File size: 25,488 Bytes
d922ac9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Planet-independent terrain feature machinery for the latband5km-v1 grid.

Everything here is pure array work plus Parquet assembly: it knows nothing
about a particular planet, raster or storage backend. The planet-specific
modules (:mod:`mars_terrain`, :mod:`earth_terrain`) own source discovery, read
windows, ground-spacing geometry and quality semantics; they call into this
module for the parts that must stay identical across planets so that the two
feature tables mean the same thing and join on ``region_id``.

What lives here:

* block averaging east-west (and north-south), with exact native pixel
  accounting;
* the +-180 degree wrap halo;
* Horn 3x3 slope with caller-supplied true ground spacing;
* the 1 km local-relief window;
* grouped per-cell reductions (percentiles, extremes, detrended roughness);
* the ``terrain-v0.2`` feature definitions and the per-cell reduction that
  produces them;
* advisory quality flags, Parquet table assembly, atomic writes and the
  band-group bookkeeping shared by both drivers.

The feature contract is versioned by ``FEATURE_VERSION``. Changing any number
produced here changes both planets at once and requires a version bump.
"""
from __future__ import annotations

import hashlib
import json
import os
import math
import uuid
from dataclasses import dataclass
from pathlib import Path

import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq

from planetary_grid import PLANETS, VERSION as GRID_VERSION

FEATURE_VERSION = 'terrain-v0.2'
ANALYSIS_SPACING_M = 200.0
BANDS_PER_GROUP = 100
ROW_HALO = 2
BLOCK_HALO = 2
RELIEF_WINDOW = 5                      # 5 x 5 analysis pixels ~ 1 km
SLOPE_THRESHOLDS = (2, 5, 10, 20, 30)  # degrees, for slope_frac_gt_*
POLAR_DISTORTION_LAT = 88.0
LOW_COVERAGE_FRACTION = 0.8

ELEV_PERCENTILES = (5, 95)
SLOPE_PERCENTILES = (50, 90, 99)

FLOAT_FEATURES = [
    'n_samples_200m', 'n_valid_samples_200m', 'valid_frac_200m',
    'elev_min', 'elev_max', 'elev_mean', 'elev_p05', 'elev_p95',
    'elev_range', 'elev_range_robust',
    'slope_mean', 'slope_median', 'slope_std', 'slope_p90', 'slope_p99',
    *[f'slope_frac_gt_{t}' for t in SLOPE_THRESHOLDS],
    'roughness_detrended_std', 'local_relief_1km', 'ew_samples_averaged',
]
COUNT_FEATURES = ['n_native_samples', 'n_valid_native_samples']


# --------------------------------------------------------------------------
# provenance
# --------------------------------------------------------------------------

def code_version(*paths: Path | str) -> str:
    """SHA-256 over the named source files, in the order given.

    Every row names the code that made it. A single path reproduces the
    original single-module digest exactly; passing several files lets a driver
    cover itself and this module together.
    """
    digest = hashlib.sha256()
    for path in paths:
        digest.update(Path(path).read_bytes())
    return digest.hexdigest()


# --------------------------------------------------------------------------
# block factors
# --------------------------------------------------------------------------

def block_factor(ratio: float, max_factor: int | None = None) -> int:
    """max(1, round(ratio)) with deterministic half-up rounding.

    Half-way values round up via ``floor(x + 0.5)`` rather than to even, so the
    factor is reproducible across platforms. ``ratio`` is target spacing over
    source spacing; a non-finite or non-positive ratio means the source pixel
    is degenerate (at a pole) and the whole row collapses to ``max_factor``.
    """
    if not math.isfinite(ratio) or ratio <= 0:
        return max_factor if max_factor else 1
    factor = max(1, math.floor(ratio + 0.5))
    return min(factor, max_factor) if max_factor else factor


def ew_block_factor(lat_deg: float, max_factor: int | None = None) -> int:
    """k = max(1, round(1/cos(lat))): pixels to average east-west for ~200 m ground.

    This is the square-pixel, simple-cylindrical case (Mars): the source pixel
    already spans ~200 m north-south, so only the 1/cos(lat) convergence of the
    meridians has to be undone.
    """
    cos_lat = math.cos(math.radians(lat_deg))
    if cos_lat <= 0:
        return max_factor if max_factor else 1
    return block_factor(1.0 / cos_lat, max_factor)


# --------------------------------------------------------------------------
# array stages
# --------------------------------------------------------------------------

def block_average_ew(z: np.ndarray, k: int) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    """NaN-aware mean of k adjacent east-west pixels.

    Returns the block means, the number of native pixels behind each block
    column (the final block is short when k does not divide the width), and the
    number of valid native pixels per block. Blocks partition the native
    columns exactly, so the native counts stay exact.
    """
    rows, width = z.shape
    if k < 1:
        raise ValueError('block factor must be >= 1')
    n_blocks = -(-width // k)
    pad = n_blocks * k - width
    valid = np.isfinite(z)
    filled = np.where(valid, z, 0.0).astype(np.float64)
    if pad:
        filled = np.concatenate([filled, np.zeros((rows, pad))], axis=1)
        valid = np.concatenate([valid, np.zeros((rows, pad), dtype=bool)], axis=1)
    total = filled.reshape(rows, n_blocks, k).sum(axis=2)
    count = valid.reshape(rows, n_blocks, k).sum(axis=2)
    with np.errstate(invalid='ignore'):
        mean = np.where(count > 0, total / np.maximum(count, 1), np.nan)
    source = np.full(n_blocks, k, dtype=np.int64)
    source[-1] = k - pad
    return mean, source, count.astype(np.int64)


def block_average(z: np.ndarray, row_starts: np.ndarray,
                  col_starts: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
    """NaN-aware mean over arbitrary rectangular blocks, plus the valid counts.

    ``row_starts`` and ``col_starts`` are the first index of each block; each
    block runs to the next start (the last to the end of the array), so blocks
    partition the input exactly and a short final block is allowed. Explicit
    starts, rather than a single factor, are what lets a band keep whole blocks
    on both sides of a halo whose own blocks must stay aligned.

    The mean is taken over all valid native pixels in the two-dimensional
    block at once, which is not the same as averaging rows and then columns
    whenever some pixels are NaN.
    """
    valid = np.isfinite(z)
    filled = np.where(valid, z, 0.0).astype(np.float64)
    total = np.add.reduceat(np.add.reduceat(filled, row_starts, axis=0), col_starts, axis=1)
    count = np.add.reduceat(np.add.reduceat(valid.astype(np.int64), row_starts, axis=0),
                            col_starts, axis=1)
    with np.errstate(invalid='ignore'):
        mean = np.where(count > 0, total / np.maximum(count, 1), np.nan)
    return mean, count


def block_starts(length: int, k: int, offset: int = 0) -> np.ndarray:
    """Block start indices for ``length`` items taken ``k`` at a time."""
    if k < 1:
        raise ValueError('block factor must be >= 1')
    return np.arange(offset, offset + length, k, dtype=np.int64)


def wrap_blocks(a: np.ndarray, halo: int = BLOCK_HALO) -> np.ndarray:
    """Pad east-west with the opposite edge so +-180 degrees is continuous."""
    if a.shape[1] < halo:
        raise ValueError('too few block columns to wrap')
    return np.concatenate([a[:, -halo:], a, a[:, :halo]], axis=1)


def horn_slope_deg(z: np.ndarray, dx_m: np.ndarray, dy_m: float | np.ndarray) -> np.ndarray:
    """Horn 3x3 slope in degrees, NaN wherever any stencil pixel is NaN.

    ``z`` is north-up and the result is smaller by one pixel on every side.
    ``dx_m`` is the east-west ground spacing, broadcastable to the output
    shape: half of the true separation between the stencil's outer columns.
    Passing it per column is what keeps the +-180 seam honest, because the
    published raster stops 0.46 pixels short of closing the longitude ring.
    ``dy_m`` is the matching north-south spacing, scalar or per row.
    """
    north, mid, south = z[:-2], z[1:-1], z[2:]
    dzdx = ((north[:, 2:] + 2 * mid[:, 2:] + south[:, 2:])
            - (north[:, :-2] + 2 * mid[:, :-2] + south[:, :-2])) / (8 * dx_m)
    dzdy = ((north[:, :-2] + 2 * north[:, 1:-1] + north[:, 2:])
            - (south[:, :-2] + 2 * south[:, 1:-1] + south[:, 2:])) / (8 * dy_m)
    slope = np.degrees(np.arctan(np.hypot(dzdx, dzdy)))
    return np.where(np.isfinite(mid[:, 1:-1]), slope, np.nan)


def _window_extreme(a: np.ndarray, size: int, op) -> np.ndarray:
    """Separable moving max/min; NaN propagates, so partial windows are NaN."""
    for axis in (0, 1):
        length = a.shape[axis]
        parts = []
        for shift in range(size):
            index = [slice(None)] * a.ndim
            index[axis] = slice(shift, length - (size - 1 - shift))
            parts.append(a[tuple(index)])
        a = op.reduce(parts)
    return a


def local_relief(z: np.ndarray, size: int = RELIEF_WINDOW) -> np.ndarray:
    """max - min over a size x size window; shrinks by (size-1) on each axis."""
    return _window_extreme(z, size, np.maximum) - _window_extreme(z, size, np.minimum)


# --------------------------------------------------------------------------
# grouped reductions (no per-cell loops)
# --------------------------------------------------------------------------

def _sum(cells: np.ndarray, values: np.ndarray, n_cells: int) -> np.ndarray:
    return np.bincount(cells, weights=values, minlength=n_cells)[:n_cells]


def grouped_percentiles(values: np.ndarray, cells: np.ndarray, n_cells: int,
                        probs) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    """Per-cell linear-interpolated percentiles via one sort by (cell, value).

    Also returns per-cell counts and the sorted values, from which min and max
    are the first and last element of each run.
    """
    order = np.lexsort((values, cells))
    ordered = values[order]
    counts = np.bincount(cells, minlength=n_cells)[:n_cells]
    starts = np.concatenate(([0], np.cumsum(counts)[:-1]))
    out = np.full((len(probs), n_cells), np.nan)
    present = counts > 0
    if present.any():
        base = starts[present]
        size = counts[present]
        for i, prob in enumerate(probs):
            position = (size - 1) * (prob / 100.0)
            low = np.floor(position).astype(np.int64)
            frac = position - low
            lower = ordered[base + low]
            upper = ordered[base + np.minimum(low + 1, size - 1)]
            out[i, present] = lower * (1.0 - frac) + upper * frac
    extremes = np.full((2, n_cells), np.nan)
    if present.any():
        extremes[0, present] = ordered[starts[present]]
        extremes[1, present] = ordered[starts[present] + counts[present] - 1]
    return out, counts, extremes


def detrended_roughness(z: np.ndarray, x_m: np.ndarray, y_m: np.ndarray,
                        cells: np.ndarray, n_cells: int) -> np.ndarray:
    """Std of residuals from a per-cell least-squares plane in local metres.

    Batched 3x3 normal equations built with bincount, then an exact second pass
    for the residuals (the algebraic shortcut loses precision on large relief).
    """
    ones = np.ones_like(z)
    sums = {name: _sum(cells, value, n_cells) for name, value in (
        ('n', ones), ('x', x_m), ('y', y_m), ('z', z),
        ('xx', x_m * x_m), ('yy', y_m * y_m), ('xy', x_m * y_m),
        ('xz', x_m * z), ('yz', y_m * z))}
    normal = np.empty((n_cells, 3, 3))
    normal[:, 0, 0] = sums['n']
    normal[:, 0, 1] = normal[:, 1, 0] = sums['x']
    normal[:, 0, 2] = normal[:, 2, 0] = sums['y']
    normal[:, 1, 1] = sums['xx']
    normal[:, 1, 2] = normal[:, 2, 1] = sums['xy']
    normal[:, 2, 2] = sums['yy']
    rhs = np.stack([sums['z'], sums['xz'], sums['yz']], axis=1)
    enough = sums['n'] >= 4
    coefficients = np.zeros((n_cells, 3))
    if enough.any():
        # pinv keeps rank-deficient cells (single row or column of pixels)
        # solvable; their residuals are still measured exactly below.
        coefficients[enough] = np.einsum(
            'cij,cj->ci', np.linalg.pinv(normal[enough]), rhs[enough])
    fitted = (coefficients[cells, 0] + coefficients[cells, 1] * x_m
              + coefficients[cells, 2] * y_m)
    residual = z - fitted
    sum_squares = _sum(cells, residual * residual, n_cells)
    with np.errstate(invalid='ignore', divide='ignore'):
        roughness = np.sqrt(np.maximum(sum_squares, 0.0) / np.maximum(sums['n'], 1))
    return np.where(enough, roughness, np.nan)


# --------------------------------------------------------------------------
# the terrain-v0.2 per-cell reduction
# --------------------------------------------------------------------------

@dataclass(frozen=True)
class AnalysisSegment:
    """One rectangular analysis grid belonging to a band, halos already dropped.

    A band normally has exactly one segment. It has more than one only when the
    source's pixel geometry changes inside the band (the Copernicus GLO-30
    east-west pixel-width zones), in which case each segment carries its own
    block factors and its own column grid. Segments are reduced together, so a
    grid cell's statistics cover every pixel in it regardless of segment.
    """
    elevation: np.ndarray        # (n_rows, n_blocks) analysis elevations
    slope: np.ndarray            # (n_rows, n_blocks) Horn slope, degrees
    relief: np.ndarray           # (n_rows, n_blocks) local relief, metres
    column_of_block: np.ndarray  # (n_blocks,) analysis column -> grid column
    block_lons: np.ndarray       # (n_blocks,) centre longitude in [-180, 180)
    row_lats: np.ndarray         # (n_rows,) centre latitude

    def validate(self) -> None:
        shapes = (self.elevation.shape, self.slope.shape, self.relief.shape)
        if len(set(shapes)) != 1:
            raise ValueError(f'misaligned stages: {shapes}')
        n_rows, n_blocks = self.elevation.shape
        if self.column_of_block.shape != (n_blocks,) or self.block_lons.shape != (n_blocks,):
            raise ValueError('column metadata does not match the analysis width')
        if self.row_lats.shape != (n_rows,):
            raise ValueError('row metadata does not match the analysis height')


def cell_features(segments, n_cells: int, *,
                  center_lat: float, lon_step: float,
                  n_native_samples: np.ndarray, n_valid_native_samples: np.ndarray,
                  local_metres) -> dict:
    """Reduce one band's analysis segments to the ``terrain-v0.2`` feature set.

    ``segments`` is a sequence of :class:`AnalysisSegment` covering the band's
    interior. ``local_metres(d_lon_deg, lat_deg, center_lat)`` returns the
    east/north offsets in metres used by the plane fit, which is the only place
    the body's shape enters.

    ``n_native_samples`` and ``n_valid_native_samples`` are the caller's exact native
    pixel accounting per grid column; they are passed through untouched because
    only the caller knows how its read window partitions the source raster.
    """
    segments = list(segments)
    for segment in segments:
        segment.validate()

    row = {name: np.full(n_cells, np.nan) for name in FLOAT_FEATURES}
    row['n_native_samples'] = n_native_samples
    row['n_valid_native_samples'] = n_valid_native_samples
    row['n_samples_200m'][:] = sum(
        np.bincount(s.column_of_block, minlength=n_cells)[:n_cells] * s.elevation.shape[0]
        for s in segments)

    flat_cells = np.concatenate([
        np.broadcast_to(s.column_of_block, s.elevation.shape).ravel() for s in segments])
    elev_flat = np.concatenate([s.elevation.ravel() for s in segments])
    finite_elev = np.isfinite(elev_flat)
    elev_cells = flat_cells[finite_elev]
    elev_values = elev_flat[finite_elev]
    row['n_valid_samples_200m'][:] = np.bincount(elev_cells, minlength=n_cells)[:n_cells]
    with np.errstate(invalid='ignore', divide='ignore'):
        row['valid_frac_200m'][:] = np.where(row['n_samples_200m'] > 0,
                                            row['n_valid_samples_200m'] / np.maximum(row['n_samples_200m'], 1), np.nan)

    if elev_values.size:
        percentiles, counts, extremes = grouped_percentiles(
            elev_values, elev_cells, n_cells, ELEV_PERCENTILES)
        has = counts > 0
        row['elev_min'][:] = extremes[0]
        row['elev_max'][:] = extremes[1]
        row['elev_p05'][:] = percentiles[0]
        row['elev_p95'][:] = percentiles[1]
        totals = _sum(elev_cells, elev_values, n_cells)
        row['elev_mean'][has] = totals[has] / counts[has]
        row['elev_range'][:] = extremes[1] - extremes[0]
        row['elev_range_robust'][:] = percentiles[1] - percentiles[0]

        # local metres for the plane fit, relative to each cell's own centre
        cell_center_lon = -180.0 + lon_step * (np.arange(n_cells) + 0.5)
        lon_grid = np.concatenate([
            np.broadcast_to(s.block_lons, s.elevation.shape).ravel() for s in segments
        ])[finite_elev]
        lat_grid = np.concatenate([
            np.repeat(s.row_lats, s.elevation.shape[1]) for s in segments])[finite_elev]
        d_lon = (lon_grid - cell_center_lon[elev_cells] + 180.0) % 360.0 - 180.0
        x_m, y_m = local_metres(d_lon, lat_grid, center_lat)
        # the intercept absorbs any constant; removing one improves conditioning
        row['roughness_detrended_std'][:] = detrended_roughness(
            elev_values - float(np.mean(elev_values)), x_m, y_m, elev_cells, n_cells)

    slope_flat = np.concatenate([s.slope.ravel() for s in segments])
    finite_slope = np.isfinite(slope_flat)
    slope_cells = flat_cells[finite_slope]
    slope_values = slope_flat[finite_slope]
    if slope_values.size:
        percentiles, counts, _ = grouped_percentiles(
            slope_values, slope_cells, n_cells, SLOPE_PERCENTILES)
        has = counts > 0
        totals = _sum(slope_cells, slope_values, n_cells)
        squares = _sum(slope_cells, slope_values * slope_values, n_cells)
        mean = np.full(n_cells, np.nan)
        mean[has] = totals[has] / counts[has]
        row['slope_mean'][:] = mean
        row['slope_median'][:] = percentiles[0]
        row['slope_p90'][:] = percentiles[1]
        row['slope_p99'][:] = percentiles[2]
        variance = np.full(n_cells, np.nan)
        variance[has] = squares[has] / counts[has] - mean[has] ** 2
        row['slope_std'][:] = np.sqrt(np.maximum(variance, 0.0))
        for threshold in SLOPE_THRESHOLDS:
            over = _sum(slope_cells, (slope_values > threshold).astype(float), n_cells)
            fraction = np.full(n_cells, np.nan)
            fraction[has] = over[has] / counts[has]
            row[f'slope_frac_gt_{threshold}'][:] = fraction

    relief_flat = np.concatenate([s.relief.ravel() for s in segments])
    finite_relief = np.isfinite(relief_flat)
    if finite_relief.any():
        relief_cells = flat_cells[finite_relief]
        counts = np.bincount(relief_cells, minlength=n_cells)[:n_cells]
        totals = _sum(relief_cells, relief_flat[finite_relief], n_cells)
        has = counts > 0
        values = np.full(n_cells, np.nan)
        values[has] = totals[has] / counts[has]
        row['local_relief_1km'][:] = values

    row['n_rows_read'] = sum(s.elevation.shape[0] for s in segments)
    return row


def quality_flags(center_lat: float, valid_frac_200m: np.ndarray, n_samples_200m: np.ndarray,
                  mixed_factor: bool, degenerate: bool,
                  shared_extra=(), per_cell_extra=None) -> list[str]:
    """Per-cell advisory flags, semicolon separated, empty string when clean.

    ``shared_extra`` are band-wide flags appended after the shared ones;
    ``per_cell_extra`` is an optional per-cell iterable of extra flag lists,
    appended last. Planet-specific modules add their own vocabulary that way
    without changing the meaning of the shared flags.
    """
    shared = []
    if abs(center_lat) > POLAR_DISTORTION_LAT:
        shared.append('polar_distortion')
    if mixed_factor:
        shared.append('mixed_block_factor')
    if degenerate:
        shared.append('degenerate_block_grid')
    shared.extend(shared_extra)
    extra = per_cell_extra if per_cell_extra is not None else [()] * len(n_samples_200m)
    out = []
    for fraction, pixels, own in zip(valid_frac_200m, n_samples_200m, extra):
        flags = list(shared)
        if not pixels:
            flags.append('no_pixels')
        elif not np.isfinite(fraction) or fraction < LOW_COVERAGE_FRACTION:
            flags.append('low_coverage')
        flags.extend(own)
        out.append(';'.join(flags))
    return out


# --------------------------------------------------------------------------
# table assembly and output
# --------------------------------------------------------------------------

def _constant(value: str, length: int) -> pa.Array:
    """Dictionary-encoded constant column; these repeat on every row."""
    return pa.DictionaryArray.from_arrays(
        pa.array(np.zeros(length, dtype=np.int32)), pa.array([value], type=pa.string()))


def build_table(rows: list[dict], provenance: dict, planet: str,
                extra_float_features=()) -> pa.Table:
    """Concatenate per-band results into one Parquet table.

    Column order is fixed: identity, the shared ``terrain-v0.2`` features, the
    native pixel counts, the flags, then provenance, then any planet-specific
    ``extra_float_features`` at the very end. Appending keeps the shared prefix
    identical between planets, so the tables stay comparable column by column.
    """
    def cat(name, dtype):
        return np.concatenate([np.asarray(r[name], dtype=dtype) for r in rows]) if rows \
            else np.zeros(0, dtype=dtype)

    bands = cat('band', np.int32)
    columns = cat('column', np.int32)
    length = len(bands)
    prefix = PLANETS[planet]['prefix']
    region_id = [f'{prefix}_{b:04d}_{c:05d}' for b, c in zip(bands.tolist(), columns.tolist())]
    flags = [flag for r in rows for flag in r['quality_flags']]

    data = {
        'grid_version': _constant(GRID_VERSION, length),
        'region_id': pa.array(region_id, type=pa.string()),
        'planet': _constant(planet, length),
        'band': pa.array(bands),
        'column': pa.array(columns),
    }
    for name in FLOAT_FEATURES:
        data[name] = pa.array(cat(name, np.float64).astype(np.float32))
    for name in COUNT_FEATURES:
        data[name] = pa.array(cat(name, np.int64))
    data['quality_flags'] = pa.array(flags, type=pa.string())
    data['source_product'] = _constant(provenance['source_product'], length)
    data['source_version'] = _constant(provenance['source_version'], length)
    data['source_checksum'] = _constant(provenance['source_checksum'], length)
    data['sample_spacing_m'] = pa.array(
        np.full(length, ANALYSIS_SPACING_M, dtype=np.float32))
    data['feature_version'] = _constant(FEATURE_VERSION, length)
    data['code_version'] = _constant(provenance['code_version'], length)
    data['computed_at'] = _constant(provenance['computed_at'], length)
    for name in extra_float_features:
        data[name] = pa.array(cat(name, np.float64).astype(np.float32))
    table = pa.table(data)
    return table.replace_schema_metadata({
        'feature_version': FEATURE_VERSION,
        'grid_version': GRID_VERSION,
        'bands': json.dumps(sorted(int(r['band'][0]) for r in rows)),
        'sample_spacing_m': str(ANALYSIS_SPACING_M),
        'provenance': json.dumps(provenance, sort_keys=True),
    })


def group_path(out_dir: Path, group: int) -> Path:
    return Path(out_dir) / FEATURE_VERSION / f'band_group={group:03d}.parquet'


def write_atomic(table: pa.Table, path: Path) -> None:
    """Write beside the target and rename, so readers never see a partial file."""
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_name(f'.{path.name}.{os.getpid()}.{uuid.uuid4().hex[:8]}.tmp')
    try:
        pq.write_table(table, temporary, compression='zstd')
        os.replace(temporary, path)
    finally:
        if temporary.exists():
            temporary.unlink()


def group_is_complete(path: Path, wanted: list[int]) -> bool:
    """A resumable group must exist and already cover every band we were asked for."""
    if not path.exists():
        return False
    try:
        metadata = pq.read_schema(path).metadata or {}
        done = set(json.loads(metadata[b'bands'].decode()))
    except Exception:
        return False
    return set(wanted).issubset(done)


# --------------------------------------------------------------------------
# band-group bookkeeping
# --------------------------------------------------------------------------

def plan_groups(bands: list[int]) -> list[tuple[int, list[int]]]:
    groups: dict[int, list[int]] = {}
    for band in bands:
        groups.setdefault(band // BANDS_PER_GROUP, []).append(band)
    return sorted(groups.items())


def parse_band_range(text: str, n_bands: int) -> list[int]:
    parts = text.split('-')
    if len(parts) == 1:
        start = stop = int(parts[0])
    elif len(parts) == 2:
        start, stop = int(parts[0]), int(parts[1])
    else:
        raise ValueError('--bands takes START-END')
    if not (0 <= start <= stop < n_bands):
        raise ValueError(f'--bands must lie inside 0-{n_bands - 1}')
    return list(range(start, stop + 1))