Corrected annual series: per-pixel run-length filter that keeps a sustained switch and overwrites flicker years

[NewGraphEnvironment]

Context

#62 ran 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). The #9 body named a "corrected series" — a per-pixel mode filter that keeps a sustained break and overwrites the flicker years — as worth building only if flicker turned out large everywhere. It did.

The numbers (data-raw/logs/break_class_groups/summary_groups.csv, one row per group)

group	changed_ha	pct_sustained	pct_endpoint	pct_flicker	stable_flicker_ha	pct_stable_flicker_of_valid
bulk	4625.0	19.7	36.4	44.0	3186.5	7.76
necr	5779.4	31.0	29.4	39.6	3828.3	9.15
lnth	1629.7	20.6	31.0	48.5	1576.7	9.85
kotl	3537.8	24.7	33.1	42.3	2235.5	3.22

Flicker is the largest category of the 2017 -> 2023 changed area in every group (39.6-48.5%), and a further 3.2-9.9% of every floodplain flickers while reading stable on the endpoints — 0.63 to 0.97 times the changed area itself (stable_flicker_over_changed), invisible to the two-epoch layer. Full reading in inst/notes/temporal-qa-groups.md.

What to build

A function that takes the same named list of classified years dft_rast_break_class() takes and returns a corrected series (a named list, same grid) in which:

• a pixel with n_flips == 0 is unchanged;
• a pixel with n_flips == 1 is unchanged — a clean switch is real by this leg's evidence, whatever its n_before / n_after (the caller thresholds, as everywhere else in the package);
• a pixel with n_flips >= 2 is overwritten year by year with a mode that respects a sustained switch: not the whole-series mode (dft_rast_consensus() already does that and votes a real mid-window switch back to its old class), but a run-length rule — a run of at least run_min consecutive years of one class is kept, and any run shorter than that is relabelled to the neighbouring long run (the earlier one on a tie, or by a documented rule).

The two-epoch transition of the corrected series is then the deliverable a report can quote as change: dft_rast_transition(corrected, from = "2017", to = "2023"). Its changed area should land near the pct_sustained + pct_endpoint share of today's, and the note's Q1 table is the acceptance check — run the corrected series through dft_rast_break_class() again and the flicker row should be near zero.

Open decisions for the plan gate

• run_min default: 2 matches the "sustained two years each side" reading used throughout Categorical breakpoint detection: sustained switch vs flicker across the annual class series (the temporal leg of change QA) #9 / Temporal QA across five watershed groups from the published annual series: does the BULK split generalise? #62.
• What to do with a flicker pixel that has no run of length run_min at all (e.g. ABABABA): leave as is and flag, or mode. Either is defensible; the count of such pixels is the number to report.
• Whether the endpoint-only breaks (n_before == 1 or n_after == 1, 29-36% of changed area) are touched. Temporal QA across five watershed groups from the published annual series: does the BULK split generalise? #62's Q2 says 2017 is the odd endpoint somewhat more often than 2023 in every group (ratio 1.09-1.25) and argues against moving the baseline; this function should leave clean switches alone and let the caller decide.

Relates to #9, #62. Scale-test on the BULK pair before the PR per CLAUDE.md; the four groups' COGs are one data-raw/break_class_groups.R <group> away.