File size: 25,317 Bytes
8b97eb8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
"""Single-file simulator runtime for the v3 benchmark.

Task simulator directories contain `state.joblib` and `sample.csv`. The public
entry point is `load(sim_dir)`, which selects the Type I or Type II runtime
from the saved schema.

Some `state.joblib` files pickle `ResidualBootstrap` under the historical
`data_aug.sim2.bootstrap` path. This module registers that path as an alias to
itself so those states unpickle without a `data_aug` dependency.
"""
from __future__ import annotations

import importlib.util
import math
import sys
import types
from dataclasses import dataclass
from functools import lru_cache
from pathlib import Path
from typing import Any, Callable, Dict, Optional

import joblib
import numpy as np
import pandas as pd


DEFAULT_MAX_ROWS = 50
DEFAULT_OVERSAMPLE = 20
HARD_LIMIT = 5000
FLOOR_FOR_LOG = 1e-12


def safe_float(value: Any) -> Optional[float]:
    """Convert finite numeric values to plain Python floats for JSON-like output."""
    try:
        number = float(value)
    except (TypeError, ValueError):
        return None
    return None if math.isnan(number) or math.isinf(number) else number


def module_from_source(source: str, module_name: str, source_name: str):
    module = types.ModuleType(module_name)
    exec(compile(source, source_name, "exec"), module.__dict__)
    return module


def load_formula_file(task_dir: Path, formula_id: str, module_name: str):
    formula_path = task_dir / "formulas" / f"{formula_id}.py"
    spec = importlib.util.spec_from_file_location(module_name, formula_path)
    if spec is None or spec.loader is None:
        raise ImportError(f"could not load formula module from {formula_path}")
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)
    return module


def group_id_key(value: Any) -> str:
    """Normalize user-supplied group ids to the string keys stored in state."""
    try:
        number = float(value)
    except (TypeError, ValueError):
        return str(value)
    if math.isfinite(number) and number.is_integer():
        return str(int(number))
    return str(value)


def group_id_value(key: str) -> Any:
    """Return a compact JSON-like group id value for output rows."""
    try:
        return int(key)
    except ValueError:
        return key


def _transform_axes(X: np.ndarray, x_axes: list[str]) -> np.ndarray:
    """Apply per-column log10 to 'log10' axes; leave 'linear' columns alone."""
    Xt = np.array(X, dtype=float, copy=True)
    for j, ax in enumerate(x_axes):
        if ax == "log10":
            Xt[:, j] = np.log10(np.maximum(Xt[:, j], FLOOR_FOR_LOG))
    return Xt


class ResidualBootstrap:
    """k-NN residual resampler over real data.

    For a query point X*, draw a residual from the empirical residuals of the
    nearest real data points in standardized input space.
    """

    def __init__(self, x_axes: list[str], k_frac: float = 0.15,
                 k_min: int = 8, k_max: int = 64):
        self.x_axes = list(x_axes)
        self.k_frac = float(k_frac)
        self.k_min = int(k_min)
        self.k_max = int(k_max)
        self._Xs: np.ndarray | None = None
        self._res: np.ndarray | None = None
        self._mean: np.ndarray | None = None
        self._std: np.ndarray | None = None
        self._tree = None

    def fit(self, X_real: np.ndarray, residuals: np.ndarray) -> "ResidualBootstrap":
        X_real = np.atleast_2d(np.asarray(X_real, dtype=float))
        res = np.asarray(residuals, dtype=float).ravel()
        finite = np.isfinite(res) & np.all(np.isfinite(X_real), axis=1)
        X_real, res = X_real[finite], res[finite]
        if len(res) == 0:
            raise ValueError("no finite (X, residual) pairs to fit bootstrap")
        Xt = _transform_axes(X_real, self.x_axes)
        self._mean = Xt.mean(axis=0)
        self._std = Xt.std(axis=0)
        self._std[self._std == 0] = 1.0
        self._Xs = (Xt - self._mean) / self._std
        self._res = res
        self._tree = None
        return self

    @property
    def n(self) -> int:
        return 0 if self._res is None else len(self._res)

    def _k(self) -> int:
        return int(np.clip(round(self.n * self.k_frac),
                           min(self.k_min, self.n), min(self.k_max, self.n)))

    def _ensure_tree(self):
        if self._tree is None and self.n >= self.k_min:
            try:
                from scipy.spatial import cKDTree
                self._tree = cKDTree(self._Xs)
            except Exception:
                self._tree = "brute"
        return self._tree

    def _neighbour_idx(self, Xq_s: np.ndarray, k: int) -> np.ndarray:
        """Return (m, k) neighbour indices for standardized queries Xq_s."""
        tree = self._ensure_tree()
        if tree is not None and tree != "brute":
            _, idx = tree.query(Xq_s, k=k)
            return np.atleast_2d(idx).reshape(len(Xq_s), k)
        out = np.empty((len(Xq_s), k), dtype=int)
        for i, q in enumerate(Xq_s):
            d = np.sum((self._Xs - q) ** 2, axis=1)
            out[i] = np.argpartition(d, min(k, self.n) - 1)[:k]
        return out

    def sample(self, X_query: np.ndarray, rng: np.random.Generator) -> np.ndarray:
        """Draw one residual per query row with replacement from its k-NN."""
        X_query = np.atleast_2d(np.asarray(X_query, dtype=float))
        m = len(X_query)
        if self.n < self.k_min:
            return self._res[rng.integers(0, self.n, size=m)]
        Xt = _transform_axes(X_query, self.x_axes)
        Xq_s = (Xt - self._mean) / self._std
        k = self._k()
        nbr = self._neighbour_idx(Xq_s, k)
        pick = rng.integers(0, k, size=m)
        chosen = nbr[np.arange(m), pick]
        return self._res[chosen]

    def predictive_interval(self, X_query: np.ndarray, lo: float = 0.1,
                            hi: float = 0.9) -> tuple[np.ndarray, np.ndarray]:
        """Per-query residual quantiles over the local neighbourhood."""
        X_query = np.atleast_2d(np.asarray(X_query, dtype=float))
        m = len(X_query)
        if self.n < self.k_min:
            ql, qh = np.quantile(self._res, [lo, hi])
            return np.full(m, ql), np.full(m, qh)
        Xt = _transform_axes(X_query, self.x_axes)
        Xq_s = (Xt - self._mean) / self._std
        nbr = self._neighbour_idx(Xq_s, self._k())
        local = self._res[nbr]
        return (np.quantile(local, lo, axis=1),
                np.quantile(local, hi, axis=1))

    def local_std(self, X_query: np.ndarray) -> np.ndarray:
        """Per-query residual std over the local neighbourhood."""
        X_query = np.atleast_2d(np.asarray(X_query, dtype=float))
        m = len(X_query)
        if self.n < self.k_min:
            return np.full(m, float(np.std(self._res)))
        Xt = _transform_axes(X_query, self.x_axes)
        Xq_s = (Xt - self._mean) / self._std
        nbr = self._neighbour_idx(Xq_s, self._k())
        return np.std(self._res[nbr], axis=1)

    def local_mean(self, X_query: np.ndarray) -> np.ndarray:
        """Per-query mean residual over the local neighbourhood."""
        X_query = np.atleast_2d(np.asarray(X_query, dtype=float))
        m = len(X_query)
        if self.n < self.k_min:
            return np.full(m, float(np.mean(self._res)))
        Xt = _transform_axes(X_query, self.x_axes)
        Xq_s = (Xt - self._mean) / self._std
        nbr = self._neighbour_idx(Xq_s, self._k())
        return np.mean(self._res[nbr], axis=1)

    def __getstate__(self):
        state = self.__dict__.copy()
        state["_tree"] = None
        return state


def _register_pickle_alias() -> None:
    data_aug = sys.modules.get("data_aug")
    if data_aug is None:
        data_aug = types.ModuleType("data_aug")
        data_aug.__path__ = []
        sys.modules["data_aug"] = data_aug

    sim2 = sys.modules.get("data_aug.sim2")
    if sim2 is None:
        sim2 = types.ModuleType("data_aug.sim2")
        sim2.__path__ = []
        sys.modules["data_aug.sim2"] = sim2

    sys.modules.setdefault("data_aug.sim2.bootstrap", sys.modules[__name__])
    setattr(data_aug, "sim2", sim2)
    setattr(sim2, "bootstrap", sys.modules["data_aug.sim2.bootstrap"])


_register_pickle_alias()


@dataclass
class TypeISimulatorBundle:
    fetch_where: Callable
    fetch_data: Callable
    load_sample: Callable
    tool_description: Callable
    formula_info: Callable
    budget_status: Callable


def _load_typeI_formula(task_dir: Path, formula_id: str):
    return load_formula_file(
        task_dir, formula_id, f"_sim2rt_{task_dir.name}_{formula_id}"
    )


def _build_typeI_simulator(wrapper_path: str | Path) -> TypeISimulatorBundle:
    sim_dir = Path(wrapper_path).resolve().parent
    formula_id = sim_dir.name
    task_dir = sim_dir.parent.parent
    state_path = sim_dir / "state.joblib"
    sample_csv = sim_dir / "sample.csv"

    @lru_cache(maxsize=1)
    def _state() -> Dict[str, Any]:
        return joblib.load(state_path)

    @lru_cache(maxsize=1)
    def _F():
        src = _state().get("formula_source")
        if src:
            return module_from_source(
                src,
                f"_sim2rt_src_{formula_id}",
                f"<formula_source:{formula_id}>",
            )
        return _load_typeI_formula(task_dir, formula_id)

    @lru_cache(maxsize=1)
    def _real_df() -> pd.DataFrame:
        return pd.DataFrame(_state()["real_rows"])

    budget = {"used": 0}

    def _remaining() -> int:
        return max(0, int(_state()["fetch_budget_rows"]) - budget["used"])

    def budget_status() -> Dict[str, Any]:
        s = _state()
        return {"fetch_budget_rows": int(s["fetch_budget_rows"]),
                "used": budget["used"], "remaining": _remaining()}

    def _generate(X_used: np.ndarray, rng) -> np.ndarray:
        s = _state()
        F = _F()
        space = s["residual_space"]
        ns = float(s["noise_scale"])
        law = dict(s["law_constants"])
        y_clean = np.asarray(F.predict(X_used, **law))
        if np.iscomplexobj(y_clean):
            y_clean = np.real(y_clean)
        y_clean = y_clean.astype(float)
        r = s["bootstrap"].sample(X_used, rng) * ns if ns != 0 else np.zeros(len(X_used))
        if space == "log":
            floor = s.get("p_min_floor") or FLOOR_FOR_LOG
            lc = np.log10(np.maximum(y_clean, floor))
            return np.maximum(10.0 ** (lc + r), floor)
        return y_clean + r

    def load_sample() -> pd.DataFrame:
        return pd.read_csv(sample_csv)

    def fetch_data(seed: Optional[int] = None, n_samples: int = 1,
                   **input_vals) -> Dict[str, Any]:
        s = _state()
        used = s["used_inputs"]
        support = s["support"]
        target = s["target"]
        unknown = [k for k in input_vals if k not in used]
        if unknown:
            return {"error": f"unknown input(s) {unknown}; valid inputs: {used}"}

        arrays = {}
        for c in used:
            v = input_vals.get(c)
            if v is None:
                return {"error": f"missing input {c!r}; pass {c}=[...]"}
            try:
                arrays[c] = np.asarray(v, dtype=float).ravel()
            except Exception as e:
                return {"error": f"could not parse {c}: {type(e).__name__}: {e}"}

        n_pts = len(arrays[used[0]])
        if any(len(a) != n_pts for a in arrays.values()):
            return {"error": "all input arrays must have equal length"}
        if n_pts == 0:
            return {"source": "synthetic", "rows": [], "n_returned": 0, "n_clipped": 0,
                    "budget": budget_status()}

        n_per = max(1, int(n_samples))
        total = n_pts * n_per
        if _remaining() <= 0:
            return {"error": "fetch budget exhausted", "budget": budget_status()}
        if total > _remaining():
            return {"error": f"request ({total} rows) exceeds remaining fetch "
                             f"budget ({_remaining()})", "budget": budget_status()}

        n_clipped = 0
        for c in used:
            lo, hi = support[c]["min"], support[c]["max"]
            a = arrays[c]
            n_clipped += int(((a < lo) | (a > hi)).sum())
            arrays[c] = np.clip(a, lo, hi)

        X = np.repeat(np.stack([arrays[c] for c in used], axis=1), n_per, axis=0)
        rng = np.random.default_rng(seed)
        y = _generate(X, rng)
        budget["used"] += total

        rows = [{target: safe_float(y[i]),
                 **{c: safe_float(X[i, j]) for j, c in enumerate(used)}}
                for i in range(len(X))]
        return {"source": "synthetic", "n_returned": len(rows), "n_clipped": n_clipped,
                "rows": rows, "budget": budget_status()}

    def fetch_where(query: str, limit: Optional[int] = None,
                    max_rows_per_request: int = DEFAULT_MAX_ROWS,
                    seed: Optional[int] = None,
                    oversample: int = DEFAULT_OVERSAMPLE) -> Dict[str, Any]:
        s = _state()
        used = s["used_inputs"]
        target = s["target"]
        all_in = s["all_input_cols"]
        if not isinstance(query, str) or not query.strip():
            return {"error": "missing 'query' expression"}

        cap = max_rows_per_request if limit is None else min(int(limit), max_rows_per_request)
        cap = max(1, min(cap, HARD_LIMIT))
        if _remaining() <= 0:
            return {"error": "fetch budget exhausted", "budget": budget_status()}
        cap = min(cap, _remaining())
        n_draw = min(cap * max(1, int(oversample)), HARD_LIMIT, len(_real_df()) * 50 or HARD_LIMIT)

        rng = np.random.default_rng(seed)
        real = _real_df()
        idx = rng.integers(0, len(real), size=n_draw)
        pool = real.iloc[idx].reset_index(drop=True).copy()
        X_used = pool[used].to_numpy(float)
        pool[target] = _generate(X_used, rng)
        pool = pool[[target] + [c for c in all_in]]

        try:
            matched = pool.query(query)
        except Exception as e:
            return {"error": f"query failed: {type(e).__name__}: {e}",
                    "available_columns": list(pool.columns)}

        n_match = int(len(matched))
        matched = matched.iloc[:cap]
        budget["used"] += int(len(matched))

        out = {"where_evaluated": query, "n_matching_in_pool": n_match,
               "n_returned": int(len(matched)),
               target: [safe_float(v) for v in matched[target]],
               **{c: [safe_float(v) for v in matched[c]] for c in all_in},
               "budget": budget_status()}
        if n_match > cap:
            out["_truncated_to"] = cap
        return out

    def formula_info() -> Dict[str, Any]:
        s = _state()
        return {"formula_id": s.get("formula_id"), "target": s["target"],
                "metric": s["metric"], "used_inputs": s["used_inputs"],
                "formula_source": s.get("formula_source"),
                "formula_doc": s.get("formula_doc"),
                "equation_loc": s.get("equation_loc"),
                "paper_ref": s.get("paper_ref"),
                "law_constants": s.get("law_constants"),
                "phase": "typeI"}

    def tool_description() -> str:
        s = _state()
        used = s["used_inputs"]
        support = s["support"]
        target = s["target"]
        lines = [
            "Synthetic data oracle for this task. Two fetch tools, both metered "
            f"against a finite budget of {int(s['fetch_budget_rows'])} total rows:",
            "  fetch_data(**inputs, n_samples=1, seed=None) — evaluate at explicit "
            "input points (pass each input as a list).",
            "  fetch_where(query, limit=20) — query a fresh synthetic pool "
            f"(pandas df.query over [{', '.join(used + [target])}]).",
            "  load_sample() — a fixed generator sample (free, does not cost budget).",
            "Queryable input ranges:",
        ]
        for c in used:
            lines.append(f"  {c} ∈ [{support[c]['min']:.4g}, {support[c]['max']:.4g}]")
        lines.append("Each fetch is a FRESH draw: formula(X) + a residual resampled "
                     "from nearby real data (heteroscedastic, non-Gaussian).")
        return "\n".join(lines)

    return TypeISimulatorBundle(
        fetch_where=fetch_where,
        fetch_data=fetch_data,
        load_sample=load_sample,
        tool_description=tool_description,
        formula_info=formula_info,
        budget_status=budget_status,
    )


@dataclass
class TypeIISimulatorBundle:
    list_groups: Callable
    fetch_data: Callable
    fetch_where: Callable
    load_sample: Callable
    budget_status: Callable
    tool_description: Callable
    form_info: Callable


def _load_typeII_form(task_dir: Path, form_id: str):
    return load_formula_file(task_dir, form_id, f"_t2rt_{form_id}")


def _build_typeII_simulator(wrapper_path: str | Path) -> TypeIISimulatorBundle:
    sim_dir = Path(wrapper_path).resolve().parent
    form_id = sim_dir.name
    task_dir = sim_dir.parent.parent
    state_path = sim_dir / "state.joblib"
    sample_csv = sim_dir / "sample.csv"

    @lru_cache(maxsize=1)
    def _state():
        return joblib.load(state_path)

    @lru_cache(maxsize=1)
    def _F():
        src = _state().get("formula_source")
        if src:
            return module_from_source(
                src,
                f"_t2rt_src_{form_id}",
                f"<formula_source:{form_id}>",
            )
        return _load_typeII_form(task_dir, form_id)

    budget = {"used": 0}

    def _remaining():
        return max(0, int(_state()["fetch_budget_rows"]) - budget["used"])

    def budget_status():
        s = _state()
        return {"fetch_budget_rows": int(s["fetch_budget_rows"]), "used": budget["used"],
                "remaining": _remaining()}

    def _gen(gid: str, X_used: np.ndarray, rng) -> np.ndarray:
        s = _state()
        F = _F()
        group = s["groups"][gid]
        space = s["residual_space"]
        ns = float(group.get("noise_scale", s["noise_scale"]))
        y = np.asarray(F.predict(X_used, **group["params"]), float)
        r = group["bootstrap"].sample(X_used, rng) * ns
        if space == "log":
            floor = s.get("p_min_floor") or FLOOR_FOR_LOG
            return np.maximum(10.0 ** (np.log10(np.maximum(y, floor)) + r), floor)
        return y + r

    def list_groups():
        s = _state()
        return {"n_groups": len(s["groups"]), "used_inputs": s["used_inputs"],
                "groups": {gid: {"n": g["n"], "support": g["support"]}
                           for gid, g in s["groups"].items()}}

    def load_sample() -> pd.DataFrame:
        return pd.read_csv(sample_csv)

    def fetch_data(group_id=None, seed: Optional[int] = None,
                   n_samples: int = 1, **inputs):
        s = _state()
        used = s["used_inputs"]
        target = s["target"]
        unknown = [k for k in inputs if k not in used]
        if unknown:
            return {"error": f"unknown input(s) {unknown}; valid inputs: {used}"}

        arrays = {}
        for c in used:
            v = inputs.get(c)
            if v is None:
                return {"error": f"missing input {c!r}; pass {c}=[...]"}
            arrays[c] = np.asarray(v, float).ravel()

        n_pts = len(arrays[used[0]])
        if any(len(a) != n_pts for a in arrays.values()):
            return {"error": "all input arrays must have equal length"}
        if n_pts == 0:
            return {"group_id": group_id, "n_returned": 0, "n_clipped": 0,
                    "rows": [], "budget": budget_status()}

        if isinstance(group_id, (list, tuple, np.ndarray, pd.Series)):
            group_ids = np.asarray(group_id, dtype=object).ravel()
            if len(group_ids) != n_pts:
                return {"error": "group_id must be scalar or have one value per input point"}
            gid_keys = [group_id_key(g) for g in group_ids]
        else:
            gid = group_id_key(group_id)
            gid_keys = [gid] * n_pts

        unknown_groups = sorted({g for g in gid_keys if g not in s["groups"]})
        if unknown_groups:
            return {"error": f"unknown group_id {unknown_groups}; call list_groups()"}

        n_per = max(1, int(n_samples))
        total = n_pts * n_per
        if total > _remaining():
            return {"error": f"request ({total}) exceeds remaining budget ({_remaining()})",
                    "budget": budget_status()}

        n_clipped = 0
        clipped = {c: arrays[c].copy() for c in used}
        for i, gid in enumerate(gid_keys):
            support = s["groups"][gid]["support"]
            for c in used:
                lo, hi = support[c]["min"], support[c]["max"]
                value = clipped[c][i]
                if value < lo or value > hi:
                    n_clipped += 1
                    clipped[c][i] = np.clip(value, lo, hi)

        X = np.repeat(np.stack([clipped[c] for c in used], axis=1), n_per, axis=0)
        gids_repeated = np.repeat(np.asarray(gid_keys, dtype=object), n_per)
        y = np.empty(len(X), dtype=float)
        rng = np.random.default_rng(seed)
        for gid in sorted(set(gid_keys)):
            mask = gids_repeated == gid
            y[mask] = _gen(gid, X[mask], rng)
        budget["used"] += total

        rows = [{target: safe_float(y[i]),
                 **{c: safe_float(X[i, j]) for j, c in enumerate(used)},
                 "group_id": group_id_value(str(gids_repeated[i]))}
                for i in range(len(X))]
        return {"group_id": group_id, "n_returned": len(rows), "n_clipped": n_clipped,
                "rows": rows, "budget": budget_status()}

    def fetch_where(group_id=None, query: str = "", limit: Optional[int] = None,
                    max_rows_per_request: int = DEFAULT_MAX_ROWS,
                    seed: Optional[int] = None, oversample: int = 20):
        s = _state()
        used = s["used_inputs"]
        target = s["target"]
        all_in = s["all_input_cols"]
        gid = str(group_id)
        if gid not in s["groups"]:
            return {"error": f"unknown group_id {group_id!r}; call list_groups()"}
        if not isinstance(query, str) or not query.strip():
            return {"error": "missing 'query'"}

        cap = max(1, min(max_rows_per_request if limit is None
                         else min(int(limit), max_rows_per_request),
                         HARD_LIMIT, _remaining()))
        if cap <= 0:
            return {"error": "fetch budget exhausted", "budget": budget_status()}

        real = pd.DataFrame(s["groups"][gid]["real_rows"])
        n_draw = min(cap * max(1, int(oversample)), HARD_LIMIT, max(len(real) * 50, HARD_LIMIT))
        rng = np.random.default_rng(seed)
        pool = real.iloc[rng.integers(0, len(real), size=n_draw)].reset_index(drop=True).copy()
        pool[target] = _gen(gid, pool[used].to_numpy(float), rng)
        pool = pool[[target] + list(all_in)]

        try:
            matched = pool.query(query).iloc[:cap]
        except Exception as e:
            return {"error": f"query failed: {type(e).__name__}: {e}", "columns": list(pool.columns)}

        budget["used"] += int(len(matched))
        return {"group_id": gid, "where": query, "n_returned": int(len(matched)),
                target: [safe_float(v) for v in matched[target]],
                **{c: [safe_float(v) for v in matched[c]] for c in all_in},
                "budget": budget_status()}

    def form_info():
        s = _state()
        return {"form_id": s.get("form_id"), "formula_source": s.get("formula_source"),
                "formula_doc": s.get("formula_doc"), "equation_loc": s.get("equation_loc"),
                "paper_ref": s.get("paper_ref"),
                "group_params": {gid: g["params"] for gid, g in s["groups"].items()}}

    def tool_description():
        s = _state()
        return ("Multi-group synthetic oracle: recover the UNIVERSAL form behind all groups.\n"
                f"  {len(s['groups'])} groups, inputs {s['used_inputs']}, "
                f"budget {int(s['fetch_budget_rows'])} rows.\n"
                "  list_groups() — group ids + per-group input ranges (free).\n"
                "  fetch_data(group_id, **inputs, n_samples=1) — eval that group at points.\n"
                "  fetch_where(group_id, query, limit=20) — query a fresh pool for that group.\n"
                "  load_sample() — fixed noisy sample across all groups (free).\n"
                "Each group has its OWN local params + heteroscedastic noise; the FORM is shared.")

    return TypeIISimulatorBundle(
        list_groups,
        fetch_data,
        fetch_where,
        load_sample,
        budget_status,
        tool_description,
        form_info,
    )


def load(sim_dir: str | Path):
    """Build a simulator bundle from a task's `simulator/` directory."""
    sim_dir = Path(sim_dir).resolve()
    state_path = sim_dir / "state.joblib"
    if not state_path.exists():
        raise FileNotFoundError(f"no state.joblib under {sim_dir}")

    schema = str(joblib.load(state_path).get("schema_version", ""))
    anchor = sim_dir / "_wrapper.py"
    if "typeII" in schema:
        return _build_typeII_simulator(anchor)
    return _build_typeI_simulator(anchor)


__all__ = ["load", "ResidualBootstrap"]