refresh(X, y): fold new rows into a fitted model (#131, slice 1)

CHANGELOG.md

@@ -4,6 +4,21 @@ All notable changes to ChimeraBoost are documented here.
 The format follows [Keep a Changelog](https://keepachangelog.com/).
 
 ## [Unreleased]
+### Added
+- **`store_training_data` and `refresh(X, y)`** on the regressor and the
+  binary classifier (#131, first slice). With `store_training_data=True` a
+  model keeps its training rows in compact form (bins for plain numeric
+  columns, raw values only for the columns that feed cross features,
+  category codes), and `refresh` folds in new rows by replaying every
+  tree's splits on the old and new rows together and refitting only the
+  leaf values. It is the same replay the default full-data refit uses,
+  with the tree count, learning rate and splits pinned; refreshing with no
+  new rows reproduces the model bit for bit. Bagged models, multiclass,
+  `loss="Quantile"` and `random_effects=True` are not supported yet. The
+  default fit's results are unchanged. On a 3.3M-row regression, folding in
+  a day of new rows took 70 s against 232 s for a full refit, at the same
+  accuracy.
+
 ### Changed
 - **The quantile head retrains its winner on all rows by default.** Called
   without an `eval_set`, `ChimeraBoostQuantileRegressor` held back
@@ -49,6 +64,11 @@ The format follows [Keep a Changelog](https://keepachangelog.com/).
   untouched: the exact-output snapshot moves only one quantile
   configuration (183 of 186 pins identical). Record:
   `benchmarks/QUANTILE_PLAN.md`, Phase 2 (Q8, Q9).
+- **Fits with linear leaves are faster, with identical results.** The
+  per-leaf linear fit used to sum each leaf on one thread, and one leaf
+  often holds 40-60% of the rows. Large leaves now spread their sums over
+  several threads, each sum still adding the rows in the same order. A
+  500k-row fit went from 15.9 to 13.4 s, and a refresh from 5.0 to 3.6 s.
 - **The quantile benchmark's NGBoost opponent now uses the RoNGBa
   settings** (Ren, Sun and Wu 2019; #163): trees of up to 31 leaves, a
   learning rate of 0.04 and at most 500 rounds, in place of NGBoost's stock

benchmarks/CAMPAIGN_PLAN.md

@@ -1277,6 +1277,11 @@ PR #170 is PARKED OPEN (not merged, not closed); issue #131 stays open.
 The loop moves on to the other issues; if he closes #170, the kernel
 speedup (435e866, bit-identical, 1.19x on linear-leaf fits) is worth
 salvaging as its own PR, since it helps every default fit.
+2026-09-26, the maintainer: "let's move forward with 170, but clean up
+the merge conflicts then i'll merge the PR". Un-parked: main merged into
+the branch; the conflicts were CHANGELOG (entries on both sides, all
+kept), this file and `REFRESH_PLAN.md` (log lines on both sides, all
+kept); no code overlapped. Awaiting his merge.
 
 #### I065 2026-09-24 issue #81 (research cascade: dead self-test anchor, stale `ideas.py` flags; BENCH tooling + test, pre-registered)
 why now: the focus rule's second issue. PR #168 merged (3e02ab1), #84

benchmarks/REFRESH_PLAN.md

@@ -131,3 +131,8 @@ Still open, for their slices:
   implemented"). Slice 1 is PARKED; slices 2-7 are not started. If #170 is
   closed, salvage the bit-identical replay-kernel speedup (435e866) as its
   own PR: it speeds up every linear-leaf fit.
+- 2026-09-26: un-parked. The maintainer: "let's move forward with 170,
+  but clean up the merge conflicts then i'll merge the PR". Main merged
+  into the branch (conflicts only in CHANGELOG and two plan files; no
+  code overlapped), so the salvage note above no longer applies. Slices
+  2-7 are not started.

chimeraboost/booster.py

@@ -824,6 +824,31 @@ def __init__(self, loss="RMSE", loss_kwargs=None, **kw):
         self.loss_name = loss
         self.loss_kwargs = loss_kwargs or {}
 
+    def replay_kwargs(self):
+        """Constructor kwargs that rebuild an equivalent booster.
+
+        Every ``_BaseBooster.__init__`` parameter (each stored under its own
+        name, so read off ``inspect.signature`` rather than a hand-kept
+        list) plus ``loss`` (from ``loss_name``) and ``loss_kwargs``; lists
+        and dicts are copied.
+        """
+        import inspect
+
+        kw = {}
+        params = inspect.signature(_BaseBooster.__init__).parameters
+        for name in params:
+            if name == "self":
+                continue
+            v = getattr(self, name)
+            if isinstance(v, list):
+                v = list(v)
+            elif isinstance(v, dict):
+                v = dict(v)
+            kw[name] = v
+        kw["loss"] = self.loss_name
+        kw["loss_kwargs"] = dict(self.loss_kwargs)
+        return kw
+
     def _fit_impl(self, X, y, cat_features=None, eval_set=None,
                   sample_weight=None, callbacks=None, prep_cache=None):
         """Fit the additive model.