What a maintainer needs beyond the build commands on the landing page. The source facts and the rules the build follows are in CLAUDE.md, and what is known about the source, and how it was measured, is in research/.
catalogue_build.py --limit N writes to data/build-limit/, never data/build/, so a development build cannot be published.
s3_sync.sh uploads only the items whose digest the bucket does not already hold, and collection.json only when an item goes or the collection changed, so publishing an unchanged build sends nothing (9 minutes to check the 174,282 items of 0.3.0, measured 2026-10-11). It deletes nothing: an object the build no longer makes is reported and left.
.github/workflows/update.yml runs on the 6th of each month, and on dispatch. It pulls the build's read caches from s3://stac-pointcloud-bc/build-cache/ (headers and class samples of every file, gzipped), rebuilds the whole catalogue from them, saves them back, and publishes what changed. A month with nothing new publishes nothing, and the collection keeps its release version. A month that adds items publishes a collection.json with no version until the next release stamps one. More than 40,000 items in one run is refused as a rewrite, which is a release, published from a local build; dispatch with dryrun to plan without uploading. Registration is not in CI, because it needs the tailnet: run stacs register --mode drift, then stacs verify, from a machine that can reach the host. The run's summary gives the commands.
A red run is usually the build refusing something a person should look at (tests, a missing cache, the build's time limit and stacs audit can also fail it). It refuses a mapsheet-year listed under 90% of its recorded count, a block directory with no NTS grid, a key or file name it cannot place, a file name whose date is no date, two files with one item id, a header that cannot be read or makes no item, an EXCLUDE or EXCLUDE_DELIVERIES entry whose files no longer show their fault, a CanElevation project listed short, a COPC pairing or non-copy count that changed, a CanElevation name matched by two files on either side, an item that fails validation, a file without a current class sample, a flagged file, or one in data/classes_required.json, without a good whole read, a whole read that does not match its file's header (point count, LAS version), a whole read of a file its sample does not flag that is not recorded (catalogue_build.py --classes-record), an empty sample, and a CLASSES_UNREADABLE entry whose file now reads, fails with another error (a re-run retries it), was re-delivered, or is not in the build (gone from the listing, or excluded). The build samples every new or re-delivered file, about 120 a minute at 16 workers on a Mac (October 2026; a runner's pace is not yet measured), so a month of more than about 20,000 such files will not finish inside the build step's 240 minutes: the run saves what it sampled and the next run, or a re-dispatch, carries on. A network error is retried by re-running, which reads only what is missing. A new mapsheet-year is built, not refused, unless its file names match CanElevation COPC (the pairing check), and is recorded with --groups-record at the next release.
A local build shares the caches: cache_sync.sh pull before it and push after. push refuses to replace a cache the monthly run has written since this tree last pulled. A change to laz_item.HEADER_VERSION makes every header be read again (about 2.5 hours in October 2026), so build that once locally and push it before the next monthly run. GitHub disables the schedule of a public repository after 60 days without repository activity; re-enable it under Actions.
README.md and index.html (the GitHub Pages landing page) are generated from README.Rmd; edit that, never the outputs. The page is R. scripts/DESCRIPTION lists its packages; it is a manifest, not a package.
Rscript -e 'pak::local_install_dev_deps("scripts")'
Rscript -e 'testthat::test_file("tests/readme_functions_test.R")'The DEM demo is Python, in its own dependency group: uv sync --group readme. The CHM demo is R. Then run the two rmarkdown::render() calls in the build chunk at the top of README.Rmd. The README.md render with update_query = TRUE queries the API, runs scripts/readme_dem.py, scripts/readme_chm.R and scripts/readme_rem.R, and refreshes data/readme_cache.rds, data/readme_dem.json, fig/kanaka_dem.png, data/readme_chm.json, fig/maxan_chm.png, data/readme_rem.json and fig/maxan_rem.png; the index.html render reads them. The CHM demo's first run downloads about 0.5 GB of .laz into data/readme_laz/ (not committed), and later runs reuse it.
The page prints no count of the collection, because the monthly update would make it stale (#24). The demos' own numbers describe the figures beside them, as of the render date the page states.