[BUGFIX] Download topography once in the tests, and fail fast when the GMT server is down

[boriskaus]

Whats the purpose of this PR?

• Bug fix
• New feature
• Documentation update
• Other, please explain

Describe it in more detail below:

CI jobs on main have been running for two hours (and then failing) whenever the GMT data server throttled a runner; see the attempt-1 jobs of run 36567599426. A healthy run downloads only ~13 MB of tiles in ~10 s, so the volume is not the problem. Three things combined:

1. The topo cache key is the hash of test/topo_prefetch.jl, which Download topography without GMT #216 changed, so all eleven matrix jobs started with an empty cache and hit the server at the same moment. Eight got through, three were throttled.
2. Without the server index (gmt_data_server.txt), import_topo treated a tiled resolution as a single global file (earth_relief_01m_g.grd, which does not exist).
3. Nothing remembered that the server was down: the index was re-requested three times per import_topo call, 4 mirrors × 60 s each. That is the exact 12m24 per testset in the log, ten calls per run.

Also: the scheduled run was not actually testing the download, because julia-actions/cache restores ~/.julia including scratch spaces, so the tiles were already there.

Library (src/import_topo.jl)

• A copy of the server index is bundled in assets/gmt_data_server.txt (125 KB, changes rarely). The download is attempted once per session with a 15 s timeout; otherwise the bundled copy is used, so tile sizes, registration, scale and filler never depend on the network.
• After a full round of mirrors fails, download_tile records the server as unreachable (a marker file in the cache, shared with the parallel test workers) and refuses further requests for 10 minutes with a clear error naming GeophysicalModelGenerator.reset_topo_server(). A timeout is no longer treated like a missing tile and silently filled with sea level.

Tests: download once, reuse everywhere

• test/topo_prefetch.jl now lists every region/resolution any test or tutorial uses (the four La Palma sets from test_import_topo.jl were missing), so it is the only place that downloads; the tests read from the cache.
• It records the outcome in ENV["GMG_TOPO_PREFETCH_OK"], which the workers inherit. test_import_topo.jl, test_GMT.jl, test_WaterFlow.jl and the La Palma tutorial skip their topography parts (@test_skip) if the server was unreachable; the scheduled run insists, so a real breakage is still caught.
• New unit tests check the bundled index knows the tiled sets.

CI

• restore-keys prefix, so a changed key still restores the previous tiles and only new ones are fetched.
• The scheduled run deletes the tile directory before testing, so the download is genuinely exercised every two weeks.
• .raw/.part/marker files excluded from the cache; timeout-minutes: 90 as a backstop.

Worst case now if the server is down on a PR: ~10 minutes for the prefetch's one attempt, then everything else fails or skips instantly, instead of two hours.

Verified locally: cached path (0 downloads, import_topo 32/32 in 2 s) and a simulated outage (unreachable mirrors, empty cache: first call errors after one mirror round, second call refused in 10 ms, download tests recorded as skipped).

Checklist

• The PR title is descriptive and starts with the appropriate tag: [BUGFIX], [ADDITION], [DOC], etc.
• New tests (either assessing the correct behaviour of new internal functions or the correctness of a tutorial) were added, or old tests were updated
• Affected tutorials have also been updated (none affected)
• The new feature was added in a way that does not break public API
• New documentation related to the new feature was added (docstrings)
• The new code follows the contributor guidelines, in particular the Runic Style

🤖 Generated with Claude Code

[mthielma]

Test now do not run for more than 30 minutes, so this seems to work.