feat(replicate): dual-track classic + modern guides for 4 papers

Promote sp.replicate entries for Card 1995, Abadie 2010 (Prop 99),
Lalonde 1986 (DW NSW+PSID-1, MatchIt subset), and Lee 2008 (Senate RD)
from single-track stubs to full classic + modern recipes on real
public-domain CSVs bundled with the package.

- Bundle real CSVs: card_1995, california_prop99, lalonde_matchit (n=614),
  lee_2008_senate (n=1390) under src/statspai/datasets/data/.
- sp.datasets.nsw_lalonde / lee_2008_senate gain a `simulated=False` real-
  data branch with pinned StatsPAI estimates and published benchmarks
  exposed via df.attrs; synthetic generators kept reachable as private
  helpers.
- Lalonde classic track: Naive OLS / Adjusted OLS / 1:1 NN PSM with
  golden numbers vs DW (1999) Tables 3-4. Modern track: DML PLR + entropy
  balancing.
- Lee classic track: conventional jump 7.414 (paper 7.99) at CCT bandwidth.
  Modern track: CCT bias-corrected robust 7.507, SE 1.741, h 17.754.
- Replicate guide formatter splits regression-test drift tolerance
  (StatsPAI version-to-version) from paper alignment delta and tightens
  the Lalonde drift tolerance to $5.
- Tests: new TestReplicateLalonde and TestReplicateLee2008 dual-track
  suites; legacy single-track suite trimmed. 28/28 dual-track tests pass;
  broader sweep across replicate/external_parity/registry/iv stays green.
- All 13 BibTeX keys cited across the four entries verified in paper.bib.

Author: brycewang-stanford

src/statspai/datasets/_canonical.py

@@ -272,7 +272,65 @@ def card_1995(seed: int = 42, simulated: bool = True) -> pd.DataFrame:
 # LaLonde (1986) — NSW experimental
 # ---------------------------------------------------------------------------
 
-def nsw_lalonde(seed: int = 42) -> pd.DataFrame:
+def nsw_lalonde(seed: int = 42, simulated: bool = True) -> pd.DataFrame:
+    """LaLonde NSW data — simulated replica or real MatchIt extract.
+
+    Parameters
+    ----------
+    seed : int, default 42
+        RNG seed for the simulated replica (ignored when ``simulated=False``).
+    simulated : bool, default True
+        If True, return a deterministic simulated NSW experimental
+        subset (185 + 260 = 445 rows) calibrated so naive OLS
+        recovers the Dehejia-Wahba experimental ATT of about $1,794.
+        If False, load the real ``MatchIt::lalonde`` extract bundled
+        in ``statspai/datasets/data/lalonde_matchit.csv`` — the DW NSW
+        treated cohort (185) plus a 429-unit PSID-1 subset for
+        observational comparisons (n=614 total, with race factor
+        already split into ``black`` and ``hispanic`` indicators).
+
+    Notes
+    -----
+    The bundled real data is ``MatchIt::lalonde`` (n=614), NOT the
+    larger DW (1999) NSW + PSID-1 sample (n=2,675).  On this smaller
+    subset, naive OLS gives ATT roughly -$635 (less negative than DW
+    Table 3's headline -$8,498, which uses the full PSID-1).  For
+    the headline naive-bias demonstration, use the simulated
+    ``nsw_dw()`` panel instead.
+
+    Simulated replica calibration
+    -----------------------------
+    """
+    if not simulated:
+        df = _load_bundled_csv("lalonde_matchit.csv")
+        df.attrs['paper'] = (
+            "Dehejia, R. & Wahba, S. (1999). Causal Effects in "
+            "Nonexperimental Studies: Reevaluating the Evaluation of "
+            "Training Programs."
+        )
+        df.attrs['data_source'] = 'real'
+        df.attrs['simulated'] = False
+        df.attrs['source_origin'] = (
+            "R MatchIt::lalonde (n=614): 185 NSW treated + 429 PSID-1 "
+            "controls.  race factor split into black + hispanic dummies."
+        )
+        # StatsPAI-pinned values on this real extract.
+        df.attrs['statspai_pinned_naive_ols_att'] = -635.0
+        df.attrs['statspai_pinned_adj_ols_att']   = 1548.2
+        df.attrs['statspai_pinned_psm_att']       = 2012.5
+        df.attrs['published_dehejia_wahba_psm']   = 1794
+        df.attrs['notes'] = (
+            "Real MatchIt::lalonde extract (n=614). Naive OLS recovers "
+            "-$635 because PSID-1 is truncated to 429 controls; "
+            "covariate-adjusted OLS recovers $1,548 and 1:1 NN PSM "
+            "recovers ~$2,012, both close to the DW (1999) Table 4 "
+            "experimental benchmark of $1,794."
+        )
+        return df
+    return _nsw_lalonde_simulated(seed)
+
+
+def _nsw_lalonde_simulated(seed: int = 42) -> pd.DataFrame:
     """Simulated replica of the NSW experimental subset (185 treated + 260
     control).
 
@@ -451,25 +509,65 @@ def _re78(age, educ, black, hisp, married, re74, re75, treat):
 # Lee (2008) — US Senate RD
 # ---------------------------------------------------------------------------
 
-def lee_2008_senate(seed: int = 42) -> pd.DataFrame:
-    """Simulated replica of Lee (2008) US Senate RD.
+def lee_2008_senate(seed: int = 42, simulated: bool = True) -> pd.DataFrame:
+    """Lee (2008) US Senate RD — simulated replica or real extract.
 
-    Running variable: Democratic vote margin in election t (continuous,
-    centred at 0).  Treatment: Democratic win (margin >= 0).  Outcome:
-    Democratic vote share in election t+1.
+    Parameters
+    ----------
+    seed : int, default 42
+        RNG seed for the simulated DGP (ignored when ``simulated=False``).
+    simulated : bool, default True
+        If True, return a deterministic simulated panel (n=6558,
+        ``voteshare_next, margin, win``) on a 0-1 vote-share scale,
+        calibrated to a 0.08 jump at the cutoff.
+        If False, load the real ``rdrobust::rdrobust_RDsenate`` extract
+        (n=1390, ``x, y`` where ``y`` is vote share in **percent
+        points** 0-100 and ``x`` is the lagged Democratic margin).
 
-    Published incumbent-advantage estimate: ≈ 0.08 voteshare points
-    (Lee 2008, Table 2).
+    Notes
+    -----
+    The real-data branch lets you reproduce Lee (2008) Table 1 /
+    CCT (2014) Table 4 numbers exactly.  StatsPAI's
+    ``sp.rdrobust(df, y='y', x='x', c=0, kernel='triangular',
+    bwselect='cct')`` recovers Conventional ≈ 7.41 and Robust ≈ 7.51
+    on this dataset (paper headline ≈ 7.99).
 
     Returns
     -------
-    pd.DataFrame with columns: voteshare_next, margin, win.
+    pd.DataFrame
+        Simulated columns: ``voteshare_next, margin, win`` (0-1 scale).
+        Real columns: ``x, y`` (running variable; vote share 0-100).
 
     References
     ----------
     Lee, D. (2008). Randomized experiments from non-random selection in
     U.S. House elections. Journal of Econometrics 142, 675-697. [@lee2008randomized]
+    Calonico, S., Cattaneo, M.D. & Titiunik, R. (2014). Robust
+    nonparametric confidence intervals for regression-discontinuity
+    designs. Econometrica 82(6), 2295-2326. [@calonico2014robust]
     """
+    if not simulated:
+        df = _load_bundled_csv("lee_2008_senate.csv")
+        df.attrs['paper'] = (
+            "Lee, D. (2008). Randomized experiments from non-random "
+            "selection in U.S. House elections."
+        )
+        df.attrs['data_source'] = 'real'
+        df.attrs['simulated'] = False
+        df.attrs['source_origin'] = (
+            "R rdrobust::rdrobust_RDsenate (n=1390): lagged Democratic "
+            "vote margin (x) and current Democratic vote share (y, "
+            "percent points 0-100)."
+        )
+        df.attrs['statspai_pinned_conv_estimate_cct_bw'] = 7.414
+        df.attrs['statspai_pinned_robust_estimate_cct_bw'] = 7.507
+        df.attrs['published_lee2008_table1'] = 7.99
+        df.attrs['notes'] = (
+            "Real Lee Senate RD panel (n=1390).  Use kernel='triangular' "
+            "and bwselect='cct' for R-parity with rdrobust."
+        )
+        return df
+
     rng = np.random.default_rng(seed)
     n = 6558
     margin = rng.normal(0, 0.25, n)