Feat: Complete adaptive geodesic tracing with DP54

Add error-controlled DP5(4) integration and trusted first-crossing localization, including non-monotonic energy thresholds and representable-time stepping.

Preserve adaptive state and independent step/time retry grants across RayPool, refinement and movie scheduling. Expose numerical controls, record actual persistent-sample costs, and add v3 lens-map provenance with legacy v2 RK4 import.

Use DP54 by default and select an 8M Schwarzschild maximum step from bounded scans and a two-run 4K comparison. Retain the conservative minimum-step guard and document critical-ray and backend capability limits. Archive self-contained benchmark inputs and raw output; keep fixed RK4 HDR references explicit.

Validation: make -B -j4 BUILD_TYPE=Debug test passed; explicit RK4 HDR references have zero differences. Bounded convergence checks, benchmark reproduction, Release build and focused reviews passed. No numerical-relativity backend is added.
This commit is contained in:
wyj committed 2026-10-05 20:27:42 -04:00
1 parent c894fdb11a
commit 0a46a7095b
48 files changed
+11491 -380

No files matched your search

+545
View File
@@ -0,0 +1,545 @@
#!/usr/bin/env python3
"""Exercise the production DP54 default, v3 lens-map provenance/replay and the
adaptive retry/budget policy from the CLI.
The production default is now adaptive Dormand-Prince 5(4). These checks
require that a run without --integrator exports wire code 1 and is physically
and bit-for-bit equal to an explicit --integrator dp54 run, so `make test`
truly covers the production default rather than only the explicit path.
--integrator rk4 stays available for the legacy wire code and for a
fixed-step reference convergence check. Small CPU 16x8/32x16 scenes keep the
runtime short; physical comparisons use stored lens-map endpoints, not only the
rendered PNG.
"""
import os
import re
import struct
import subprocess
import sys
import tempfile
import zlib
from pathlib import Path
# Keep scratch data inside the pre-approved OpenCode scratch directory instead
# of creating directories directly under the system temporary root.
TMP_ROOT = Path('/tmp/opencode')
TMP_ROOT.mkdir(parents=True, exist_ok=True)
BUILD = Path(sys.argv[1] if len(sys.argv) > 1 else 'build/Release').resolve()
TESTDIR = Path(sys.argv[2]).resolve() if len(sys.argv) > 2 else BUILD
ENV = dict(os.environ, OMP_NUM_THREADS='4')
# v3 wire offsets (see src/lens_map.c). The header is deliberately not part of
# the payload CRC.
VERSION_OFFSET = 8
FRAME_COUNT_OFFSET = 32
PROVENANCE_OFFSET = 40
PROVENANCE_V3_OFFSET = 100
VERTEX_COUNT_OFFSET = 200
TRIANGLE_COUNT_OFFSET = 208
VERTEX_START = 224
VERTEX_SIZE = 108
TRIANGLE_SIZE = 32
ATOL_FIELDS = ('atol_x', 'atol_Pi', 'atol_L', 'rtol', 'min_step', 'max_step',
'max_lookback_time', 'retry_lookback_increment',
'max_total_lookback_time')
# Machine-roundoff floor for comparing two adaptive integrations; below this a
# difference carries no convergence information.
ROUNDOFF_FLOOR = 1e-12
ENDPOINT_ASSERT = 1e-6
def run(binary, *args, ok=True, env=ENV):
result = subprocess.run([str(binary), *map(str, args)], env=env,
capture_output=True, text=True)
if (result.returncode == 0) != ok:
raise AssertionError(
f'{binary.name} {args}: rc={result.returncode}\n{result.stderr}')
return result
def image_payload(path):
data = path.read_bytes()
assert data[:8] == b'\x89PNG\r\n\x1a\n', f'not a PNG: {path}'
offset, compressed = 8, bytearray()
while offset < len(data):
count, kind = struct.unpack_from('>I4s', data, offset)
payload = data[offset + 8:offset + 8 + count]
if kind == b'IDAT':
compressed.extend(payload)
offset += count + 12
return bytes(zlib.decompress(compressed))
def map_provenance(path):
data = path.read_bytes()
assert data[:8] == b'GRLENS\x01\x00'
version = struct.unpack_from('<I', data, VERSION_OFFSET)[0]
frame_count = struct.unpack_from('<Q', data, FRAME_COUNT_OFFSET)[0]
base = PROVENANCE_OFFSET
prov = {}
prov['threshold_kind'], prov['threshold_policy_version'] = struct.unpack_from(
'<II', data, base)
prov['threshold_value'] = struct.unpack_from('<d', data, base + 8)[0]
(prov['retry_step_increment'], prov['max_total_steps'], prov['max_level'],
prov['integrator']) = struct.unpack_from('<IIII', data, base + 16)
prov['min_edge_pixels'] = struct.unpack_from('<d', data, base + 32)[0]
prov['min_area_pixels2'] = struct.unpack_from('<d', data, base + 40)[0]
prov['coordinate_time_step'] = struct.unpack_from('<d', data, base + 48)[0]
prov['initial_max_steps'] = struct.unpack_from('<I', data, base + 56)[0]
if version == 3:
off = PROVENANCE_V3_OFFSET
for name in ATOL_FIELDS:
prov[name] = struct.unpack_from('<d', data, off)[0]
off += 8
prov['max_consecutive_rejections'] = struct.unpack_from('<I', data, off)[0]
return version, frame_count, prov
def map_vertices(path):
data = path.read_bytes()
version = struct.unpack_from('<I', data, VERSION_OFFSET)[0]
assert version == 3, f'expected a v3 map, got v{version}'
vertices = struct.unpack_from('<Q', data, VERTEX_COUNT_OFFSET)[0]
triangles = struct.unpack_from('<Q', data, TRIANGLE_COUNT_OFFSET)[0]
offset = VERTEX_START
values = []
for _ in range(vertices):
values.append(struct.unpack_from('<9dIIIQQQ', data, offset))
offset += VERTEX_SIZE
return values, data[offset:offset + triangles * TRIANGLE_SIZE]
def endpoint_deviation(a, b):
"""(mismatched provenance count, max direction/log-g deviation).
Vertices whose end_id/outcome/reason differ are counted as mismatches and
excluded from the numeric deviation; the caller requires zero mismatches.
"""
mismatches = 0
worst = 0.0
assert len(a) == len(b)
for x, y in zip(a, b):
if x[9:12] != y[9:12]:
mismatches += 1
continue
for k in (3, 4, 5, 6, 7, 8):
worst = max(worst, abs(x[k] - y[k]))
return mismatches, worst
def trace_cost(stderr, label):
match = re.search(label + r' trace cost: accepted=(\d+) rejected=(\d+) '
r'rhs=(\d+)', stderr)
return None if match is None else match.groups()
def sum_rhs(vertices):
return sum(v[14] for v in vertices)
with tempfile.TemporaryDirectory(prefix='gr-adaptive-cli-',
dir=str(TMP_ROOT)) as directory:
tmp = Path(directory)
for backend in ('minkowski', 'schwarzschild'):
binary = BUILD / f'{backend}_sky'
if not binary.exists():
print(f'{backend}: binary absent, skipping', flush=True)
continue
help_text = run(binary, '--help').stdout
for option in ('--integrator', '--ode-rtol', '--ode-atol-x',
'--ode-atol-pi', '--ode-atol-l', '--ode-initial-step',
'--ode-min-step', '--ode-max-step',
'--ode-max-rejections', '--trace-max-steps',
'--trace-lookback-time', '--retry-step-increment',
'--max-total-steps', '--retry-lookback-increment',
'--max-total-lookback-time'):
assert option in help_text, (backend, option)
ext = 'png' if '.png' in help_text else 'ppm'
hdr_available = '--hdr-output' in help_text
hdr_args = ['--hdr-output'] if hdr_available else []
common = ['--catalog', 'assets/sky_grid_5deg.csv', '--width', 32,
'--height', 16, '--fov-deg', 80, '--exposure', 1e-3,
'--coarse-cell-pixels', 8, '--refine-max-level', 0,
'--psf-relative-tail', 1e-4]
def single(name, *options, ok=True, env=ENV, use_common=common):
out = tmp / f'{backend}_{name}.{ext}'
result = run(binary, *use_common, '--output', out, *options,
ok=ok, env=env)
return out, result
# 1) The production default (no --integrator) must be DP54 and must
# match an explicit --integrator dp54 run physically and in its
# PNG/HDR output.
dflt_map = tmp / f'{backend}_dflt.grlens'
dflt_out, dflt_run = single('dflt', *hdr_args, '--verbose',
'--lens-map-output', dflt_map)
version, frame_count, dflt_prov = map_provenance(dflt_map)
assert version == 3 and frame_count == 1
assert dflt_prov['integrator'] == 1, dflt_prov
assert dflt_prov['min_step'] == 1e-12, dflt_prov
assert dflt_prov['max_step'] == {'minkowski': 16.0,
'schwarzschild': 8.0}[backend], dflt_prov
assert dflt_prov['min_step'] <= dflt_prov['coordinate_time_step'] \
<= dflt_prov['max_step']
assert dflt_prov['atol_x'] > 0 and dflt_prov['rtol'] > 0
assert dflt_prov['max_lookback_time'] > 0
assert dflt_prov['max_consecutive_rejections'] > 0
expl_map = tmp / f'{backend}_expl.grlens'
expl_out, _ = single('expl', *hdr_args, '--integrator', 'dp54',
'--lens-map-output', expl_map)
assert dflt_map.read_bytes() == expl_map.read_bytes(), \
'default map differs from explicit dp54'
assert image_payload(dflt_out) == image_payload(expl_out)
if hdr_available:
dflt_hdr = dflt_out.with_name(dflt_out.stem + '_HDR.fits')
expl_hdr = expl_out.with_name(expl_out.stem + '_HDR.fits')
assert dflt_hdr.read_bytes() == expl_hdr.read_bytes()
dflt_vertices, _ = map_vertices(dflt_map)
# 2) The legacy RK4 wire code must be 0 and its cost counters must be
# real (nonzero RHS evaluations), not legacy zeros.
rk4_map = tmp / f'{backend}_rk4.grlens'
single('rk4', '--integrator', 'rk4', '--lens-map-output', rk4_map)
_, _, rk4_prov = map_provenance(rk4_map)
assert rk4_prov['integrator'] == 0, rk4_prov
rk4_vertices, _ = map_vertices(rk4_map)
assert sum_rhs(rk4_vertices) > 0, 'RK4 RHS cost counters are not real'
# 3) Same-camera tolerance convergence with three levels. Outcome,
# reason and end must not change between levels, and the deviation
# from the tightest reference must shrink as the tolerance tightens.
# These are local ODE tolerances, not a global sky-error bound.
tol_maps = {}
for tol in ('1e-7', '1e-9', '1e-12'):
m = tmp / f'{backend}_tol_{tol}.grlens'
single(f'tol_{tol}', '--integrator', 'dp54', '--ode-rtol', tol,
'--ode-atol-x', tol, '--ode-atol-pi', tol,
'--ode-atol-l', tol, '--lens-map-output', m)
tol_maps[tol] = map_vertices(m)[0]
mism7, err7 = endpoint_deviation(tol_maps['1e-7'], tol_maps['1e-12'])
mism9, err9 = endpoint_deviation(tol_maps['1e-9'], tol_maps['1e-12'])
assert mism7 == 0 and mism9 == 0, \
f'{backend}: tolerance levels disagree on outcome/end'
assert err9 < ENDPOINT_ASSERT, (backend, 'default vs tight', err9)
assert err7 + ROUNDOFF_FLOOR >= err9, \
f'{backend}: tightening tolerance did not reduce error ' \
f'({err7} -> {err9})'
print(f'{backend}: default==dp54, tol errors 1e-7={err7:.3g} '
f'1e-9={err9:.3g}', flush=True)
# 4) Render-only replay of the default map must be bit-identical and
# its stored statistics must equal the live trace cost.
replay_out = tmp / f'{backend}_replay.{ext}'
replay_run = run(binary, *common, *hdr_args, '--lens-map-input',
dflt_map, '--verbose', '--output', replay_out)
assert image_payload(replay_out) == image_payload(dflt_out)
if hdr_available:
replay_hdr = replay_out.with_name(replay_out.stem + '_HDR.fits')
assert dflt_out.with_name(dflt_out.stem + '_HDR.fits').read_bytes() \
== replay_hdr.read_bytes()
live_cost = trace_cost(dflt_run.stderr, 'Frame 0')
replay_cost = trace_cost(replay_run.stderr, 'Imported map')
assert live_cost is not None and replay_cost is not None
assert live_cost == replay_cost, (live_cost, replay_cost)
# 5) Explicit zeros in the retry policy must survive, not be filled in
# by the derived defaults.
zero_step_map = tmp / f'{backend}_zero_step.grlens'
single('zero_step', '--integrator', 'dp54', '--retry-step-increment',
'0', '--lens-map-output', zero_step_map)
_, _, zero_step = map_provenance(zero_step_map)
assert zero_step['retry_step_increment'] == 0, zero_step
assert zero_step['max_total_steps'] > 0, zero_step
zero_time_map = tmp / f'{backend}_zero_time.grlens'
single('zero_time', '--integrator', 'dp54',
'--retry-lookback-increment', '0', '--lens-map-output',
zero_time_map)
_, _, zero_time = map_provenance(zero_time_map)
assert zero_time['retry_lookback_increment'] == 0, zero_time
assert zero_time['max_total_lookback_time'] >= \
zero_time['max_lookback_time'], zero_time
# 6) RK4 rejects every DP-only option rather than silently ignoring it.
rk4_errors = [
(['--integrator', 'rk4', '--ode-rtol', 1e-9], 'applies only'),
(['--integrator', 'rk4', '--ode-min-step', 1e-9], 'applies only'),
(['--integrator', 'rk4', '--ode-max-step', 1e-3], 'applies only'),
(['--integrator', 'rk4', '--ode-max-rejections', 4], 'applies only'),
(['--integrator', 'rk4', '--trace-lookback-time', 1], 'applies only'),
(['--integrator', 'rk4', '--retry-lookback-increment', 1], 'applies only'),
(['--integrator', 'rk4', '--max-total-lookback-time', 2], 'applies only'),
]
# DP cross-field validation and malformed values fail immediately.
invalid = [
(['--integrator', 'bogus'], None),
(['--integrator'], None),
(['--integrator', 'dp54', '--ode-rtol', 0], None),
(['--integrator', 'dp54', '--ode-rtol', -1], None),
(['--integrator', 'dp54', '--ode-atol-x', 'nan'], None),
(['--integrator', 'dp54', '--ode-max-rejections', 0], None),
(['--integrator', 'dp54', '--trace-max-steps', 0], None),
(['--integrator', 'dp54', '--ode-initial-step', 5,
'--ode-max-step', 1], 'min <= initial <= max'),
(['--integrator', 'dp54', '--ode-min-step', 4,
'--ode-initial-step', 2], 'min <= initial <= max'),
(['--integrator', 'dp54', '--trace-lookback-time', 0], None),
(['--integrator', 'dp54', '--retry-step-increment', -1], None),
]
for options, message in rk4_errors + invalid:
missing = tmp / 'adaptive_should_not_exist.csv'
result = run(binary, '--catalog', missing, *options, ok=False)
if message:
assert message in result.stderr, (options, result.stderr)
assert not missing.exists(), result.stderr
# 7) Replay must consume the stored policy: explicit tracing options
# cannot be layered on top of --lens-map-input.
conflict = run(binary, *common, '--lens-map-input', dflt_map,
'--integrator', 'rk4', '--output',
tmp / 'conflict.png', ok=False)
assert 'cannot be combined with --lens-map-input' in conflict.stderr, \
conflict.stderr
conflict2 = run(binary, *common, '--lens-map-input', dflt_map,
'--trace-max-steps', 100, '--output',
tmp / 'conflict2.png', ok=False)
assert 'cannot be combined with --lens-map-input' in conflict2.stderr, \
conflict2.stderr
conflict3 = run(binary, *common, '--lens-map-input', dflt_map,
'--integrator', 'dp54', '--ode-rtol', 1e-9,
'--output', tmp / 'conflict3.png', ok=False)
assert 'cannot be combined with --lens-map-input' in conflict3.stderr, \
conflict3.stderr
# 8) Single vs movie: the same physical observer event must agree, and
# threads/slabs must not change the stored adaptive endpoints.
track = tmp / f'{backend}.csv'
observer_test = TESTDIR / f'test_observer_{backend}'
if observer_test.exists():
run(observer_test, track)
moving_single = tmp / f'{backend}_moving.grlens'
single('moving', '--observer-position', 3, -4, 5,
'--observer-velocity', 0.2, -0.1, 0.3,
'--look-ra-deg', 37, '--look-dec-deg', -23,
'--camera-roll-deg', 19, '--lens-map-output', moving_single)
movie_map = tmp / f'{backend}_movie.grlens'
run(binary, *common, '--observer-track', track, '--frames-dir', tmp,
'--frames-prefix', f'{backend}_movie', '--duration', 0,
'--fps', 1, '--lens-map-output', movie_map)
single_v, _ = map_vertices(moving_single)
movie_v, _ = map_vertices(movie_map)
mism, dev = endpoint_deviation(single_v, movie_v)
assert mism == 0 and dev < ENDPOINT_ASSERT, (mism, dev)
thread_maps = []
for threads in (1, 2, 4):
m = tmp / f'{backend}_movie_{threads}thr.grlens'
run(binary, *common, '--observer-track', track, '--frames-dir',
tmp, '--frames-prefix', f'{backend}_m{threads}',
'--duration', 0, '--fps', 1, '--lens-map-output', m,
env=dict(ENV, OMP_NUM_THREADS=str(threads)))
thread_maps.append(m)
reference = thread_maps[0].read_bytes()
for m in thread_maps[1:]:
assert m.read_bytes() == reference, \
f'{backend}: DP movie map changed across threads'
slab_maps = []
for slab in (2, 8, 64):
m = tmp / f'{backend}_slab_{slab}.grlens'
run(binary, *common, '--observer-track', track, '--frames-dir',
tmp, '--frames-prefix', f'{backend}_s{slab}',
'--slab-duration', slab, '--duration', 0, '--fps', 1,
'--lens-map-output', m)
slab_maps.append(m)
base_v, _ = map_vertices(slab_maps[0])
for m in slab_maps[1:]:
other_v, _ = map_vertices(m)
mism, dev = endpoint_deviation(base_v, other_v)
assert mism == 0 and dev < ENDPOINT_ASSERT, (mism, dev)
print(f'{backend}: DP movie threads/slabs agree', flush=True)
# 9) Budget: a tiny initial coordinate-time budget leaves UNRESOLVED
# rays; with no room to grow the publication gate refuses the frame,
# while retry increments that can grow resolve it.
if backend == 'schwarzschild':
refused = tmp / f'{backend}_refused.{ext}'
small = ['--integrator', 'dp54', '--trace-lookback-time', 1e-6,
'--retry-lookback-increment', 0,
'--max-total-lookback-time', 1e-6]
result = run(binary, *common, '--output', refused, *small, ok=False)
assert 'Incomplete render refused' in result.stderr, result.stderr
assert not refused.exists()
allow = tmp / f'{backend}_allow.{ext}'
run(binary, *common, '--output', allow, '--allow-incomplete', *small)
assert allow.exists()
direct = tmp / f'{backend}_direct.{ext}'
direct_result = subprocess.run(
[str(binary), *map(str, common), '--output', str(direct),
'--integrator', 'dp54', '--trace-lookback-time', '4000'],
env=ENV, capture_output=True, text=True)
if direct_result.returncode == 0:
retried = tmp / f'{backend}_retried.{ext}'
run(binary, *common, '--output', retried, '--integrator', 'dp54',
'--trace-lookback-time', 1e-6,
'--retry-lookback-increment', 25,
'--max-total-lookback-time', 4000)
assert image_payload(retried)
print('schwarzschild: retry budget resolves previously '
'unresolved rays', flush=True)
else:
print('schwarzschild: direct budget scene still unresolved; '
'retry-resolution case skipped', flush=True)
# 10) Fixed-step reference convergence on a small scene. The
# accepted budget is explicit and large so the finer step does
# not silently shorten the traced history. Both fixed-step
# levels and the DP54 default must agree below ENDPOINT_ASSERT.
# If the reference itself does not converge this must FAIL and
# be reported, never loosened into a false zero.
ref_common = ['--catalog', 'assets/sky_grid_5deg.csv', '--width',
16, '--height', 8, '--fov-deg', 80, '--exposure',
1e-3, '--coarse-cell-pixels', 8, '--refine-max-level',
0, '--psf-relative-tail', 1e-4]
def ref_map(tag, *options):
m = tmp / f'{backend}_ref_{tag}.grlens'
run(binary, *ref_common, '--output',
tmp / f'{backend}_ref_{tag}.{ext}', '--lens-map-output', m,
*options)
return map_vertices(m)[0]
rk4_04 = ref_map('rk4_04', '--integrator', 'rk4',
'--ode-initial-step', 0.04, '--trace-max-steps',
'262144')
rk4_02 = ref_map('rk4_02', '--integrator', 'rk4',
'--ode-initial-step', 0.02, '--trace-max-steps',
'262144')
dp_tight = ref_map('dp_tight', '--ode-rtol', '1e-12',
'--ode-atol-x', '1e-12', '--ode-atol-pi',
'1e-12', '--ode-atol-l', '1e-12')
assert sum(1 for v in rk4_02 if v[10] == 0) > 0, \
'reference scene has no escaped ray'
mism, ref_dev = endpoint_deviation(rk4_04, rk4_02)
assert mism == 0, 'RK4 reference levels disagree on outcome/end'
assert ref_dev < ENDPOINT_ASSERT, \
f'RK4 reference not converged: {ref_dev}; report to parent'
mism, dp_dev = endpoint_deviation(dp_tight, rk4_02)
assert mism == 0, 'DP54 default disagrees with RK4 reference'
assert dp_dev < ENDPOINT_ASSERT, (dp_dev,)
print(f'schwarzschild: RK4 ref convergence {ref_dev:.3g}, '
f'DP default vs fine RK4 {dp_dev:.3g}', flush=True)
print(f'{backend}: adaptive CLI checks passed', flush=True)
# Alcubierre production-default smoke: the DP54 default policy must
# validate and the migrated lookback budget must actually cover the warp
# bubble feature (rays integrate and escape) instead of pre-routing all
# misses. No movie/thread sweep for this third backend.
alc = BUILD / 'alcubierre_sky'
if not alc.exists():
print('alcubierre: binary absent, skipping', flush=True)
else:
alc_help = run(alc, '--help').stdout
ext = 'png' if '.png' in alc_help else 'ppm'
alc_map = tmp / 'alcubierre_default.grlens'
alc_out = tmp / f'alcubierre_default.{ext}'
run(alc, '--alcubierre-vs', 0.3, '--alcubierre-radius', 1,
'--alcubierre-sigma', 1, '--catalog', 'assets/sky_grid_5deg.csv',
'--width', 16, '--height', 8, '--fov-deg', 80, '--exposure', 1e-3,
'--coarse-cell-pixels', 8, '--refine-max-level', 0,
'--psf-relative-tail', 1e-4, '--verbose', '--output', alc_out,
'--lens-map-output', alc_map)
_, _, alc_prov = map_provenance(alc_map)
assert alc_prov['integrator'] == 1, alc_prov
assert alc_prov['min_step'] <= alc_prov['coordinate_time_step'] \
<= alc_prov['max_step']
assert alc_prov['max_lookback_time'] > 0
alc_vertices, _ = map_vertices(alc_map)
outcomes = {v[10] for v in alc_vertices}
assert 3 not in outcomes, f'Alcubierre default left INCOMPLETE rays'
assert any(v[10] == 0 for v in alc_vertices), \
'Alcubierre default produced no escaped ray'
assert sum_rhs(alc_vertices) > 0, \
'Alcubierre default pre-routed every ray; lookback misses feature'
assert image_payload(alc_out)
# The migrated coordinate-time coverage is the physical geometry budget
# margin*4*escape/(1-|v_s|) with escape = R + 20/sigma, and it is
# independent of the accepted-step count and of the chosen initial
# step. Two cheap maps with different resource overrides must keep the
# same lookback.
alc_escape = 1.0 + 20.0 / 1.0 # R + 20/sigma for R=sigma=1
alc_sep = 1.0 - abs(0.3)
alc_expected_lookback = 1.25 * 4.0 * alc_escape / alc_sep
assert abs(alc_prov['max_lookback_time'] - alc_expected_lookback) \
< 1e-12, alc_prov['max_lookback_time']
def alc_map_with(tag, *options):
m = tmp / f'alcubierre_{tag}.grlens'
run(alc, '--alcubierre-vs', 0.3, '--alcubierre-radius', 1,
'--alcubierre-sigma', 1, '--catalog', 'assets/sky_grid_5deg.csv',
'--width', 16, '--height', 8, '--fov-deg', 80, '--exposure',
1e-3, '--coarse-cell-pixels', 8, '--refine-max-level', 0,
'--psf-relative-tail', 1e-4, '--allow-incomplete',
'--output', tmp / f'alcubierre_{tag}.{ext}',
'--lens-map-output', m, *options)
return map_provenance(m)[2]
steps_override = alc_map_with('steps_override', '--trace-max-steps', 8)
assert steps_override['initial_max_steps'] == 8
assert abs(steps_override['max_lookback_time'] - alc_expected_lookback) \
< 1e-12, steps_override
step_override = alc_map_with('step_override', '--ode-initial-step', 0.02)
assert abs(step_override['coordinate_time_step'] - 0.02) < 1e-15
assert abs(step_override['max_lookback_time'] - alc_expected_lookback) \
< 1e-12, step_override
# Extreme separation (v_s = 0.99999999) where the legacy fixed-step
# estimate far exceeds the cap. The DP path must accept an explicit
# tiny coordinate-time budget instead of being rejected at startup by
# the fixed-step guard, and it must actually exercise the quota path
# (UNRESOLVED rays or real RHS work), not pre-route everything to
# escapes. --allow-incomplete publishes the diagnostic frame.
extreme = ['--alcubierre-vs', 0.99999999, '--alcubierre-radius', 1,
'--alcubierre-sigma', 1, '--catalog',
'assets/sky_grid_5deg.csv', '--width', 8, '--height', 4,
'--fov-deg', 80, '--exposure', 1e-3,
'--coarse-cell-pixels', 4, '--refine-max-level', 0,
'--psf-relative-tail', 1e-4,
'--observer-position', 0, 0, 0,
'--observer-velocity', 0.99999999, 0, 0,
'--look-ra-deg', 0, '--look-dec-deg', 0]
ext_map = tmp / 'alcubierre_extreme.grlens'
run(alc, *extreme, '--integrator', 'dp54',
'--trace-lookback-time', 1, '--trace-max-steps', 4,
'--max-total-steps', 4, '--retry-step-increment', 0,
'--max-total-lookback-time', 1, '--retry-lookback-increment', 0,
'--allow-incomplete', '--output',
tmp / f'alcubierre_extreme.{ext}', '--lens-map-output', ext_map)
_, _, ext_prov = map_provenance(ext_map)
assert ext_prov['integrator'] == 1
assert abs(ext_prov['max_lookback_time'] - 1.0) < 1e-15, ext_prov
assert ext_prov['initial_max_steps'] == 4, ext_prov
ext_vertices, _ = map_vertices(ext_map)
assert any(v[10] == 2 for v in ext_vertices) or \
sum_rhs(ext_vertices) > 0, \
'extreme Alcubierre case did not exercise the quota path'
# The same parameters without explicit DP budgets keep the legacy
# fixed-step startup guard, which still rejects the estimated domain.
rk4_extreme = run(
alc, '--integrator', 'rk4', '--alcubierre-vs', 0.99999999,
'--alcubierre-radius', 1, '--alcubierre-sigma', 1, '--catalog',
'assets/sky_grid_5deg.csv', '--width', 8, '--height', 4,
'--fov-deg', 80, '--exposure', 1e-3, '--coarse-cell-pixels', 4,
'--refine-max-level', 0, '--psf-relative-tail', 1e-4,
'--output', tmp / f'alcubierre_rk4_extreme.{ext}', ok=False)
assert rk4_extreme.returncode == 2, rk4_extreme.stderr
assert 'cap' in rk4_extreme.stderr, rk4_extreme.stderr
print('alcubierre: default DP config validated and produced escapes; '
'extreme DP quota path ok, RK4 guard still rejects', flush=True)