6 KiB
Supplied-dataset findings — spatial-reference engine
The first algorithm failed the owner's screenshots and is superseded. These results are provisional automatic decisions, not ground-truth accuracy. Rejection counts are not tuned to the owner's approximate manual count. No controlled Siril stack comparison has been performed.
| Target / FITS UTC date | Lights | KEEP | REVIEW | REJECT | KEEP integration |
|---|---|---|---|---|---|
| IC1805 / Aug 26 | 309 | 301 | 3 | 5 | 150.5 min |
| M33 / Oct 2 | 254 | 243 | 11 | 0 | 121.5 min |
| Pleiades / Sep 20 | 490 | 358 | 55 | 77 | 179 min |
| Pleiades / Oct 2 | 563 | 547 | 16 | 0 | 273.5 min |
| NGC281 / Sep 26 | 304 | 293 | 11 | 0 | 146.5 min |
| NGC7000 / Aug 17 | 147 | 141 | 6 | 0 | 70.5 min |
| Total | 2,067 | 1,883 | 102 | 82 | 941.5 min |
Across both Pleiades nights: 905 KEEP, 71 REVIEW, 77 REJECT. KEEP-only export excludes 148 exposures (74 minutes); REVIEW integration is not declared unusable. All sessions together retain 941.5 of 1,033.5 supplied minutes by default. Original FITS files are untouched. The added NGC 7000 sample is documented separately in preview and morphology findings; its two reported REVIEW exposures have visibly elongated stars, rather than cloud flags.
Reproduced failures and regression results
The five owner-reviewed examples originally passed only 1/5. The spatial-reference engine passes 5/5:
| Filename suffix | Owner judgment | New automatic |
|---|---|---|
| 20260919-235146 | Extensive upper cloud | REJECT |
| 20260919-235218 | Extensive upper cloud, previously KEEP | REJECT |
| 20260919-235652 | Clear, previously KEEP | KEEP |
| 20260919-235856 | Clear, falsely REVIEW | KEEP |
| 20260920-001734 | Extensive cloud bands, previously KEEP | REJECT |
A one-frame-at-a-time diagnostic aligned 24×40 sky maps and removed offsets/planar gradients. The three cloudy examples had 95th-percentile residual amplitudes about 250–636 ADU, compared with 10–13 ADU for the clear examples. The production engine additionally uses per-cell reference scatter, local sky subtraction, refined alignment and forced stellar apertures. Exact diagnostic amplitudes differ from production measurements because sampling and registration differ.
Before seeing v2 decisions, 56 fixed capture-order previews were screened across the five sessions: three obvious additional cloudy frames and 53 without obvious cloud structure. The first v2 policy rejected one clear-looking boundary sample based on sky structure alone despite almost normal stellar flux. The final policy sends such modest uncorroborated structure to REVIEW; overwhelming sky structure still rejects. The resulting cloud-only screen passes 56/56. This policy adjustment used the screen, so it is no longer an untouched test set.
The provisional clear-looking labels only assert not REJECT. They do not certify KEEP, PSF quality or uniform transparency. Independent display stretches can hide faint clouds. These preview observations are not user ground truth, an unbiased accuracy estimate, or proof that every bad frame was found.
Additional boundary inspection covered the 12 KEEP frames with highest cloud significance, the 12 REJECT frames with lowest significance and the 12 REVIEW frames with highest significance in the first Pleiades night. The sampled rejects show cloud bands; weaker structure mainly routes to REVIEW. One nearly opaque frame remains REVIEW because alignment is uncertain. These selected samples expose edge cases but cannot measure general error rates.
The validation executable passes 39 checks, including image-level persistent nebulosity/gradients, gradual flux drift, a prolonged moving-cloud stretch, modest sky structure without stellar loss, a wholly unstable cloudy reference, cache/label migration, source-safe export and display/source independence. These validate mechanics and stated invariants rather than stack quality.
Resource verification
The fresh measurement pass for all 1,920 frames used one worker and took 436.8 seconds, peaking at 589.8 MiB resident RAM and 559.3 MiB committed memory. A subsequent forced-photometry cache refresh with the final reference-stability guard peaked at 616.1 MiB resident RAM, 586.3 MiB committed, and took 58.6 seconds. Cached reclassification avoids image reads.
Those historical one-worker runs enforced and verified the Windows 2 GiB process allocation limit and 768 MiB managed-heap cap. The current package allows up to four workers, with a verified 6 GiB process limit and 3 GiB managed-heap limit. Analysis still checks physical headroom before each image read, preserving about one quarter of physical RAM up to 8 GiB. The page file does not count as free physical RAM. Large raw images are never retained as a project-wide image collection. See the NGC 7000 findings for current parallel benchmarks and the 2,067-frame cached refresh; that refresh peaked at 661.2 MiB resident RAM.
Current sequential packaged desktop checks passed at two window sizes, including graphs, dense spatial overlay, comparison, 40 exposure switches, filters, sorting, manual overrides, saved notes, reset, latest-preview cancellation and the nebula/stretch controls. Peak desktop RAM was 194.8 MiB. Larger measurements remain in binary caches rather than review metadata. Windows and Linux packages compile; actual Linux execution is untested.
The native memory-access crash reported earlier still has no confirmed root-cause diagnosis. Resource enforcement and successful sequential tests are reported separately from crash-cause claims.
Evidence: diagnostics/spatial-fresh-resources.json, screenshots/resource-checks.json, screenshots/smoke-checks.txt, validation-sheets/, ../tests/real-frame-labels.json, ../tests/visual-cloud-holdout.json, and ../projects/Sample Survey/report.json. The failed v1 metadata/report are retained in diagnostics/v1-project.json and diagnostics/v1-report.json. Complete manual labels and controlled all/manual/automatic Siril stacks remain necessary.