Add topo/dtopo input roadmap to dev/design

[mandli]

Design document for tracking and future plans for the heavy redo of the input file formats, notably topo, dtopo and met input.

[mandli]

PR Detailed Description -- Summarized by Claude

Why this is tracked rather than in .plans/

dev/design/README.md makes the case already: a source of truth in a gitignored
path cannot be reviewed, cannot be updated in the same PR as the code, cannot be
git blamed, and can be lost — several met-forcing docs were lost that way.
This roadmap is the topo/dtopo equivalent, and until now there was no topo
design doc at all; only the met ones existed.

The Contents entry in dev/design/README.md is updated in the same commit,
which that file's own same-PR rule requires.

What it records

§1 — merge order. The topo stack is strictly linear (#739 → #741 →
topo-input-parity), verified with git merge-base --is-ancestor rather than
assumed, plus the five independent PRs and the geoclaw-before-clawutil ordering
for #745 / clawutil#208.

§1 — supersession, with the evidence inline. #727 and topo-input-unify are
being closed and deleted, so the reasoning has to survive them:

• Combine buffer and anti-meridional wrapping #727's mechanism was reimplemented, not dropped — a four-row table maps each
  piece (_expand_wrapped_topo → _resolve_topo_records; Python-side buffer
  baking → Fortran nbuf4 = topo_buffer(topo_idx)) and each of its three tests
  to a named, stronger equivalent in Topo crop_extent / antimeridian: stop failing silently, and make the working paths reachable #739.
• topo-input-unify's six phase commits are listed against their 1:1 rebased
  equivalents on parity (19573e4e→994e6efb, and so on).

This is what makes the close/delete auditable after the fact. It also records
that c893f238 fixed a bug in the same area: leaving nbuf4 = 0 on the
nc_has_crop branch silently dropped topo_buffer for every file
topo_entries() writes. master still has the single assignment; #739 has both.

§2 — five deferred items, each with what blocks it rather than a bare TODO:

±180 seam gap	design decision; three specific questions written out
dtopo Fortran crop/wrap parity	mandli#15
1D preprocessing scope	opened by #745; needs a decision, not a default
build-flag hygiene	radial_flat hardcode; the ALL_LFLAGS question
-DNETCDF / make new	documented, with both rejected fixes recorded

The seam-gap entry is the one worth reading: it explains why buffer cannot
close the hole (after #739 the buffer is applied, but clamped to the file's own
extent, and neither entry has data past its edge), so the next person does not
re-try tuning it.

§3 — the #740 correction on record. The topo0save bug in #742 was
initially believed to explain the missing tsunami. It does not: #740 coarsens
GEBCO 20× to 0.167° while dtopo_usgs100227.tt3 is 0.101° (read from the
file header), so the finer topo_for_dtopo grid outranks the topo entries and
is deformed regardless. Recorded so the theory is not re-tried.

§4 — how each claim was checked, including what was not: that parity
passes CI on a machine other than the author's, and the remote tests.

The CI correction in the second commit

The first version of §2.3 said CI selects only python and slow markers.
That is wrong — testing.yml also has a regression-tests job, and
slow-tests.yml has matrix regression/python jobs. Found while checking
whether parity's tests/regression/topo_crop/ suite would actually run in CI.

The conclusion was unaffected and still holds — git grep -l 1d_classic -- tests/ .github/ is empty, so no 1D example is built or run anywhere in CI —
but the supporting detail would have misled anyone auditing coverage, and it
mattered in the other direction too: regression is a first-class CI selector,
which is what #742's topo0save test and parity's topo_crop suite rely on.

The paragraph now carries a table of all five jobs with their marker strings
quoted verbatim from the workflow files, each checked against
grep -hoE '\-m "[^"]+"' .github/workflows/*.yml.

Verification

Documentation only, so the check is that the references are real. Each was
confirmed against the repositories at 2026-09-07:

• a3ee4b27 (Add fetch_remote_topo; deprecate read_netcdf; refactor etopotools #726 merge), the two merge-base ancestry relations, parity's
  12 commits.
• nbuf4 = topo_buffer occurrence counts: master 1, Topo crop_extent / antimeridian: stop failing silently, and make the working paths reachable #739 2.
• Makefile.common:66 / :84, examples/bouss/radial_flat/Makefile:76.
• dtopo_usgs100227.tt3 dx = 1.01010101010104e-01 from its header.
• Complete dtopo crop/wrap parity with topo (ASCII + netCDF) in the Fortran reader mandli/geoclaw#15 open, with the title quoted.
• All five CI marker strings verbatim; 1d_classic absent from tests/ and
  .github/.