Fuse ordered LiDAR residual error summaries

Native compiled factors with at least 8,192 observations compute residuals, squares and their ordered sum in one optional CPU Numba pass. No per-row residual array or Python float list is created for optimizer/holdout RMSE. Fastmath stays disabled and the kernel releases the GIL. CPython 3.10/3.11 retain plain ordered addition; CPython 3.12 retains Neumaier compensation, including native underflow and overflow handling. Python computes the final division and square root. Other runtimes use the original residual/list path.

The summation policy follows the CPython 3.12 implementation. Bitwise tests compare the running Python reduction, general corrections, weights, signed zero, extreme dynamic range, underflow and overflow.

The native optimizer also requests normal equations without converting their discarded residual vector to a Python list. Direct callers still receive complete residuals by default. Custom compiled-system hooks keep their existing arguments and return contract. Public residual/correction hooks, replaced reductions, small factors and unsupported runtimes keep the original path. Public residuals and final observability diagnostics remain complete.

Complete fresh-process comparisons

Each workload has three alternating pairs and independent initially empty persistent Numba/CuPy caches. Every initial compilation and slower trial is retained. Whole wall includes startup, imports, replay, all clock candidates and refinements, transfers, diagnostics, hashes, YAML and measurement saves. Parent audits are excluded; OS caches persist. Separate cohorts are neither pooled nor multiplied by earlier speedup ratios.

Comparison	Baseline median	Candidate median	Ratio of medians	Median paired ratio
Previous GPU, 960,000 captured points	69.805556 s	50.653481 s	1.378x	1.380x
Previous GPU, 48,000 captured points	13.870382 s	13.226075 s	1.049x	1.034x
Previous GPU, default sampling	7.703580 s	8.765065 s	0.879x	0.965x
Packed CPU, separate dense control	196.076865 s	39.026530 s	5.024x	5.024x

Default sampling regresses by about 14% in the ratio of medians; its median paired ratio is 0.965x. The cause of timing variation remains unresolved. Default factor sizes retain the original residual/RMSE route, without dispatching the new summary kernel. No consistent small-batch gain is claimed. Cold dense CPU wall of 289.691684 s is retained in full.

Workload	Pair	Order	Baseline whole wall	Candidate whole wall
dense	1	baseline, candidate	60.541125 s	57.133829 s
dense	2	candidate, baseline	69.805556 s	35.290130 s
dense	3	baseline, candidate	69.896860 s	50.653481 s
full	1	baseline, candidate	25.900847 s	25.043427 s
full	2	candidate, baseline	13.870382 s	10.238290 s
full	3	baseline, candidate	13.174195 s	13.226075 s
default	1	baseline, candidate	15.982607 s	16.568270 s
default	2	candidate, baseline	6.441744 s	8.765065 s
default	3	baseline, candidate	7.703580 s	6.228300 s
cpu-dense	1	baseline, candidate	289.691684 s	50.459870 s
cpu-dense	2	candidate, baseline	149.829356 s	35.999593 s
cpu-dense	3	baseline, candidate	196.076865 s	39.026530 s

All 24 formal results match after normalizing path separators and excluding only strictly validated execution identity notes. Full API results, saved YAML bytes, histories, diagnostics and logical counters are retained. Dense controls use 768,000 train and 192,000 holdout targets, rank 6 and max_iterations.

The preceding GPU comparison archives fourteen modules. The separate CPU control archives ten CPU modules and shares the packed builder, with neither GPU position queries nor the compiled packing callback supplied by that archived CPU solver. Settings, samples, splits and writers are unchanged.

Actual work outside formal timing

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	53
cpu_summary_calls	154
cpu_summary_rows	76,674,615
discarded_residual_lists_avoided	102
discarded_residual_rows_avoided	59,724,347
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	Pre-change call sum	Current call sum
CudaPlaneQueries.query	5.719896 s	4.749705 s
compiled_packing	2.759657 s	2.238720 s
_PackedObservationBuilder.build	8.894645 s	7.411736 s
cpu_residual_kernel	9.421330 s	—
cpu_residual_calls	12.476100 s	2.508685 s
fixed_rmse	7.442820 s	—
gpu_linearization	12.107272 s	5.372967 s
cpu_normal_calls	2.437897 s	2.064291 s
pair_rmse	0.119885 s	—
cpu_summary	—	10.066799 s

Call sums nest and overlap across candidates, and instrumentation differs; they are not critical-path percentages or formal phase speedups. GPU queries, finite differences, cache hits and ordered normal equations keep their work. CPU residual computation is retained; the new summary avoids object conversion and Python iteration. GPU workload coverage does not increase.

Provenance and limits

The complete pre-change profile is retained with source bytes verified against the preceding commit. Its historical comparison excludes only git commit fields; full saved provenance is retained. The current diagnostic matches the complete first formal candidate result and checks source/input bytes before and after.

The public GLIM sensor is copied into two topics as an execution parity control. This does not establish independent two-sensor calibration accuracy. Three local pairs are descriptive observations, without portable confidence estimates. Peak memory, other interpreters and Python 3.13+ acceleration are not quantified.

Measured implementation: a514397f5219b1423f3aedbc8b79be9b581f928b. The publication index binds every measured source, baseline module, config, launcher, full result and portable location metadata. Source archives preserve measured checkout bytes. API values and saved YAML bytes remain unchanged.