stac-dem.Rmd ships pre-0.5.0 figures and could not be re-baked; its numbers do not reproduce even under the old units

[NewGraphEnvironment]

Problem

vignettes/stac-dem.Rmd publishes numeric output produced by the units defect fixed in #49, and it
could not be regenerated as part of that work. It ships a caveat block instead (added in 0.5.0), but
the caveat is a stopgap — the figures on the pkgdown site are still wrong.

It is the baked half of the .Rmd.orig pattern: its chunks are plain ```r blocks with #>
output already embedded, so R CMD check and the pkgdown build render them as-is. Re-baking needs
the STAC endpoint reachable and a 1 m lidar re-run, which is why it was deferred rather than done.

What is stale

line	published figure
:85	10 m DEM valley cells: 54637 / 518400 ( 10.5 %)
:189	5 m DEM valley cells: 250564 / 2073600 ( 12.1 %)
:205	Valley area (bundled 10 m DEM): 5.46 km²
:207	Valley area (STAC 5 m DEM): 6.26 km²
:317	1 m DEM valley cells: 3,883,179 / 1.4e+07 ( 27.7 %)

The published figures do not reproduce even under the old units

Worth recording, because it means a re-bake is not simply "apply the 0.5.0 delta". Running the
vignette's own 10 m baseline configuration — field = "upstream_area_ha", slope = slope_10m,
slope_threshold = 9, max_width = 2000, cost_threshold = 2500, flood_factor = 6,
precip = precip_r, no waterbodies:

	cells
published in the baked vignette	54,637
as-coded reproduction, today	53,635
corrected (0.5.0)	28,727

So roughly 1,000 cells of that gap predate this release and come from changes since the vignette
was last baked. Nobody noticed, because a baked vignette has nothing checking it.

Configuration matters too, and is easy to get wrong when quoting a figure: the same tile gives 53,635
with explicit slope.tif and no waterbodies, but 55,345 with waterbodies. A cell count quoted
without its configuration is not reproducible.

What to do

• Re-bake stac-dem.Rmd from stac-dem.Rmd.orig with the STAC endpoint reachable
• Remove the 0.5.0 caveat block once the figures are current — it lives only in the baked .Rmd,
  deliberately not in .orig, so a clean re-bake drops it automatically
• Confirm the resolution comparison still makes its argument at corrected depths — it is the
  point of the vignette, and every run in it shrank

Worth considering while open

A baked vignette drifting silently for several releases is the underlying problem, not the units fix.
Two options, neither free:

1. A staleness guard — record the package version the vignette was baked at, and fail or warn when
   it is behind. Cheap, and it converts a silent drift into a visible one.
2. Drop the pre-bake for the 10 m baseline section, which needs no network and could run live like
   valley-confinement.Rmd does. Only the STAC and 1 m lidar sections genuinely need pre-baking.

Option 2 would have caught this one, since the stale figure is in the baseline section.

Related: vignettes/valley-confinement.Rmd runs live and was correct automatically;
vignettes/pars-floodplain.Rmd had its cached artifacts regenerated in #49. This is the only one of
the three left stale.