Add LP-DiD launch carousel (local-projections crossover narrative)

[igerber]

Summary

• Add the LP-DiD launch carousel: carousel/generate_lpdid_carousel.py (generator, "Horizon" palette: violet/cyan/amber on a lavender gradient) + the rendered carousel/diff-diff-lpdid-carousel.pdf (11 slides, 4:5) + carousel/README.md (regeneration contract, carousel-only deps, claim-discipline convention).
• Narrative spine follows the paper's own novelty argument (JAE 2025, pp. 741-742), with pull quotes verbatim: macro built local projections to read dynamic effects of shocks (cover, IRF motif) -> negative-weighting contamination (pull quote + schematic) -> dark twist slide (a naive LP implementation inherits the same bias, Eqs. 6-7) -> the "underappreciated feature" fusion (LP + clean controls = LP-DiD) -> "unclean" observations mechanism grid -> non-negative weights (sum over g != 0, Eqs. 9-10) + the "subsumes" equivalence map -> symmetric pre-trends (Roth 2024, the authors' claim) -> code / production grid / validation story / CTA.
• Claim discipline documented in the module docstring: equivalence map limited to the four proven nestings (Callaway-Sant'Anna exact Sec. 3.7; Cengiz-style stacking exact Sec. 3.7 - literature design, deliberately never called "Stacked DiD" since the library's StackedDiD implements WFH 2024 Q-weights, which LP-DiD does not nest; classic 2x2 Sec. 2.2; BJS-style exact single-cohort PMD, fn. 10-11). Survey claims scoped to the variance-weighted default path; non-absorbing scoped to entry effects (Eq. 12/13); fixest parity scoped to VW variants; tolerances match the REGISTRY-recorded digits; no speed claims (the paper's Table 2 speedups are Stata-implementation artifacts and do not hold for diff-diff's vectorized CS - verified empirically).
• Footer version label parsed from pyproject.toml at generation time (stdlib regex, Python >= 3.9), so the deck cannot drift from the release it advertises.

Methodology references (required if estimator / math changes)

• Method name(s): N/A - no estimator/math changes (marketing/documentation asset only). Deck copy describes LPDiD.
• Paper / source link(s): Dube, Girardi, Jordà & Taylor (2025), "A Local Projections Approach to Difference-in-Differences," Journal of Applied Econometrics 40(5), 741-758. https://doi.org/10.1002/jae.70000
• Any intentional deviations from the source (and why): None - slide copy quotes the paper verbatim and mirrors the scoping already documented in docs/methodology/REGISTRY.md (## LPDiD).

Validation

• Tests added/updated: No test changes.
• Backtest / simulation / notebook evidence (if applicable): N/A. All numeric claims on the deck (parity tolerances, equivalence scope, survey/non-absorbing limits) restate REGISTRY-documented, test-backed results; the deck introduces no new numbers.

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

• No estimator, weighting, variance, SE, or default behavior changes were introduced.
• Affected method is LPDiD, but only through carousel/documentation copy.
• The deck’s LPDiD methodology claims match docs/methodology/REGISTRY.md and the local Dube paper review.
• The committed PDF text matches the generator output claims and uses the v3.6.1 footer from pyproject.toml.
• No security, performance, or maintainability blockers found.

Methodology

Finding: No findings
Severity: None
Impact: The LPDiD claims are scoped correctly: clean controls/non-negative weights, CS/Cengiz/2x2/BJS equivalence qualifiers, non-absorbing entry effects, and survey default-path scope are all consistent with docs/methodology/REGISTRY.md:L1799-L1895 and the generator’s claim discipline at carousel/generate_lpdid_carousel.py:L34-L55.
Concrete fix: N/A.

Code Quality

Finding: No findings
Severity: None
Impact: The generator parses successfully and follows the existing carousel pattern. Version parsing from pyproject.toml is explicit and fail-closed at carousel/generate_lpdid_carousel.py:L87-L94.
Concrete fix: N/A.

Performance

Finding: No findings
Severity: None
Impact: This is an offline asset generator; no library runtime path is affected. The deck also avoids unsupported LPDiD speed claims.
Concrete fix: N/A.

Maintainability

Finding: No findings
Severity: None
Impact: carousel/README.md:L15-L18 documents the regenerate-with-PDF contract, and carousel/README.md:L30-L39 documents claim discipline against the registry/paper review.
Concrete fix: N/A.

Tech Debt

Finding: No findings
Severity: None
Impact: Existing LPDiD deferred limitations remain documented in TODO.md, including non-absorbing exit-event dynamics and survey reweight/replicate gaps; the deck preserves those scopes.
Concrete fix: N/A.

Security

Finding: No findings
Severity: None
Impact: No secrets or PII found in the added generator, README, or extracted PDF text. The generator only reads local pyproject.toml, writes the committed PDF path, and uses temporary image files.
Concrete fix: N/A.

Documentation/Tests

Finding: No findings
Severity: None
Impact: No new tests were added, but the PR is a documentation/marketing asset. The claims are already backed by existing LPDiD parity tests, including CS/BJS equivalences at tests/test_lpdid.py:L712-L758, R parity at tests/test_methodology_lpdid.py:L227-L305, non-absorbing parity at tests/test_methodology_lpdid.py:L308-L370, and survey parity at tests/test_methodology_lpdid.py:L390-L536.
Concrete fix: N/A.