Fuse continuous-time LiDAR observation construction

The optional CPU adapter combines native raw-normal scoring, plane selection and ordered gathers for batches of at least 8,192 rows. Two Numba kernels release the GIL without an inner thread pool. They retain FP64 operation order with fastmath disabled. Exact score ties at distinct centroids retain Python math.dist; equal centroids retain the first uniform plane, including signed zeros. Output columns use independent read-only NRT allocations. Small batches and unsupported kernels keep NumPy construction. Public query and observation hooks retain their guards.

This change accelerates CPU construction around the existing GPU queries and finite differences. GPU workload coverage does not increase.

Complete fresh-process comparisons

Each workload has three alternating pairs with independent initially empty persistent Numba/CuPy caches. All cold, slow and first runs are retained. Wall includes startup, imports, compilation, replay, transfers, all clock candidates and refinements, diagnostics, hashes, YAML and measurement saves. Parent audits are outside wall; OS caches persist. Workloads and cohorts are reported separately, without pooling or multiplying speedups.

Comparison / workload	Baseline median	Candidate median	Ratio of medians	Median paired ratio
Previous GPU / 960,000 captured points	53.362673 s	50.009781 s	1.067x	1.100x
Previous GPU / 48,000 captured points	10.123346 s	9.228069 s	1.097x	1.061x
Previous GPU / default sampling	5.450336 s	5.448372 s	1.000x	0.992x
Packed CPU / 960,000 captured points	136.841544 s	45.582851 s	3.002x	2.930x

The incremental dense improvement over the preceding GPU implementation is modest. This is not evidence of a large new whole-pipeline speedup. The separate CPU control describes the cumulative hybrid implementation and must not be multiplied by earlier results.

Workload	Pair	Order	Baseline whole wall	Candidate whole wall
dense	1	baseline, candidate	53.362673 s	57.517236 s
dense	2	candidate, baseline	49.359184 s	44.867407 s
dense	3	baseline, candidate	57.592006 s	50.009781 s
full	1	baseline, candidate	17.513913 s	17.596075 s
full	2	candidate, baseline	8.637609 s	8.144351 s
full	3	baseline, candidate	10.123346 s	9.228069 s
default	1	baseline, candidate	13.378151 s	14.417411 s
default	2	candidate, baseline	5.450336 s	5.386229 s
default	3	baseline, candidate	5.405486 s	5.448372 s
cpu-dense	1	baseline, candidate	140.224374 s	48.579472 s
cpu-dense	2	candidate, baseline	136.841544 s	37.436309 s
cpu-dense	3	baseline, candidate	133.544087 s	45.582851 s

All 24 formal results match exactly after normalizing path separators and excluding only strictly validated execution identity notes. Full API results, YAML artifacts, optimization histories, diagnostics and logical counters are retained. Dense controls use 768,000 train and 192,000 holdout targets; the final result has rank 6 and status max_iterations.

GPU/GPU comparisons archive fourteen preceding modules. The separate CPU control archives ten CPU modules; the packed builder is shared with the GPU position callback and compiled packing callback disabled by the archived CPU solver. Physical settings, samples and writers are unchanged.

Observed work, outside the timing cohorts

Pre-change hybrid construction	Rows
build_calls	52
rows	24,960,000
both_present	12,878,352
score_ties	783,186
same_centroid_ties	475,516
different_centroid_ties	307,670

Current execution	Count
gpu_linearization_calls	56
gpu_linearization_requests	129
initial_rank_calls	26
initial_rank_compiled	26
gpu_linearization_cache_hits	73
max_per_context_cache_bytes	33,398,792
gpu_linearization_rows	32,768,988
gpu_residual_calls	0
gpu_residual_rows	0
gpu_payload_uploads	26
gpu_payload_bytes	2,066,730,664
cpu_linearization_calls	0
cpu_residual_calls	207
cpu_normal_calls	129
cpu_uniform_query_calls	0
cpu_adaptive_query_calls	0
cpu_adaptive_repaired_rows	0
compiled_packing_calls	52
compiled_packing_rows	24,960,000
cpu_distinct_centroid_tie_rows	307,670
compiled_replay_calls	120
compiled_replay_rows	960,000
gpu_plane_query_calls	52
gpu_uniform_query_rows	24,960,000
gpu_adaptive_query_rows	24,960,000
gpu_map_upload_bytes	105,635
gpu_query_upload_bytes	1,198,080,000

Instrumented phase	Sum of call durations
CudaPlaneQueries.query	6.008207 s
compiled_packing	2.841975 s
_PackedObservationBuilder.build	9.341024 s
cpu_residual_calls	13.457525 s
gpu_linearization	9.948843 s
cpu_normal_calls	2.564417 s

These call sums overlap across candidates and nest inside other phases; they are not whole-wall percentages. GPU queries, finite differences, cache hits and ordered CPU normal equations remain unchanged. CPU replay, standalone residuals and native adaptive boundary/tie repairs remain on CPU.

Preserved diagnostics and limitations

The first pre-change diagnostic saved a complete result but its historical comparison failed because a nested diagnostic provenance also contained a different git commit. Its result, source snapshot, launcher and failure log are retained under profiles/v1. The second diagnostic excludes historical git commit fields for that comparison and confirms all numerical values; full saved provenance is retained under profiles/v2.

The public GLIM sensor is copied into two topics as an execution parity control. This does not establish independent two-sensor calibration accuracy. Local descriptive timings vary; three pairs are not portable confidence estimates. No peak-memory reduction is quantified.

Source commit: 55e8f6217e39164bbc4ff8b717e6596b16c33455. All measured sources, baseline modules, configs, launchers, complete models and portable location metadata are bound by publication-index.json. Source archives preserve measured checkout bytes. Only module/file locations are rewritten; full result values and saved YAML bytes remain unchanged.