Temporal QA across four watershed groups from the published annual series (#62)

[NewGraphEnvironment]

Summary

• data-raw/break_class_groups.R runs dft_rast_break_class() on the published seven-year IO LULC series of every group that has one — bulk_co_ff04, necr_ch_ff04, lnth_ch_ff04, kotl_bt_ff04 — straight from the stac-floodplains-bc hrefs (verified against file:checksum), then the Detect geometric edge/misregistration artifacts in transitions (sliver width, boundary-hugging, reciprocal pairs) #44 patch pipeline and floodplain-shape rows. Its summarize stage emits every number the note quotes and refuses to finish if the note's tables differ from the generated file.
• inst/notes/temporal-qa-groups.md answers the four questions in Temporal QA across five watershed groups from the published annual series: does the BULK split generalise? #62: the sustained share of 2017-2023 change is 19.7 / 31.0 / 20.6 / 24.7 % (bulk / necr / lnth / kotl), flicker 44.0 / 39.6 / 48.5 / 42.3 % and the largest category everywhere, so transition_2017_2023 overstates sustained change 3.2-5.1x; 2017 is the odd endpoint more often than 2023 in every group (ratio 1.09-1.25); flicker moves 9 points across a 2.5x width range, so it is mostly the classifier's; the geometric and temporal legs are independent to nearly independent. Between-group spread is confounded with producer path (the two cube-cut groups flicker least) and the note says so.
• Evidence in data-raw/logs/break_class_groups/ (per-group CSVs, RSS traces, summary_groups.csv / .md, BULK reconciliation). Nothing in R/ changes. Five groups became four: PINE was dropped upstream (floodplains#76); issue Temporal QA across five watershed groups from the published annual series: does the BULK split generalise? #62's body was edited before the run.

Related Issues

• Closes Temporal QA across five watershed groups from the published annual series: does the BULK split generalise? #62
• Follow-ups filed with the numbers: Corrected annual series: per-pixel run-length filter that keeps a sustained switch and overwrites flicker years #64 (corrected annual series), Carry temporal evidence (break_n_flips, break_year) beside transition_2017_2023 for full-series items stac_floodplains_bc#67 (break_n_flips / break_year assets)
• Relates to Categorical breakpoint detection: sustained switch vs flicker across the annual class series (the temporal leg of change QA) #9, Detect geometric edge/misregistration artifacts in transitions (sliver width, boundary-hugging, reciprocal pairs) #44, Annual IO LULC series (2017-2023) for bulk, necr, lnth, kotl: rerun step 3 with every year floodplains#79, Publish all seven classified years for bulk, necr, lnth, kotl stac_floodplains_bc#59
• Relates to NewGraphEnvironment/sred-2025-2026#16

Scale run (m1, 64 GB, terra 1.9.34, file-backed inputs)

group	cells	wall s	scan s	peak RSS GiB
bulk	169,248,352	312.3	61.2	15.4
kotl	204,497,703	322.6	72.2	16.3
lnth	60,522,376	99.9	21.1	13.7
necr	56,246,160	137.9	22.3	14.2

Peak RSS sits at 14-16 GiB whatever the grid size — terra sizing its working set to available RAM — and the #9 run's 8.5-20 GB in-memory input floor is gone. BULK re-measured on the published grid against the #9 run: +71 valid cells, +4.62 ha changed, shares within 0.03 of a point (summary_bulk_reconcile.csv).

Test plan

• devtools::test(): 867 pass, 0 fail, 13 skip (no R/ change)
• Four code-check rounds plus a plan review (five agents); findings and their fixes in planning/archive/2026-09-issue-62-temporal-qa-groups/review-round{1..4}.md. The plan review rejected the per-stream segment design for Q3 before the run; rounds 2-4 each found the same mechanism one artifact further out (a prose number typed from a reading of the table, a derived cell downstream of a display rounding), reaching a filed issue body; the loop ended by enumerating every artifact that carries a number from summary_groups.csv.
• summarize re-run after every edit; note tables verified verbatim by the script

Notes

• The per-group stage writes two areas rounded (area_km2 2 dp, area_ha 1 dp) that ff06_over_ff02 and pct_area_artifact then divide; recomputed from unrounded sources all eight cells are unchanged. Recorded in the evidence README to fix on the next per-group run rather than re-running the batch for no change in any reported value.
• data-raw/break_class_groups-run.sh starts Rscript alone so the RSS sampler reads R's PID; a first sampler had caught the wrapper shell.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XcFmzirEsjTxQd3jzGdnPZ