Bump version to 3.7.0

[igerber]

Summary

• Bump version 3.6.2 -> 3.7.0 across all surfaces (diff_diff/__init__.py, pyproject.toml, rust/Cargo.toml, diff_diff/guides/llms-full.txt, CITATION.cff).
• Convert the curated ## [Unreleased] CHANGELOG section to ## [3.7.0] - 2026-07-08 and add the v3.6.2...v3.7.0 comparison link. No auto-generated content - the hand-written entries (leading with ### Removed) are preserved verbatim.

This is a minor release cutting the 41 commits merged to main since v3.6.2. Highlights:

Removed (breaking, deprecation-cycle complete):

• Deprecated HAD survey= / weights= kwargs removed from all 8 surfaces - HeterogeneousAdoptionDiD.fit (refactor(had): remove deprecated survey=/weights= kwargs from HeterogeneousAdoptionDiD.fit (3.7.0) #633) and the 7 pretest helpers (refactor(had): remove deprecated survey=/weights= kwargs from the 7 pretest helpers (3.7.x) #659). Only survey_design= remains. Deprecation warnings shipped since v3.3.1.

Added:

• ContinuousDiD: covariate support (reg + dr) under conditional parallel trends (feat(continuous-did): covariate support (reg + dr) under conditional parallel trends #616), discrete-treatment saturated regression (feat(continuous-did): discrete-treatment saturated regression (treatment_type="discrete") #618), lowest-dose-as-control (feat(continuous-did): lowest-dose-as-control (control_group="lowest_dose", Remark 3.1) #623).
• Callaway allow_unbalanced_panel RC-on-panel parity with R did (feat(callaway): allow_unbalanced_panel — RC-on-panel parity with R did #631); imputation leave-one-out conservative variance (BJS 2024 App. A.9) (feat(imputation): leave-one-out conservative variance (BJS 2024 App. A.9) #630); Rust one-way HC2 robust vcov kernel (feat(rust): one-way HC2 robust vcov kernel #651).

Changed / performance:

• EfficientDiD kcov hoisting + per-group W lifecycle (100k: 86 -> 18 min) (perf(efficient_did): cross-cell kcov-table hoisting + per-group W lifecycle (100k: 86 -> 18 min) #629), Rust batched-Cholesky ridge solve (perf(efficient_did): dispatch conditional Omega* ridge solves to a Rust batched-Cholesky kernel (fits 1.4-1.8x) #632); Callaway per-cell solver fast paths + IF-scatter perf (perf(linalg): per-cell solver fast paths - IRLS Cholesky inner solve + rank-detection certification (CS covariate fits 1.5-2x) #634/perf(callaway): fancy-+= per-(g,t) IF scatters; drop dead _compute_aggregated_se #636/perf(callaway): stop materializing per-cell unit-label arrays in influence_func_info #640); wild-bootstrap ~7x (perf(utils): precompute r-independent wild-bootstrap inversion pieces (~7x, outputs unchanged) #647).

Fixed / SE parity (behavior changes on reported standard errors):

• Systematic standard-error parity sweep vs canonical R across CS/staggered inference (fix(staggered): CS panel reg/ipw influence functions + per-cell SEs to DRDID parity #619/fix(inference): three standard-error parity fixes vs canonical R #620/CS positional base period + absorbed-FE full-K SE parity vs R #621/fix(callaway): IF-based per-cell SE for no-covariate doubly-robust #627), honest_did nested FLCI center parity (fix(honest_did): nested optimal-FLCI optimizer — R center parity (SE-audit B2b) #635), deterministic clustered vcov in Rust (fix(rust): deterministic clustered vcov — first-appearance cluster aggregation #653), survey explicit-vcov_type contract under replicate designs (fix(survey): unify explicit-vcov_type contract under replicate designs (warn + hc1 remap) #654) - backed by the se-audit machine-precision test locks (test(se-audit): CI-lock flagship + previously-unasserted SE parity (G2 + C1-C3/C5) #624/test(se-audit): fiddly-bits SE parity locks (C2 DOF, C3 df, C4 intercept SE, C5 Yatchew p/sigma2, G2 cluster band) #628/test(se-audit): G2 hetero + cluster machine-precision fixest locks on unbalanced heteroskedastic DGP #644/test(callaway): machine-precision golden tolerances — SE-audit C6 closure #658).

Methodology references (required if estimator / math changes)

• Method name(s): N/A - release chore. This PR changes only version strings and the CHANGELOG; no estimator or math logic is touched here. The underlying methodology changes being released each merged in their own PRs with paper citations and docs/methodology/REGISTRY.md updates.
• Paper / source link(s): N/A (see per-PR references and REGISTRY.md)
• Any intentional deviations from the source (and why): None in this PR

Validation

• Tests added/updated: No test changes - version-string + CHANGELOG only. (The released work carries its own test suites, merged per-PR.)
• Backtest / simulation / notebook evidence (if applicable): N/A

Release note

• Cut as a minor (3.7.0): new user-facing capabilities + a completed deprecation-cycle removal + SE-value corrections toward canonical R. Users re-running existing code against absorbed/CS doubly-robust paths may see reported standard errors move - see the CHANGELOG Fixed / Changed sections for specifics.

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

• The diff is limited to release metadata and changelog sectioning; no estimator, math, weighting, SE/variance, assumptions, or default behavior code is changed.
• Affected methods: none in executable code. The HAD API-consolidation release-note content is already documented in docs/methodology/REGISTRY.md:L3206-L3207.
• Version surfaces are aligned on 3.7.0: pyproject.toml:L7, diff_diff/__init__.py:L304, rust/Cargo.toml:L3, CITATION.cff:L10, and diff_diff/guides/llms-full.txt:L5.
• Release date metadata is consistent: CHANGELOG.md:L8 and CITATION.cff:L11 both use 2026-07-08.
• No new edge-case, NaN/Inf, inference, control-group, bootstrap, or parameter-propagation paths are introduced.

Methodology

No findings.

Impact: This PR does not change estimator implementations or methodology behavior. The only methodology-adjacent changelog text concerns previously shipped HAD survey API consolidation, which is documented in the methodology registry.

Concrete fix: None.

Code Quality

No findings.

Impact: The changed files are straightforward metadata/string updates.

Concrete fix: None.

Performance

No findings.

Impact: No runtime code paths are changed.

Concrete fix: None.

Maintainability

No findings.

Impact: Release metadata is synchronized across Python, Rust, citation, changelog, and bundled guide surfaces. No Cargo.lock exists in the repo, so there is no lockfile version drift to address.

Concrete fix: None.

Tech Debt

No findings.

Impact: The PR does not add deferred work or new limitations.

Concrete fix: None.

Security

No findings.

Impact: No secrets, credentials, network logic, or executable workflow changes are introduced.

Concrete fix: None.

Documentation/Tests

No blocking findings.

Severity: P3-informational
Impact: No tests were added, which is acceptable for a version/changelog-only release PR.
Concrete fix: None required.