existing fixture cells.

Maintainability

• No findings. The comments are verbose but useful for preserving why the tolerance choices differ across DR/reg/ipw paths.

Tech Debt

• Severity: P3 informational
  Impact: C7/C8 remain deferred, but the PR tracks that explicitly in TODO.md:L63; this is acceptable deferred work and not a blocker.
  Concrete fix: No action for this PR; re-derive C7/C8 from a fresh loose-tolerance inventory when that follow-up is picked up.

Security

• No findings. No executable production code, credentials, network calls, or secret-bearing files are changed.

Documentation/Tests

• Severity: P3 informational
  Impact: The added SE assertion and tightened tolerances are documented in CHANGELOG.md:L52-L59 and the test docstrings. I could not execute the targeted tests because pytest is unavailable (pytest: command not found; python -m pytest: no module named pytest).
  Concrete fix: Ensure CI runs the touched golden tests, especially TestCSDIDGoldenValues::{test_golden_default_dgp_dr,test_golden_default_dgp_reg,test_golden_default_dgp_ipw}.