fix: validate bbox input and guard empty geocoder results

[eklavya072]

overview

BboxParamType only checked that it got four numbers, so a fair bit of unusable
input was getting through — nan, coordinates off the globe, inverted boxes.
this adds tests for those cases and the validation to reject them.

what changed

convert() now rejects non-finite ordinates, lon/lat outside -180/180 and
-90/90, and min >= max, which handles the inverted and zero-area cases in one
rule. the finite check runs first because nan compares False against everything
and was slipping straight through the range check. bounds match the ones
cmask.py already uses.

buffer_bbox clamps latitude to ±90 so a bbox already touching a pole doesn't
get pushed past it. longitude is left alone since a bbox crossing the antimeridian
looks like a legitimate AOI, and mgrsconv wraps longitude elsewhere anyway.

get_bbox_label / get_best_semantic_label were indexing rg.search(...)[0]
with no guard, so an AOI with no geocoder match came out as an IndexError. both
raise BadParameter now. get_bbox_label's -> str annotation is also now
-> dict, which is what it actually returns.

testing

tests/rapida/test_bbox_param_type.py covers all four functions. the edge cases
went in as xfail(strict=True) so they'd flip the moment the behaviour changed —
they're plain tests now. 38 passing, and the example in admin.py still parses.

one thing i'm unsure about: rejecting a zero-area bbox is the only change here
that isn't purely a bug fix. happy to drop it if a degenerate bbox is meaningful
somewhere.

files:

• rapida/util/bbox_param_type.py
• tests/rapida/test_bbox_param_type.py

[eklavya072]

Flagging this in case it slipped past — happy to adjust anything here.

This is test-only (tests/rapida/test_bbox_param_type.py, +300/−0); no runtime code is touched. Beyond covering the paths for BboxParamType.convert, buffer_bbox, get_bbox_label and get_best_semantic_label, the xfail(strict=True) cases document six validation gaps I hit while reading the code:

1. Non-finite ordinates are accepted — nan,nan,nan,nan and 1,2,inf,4 pass float() and are returned unchanged.
2. No lon/lat range check — 200,10,210,20 and 10,100,20,110 are accepted.
3. No min < max check — an inverted bbox like 10,0,5,1 is accepted.
4. Zero-area bbox accepted — 5,5,5,5 is passed downstream rather than failing early.
5. buffer_bbox can push latitude past 90° — max(cos, 1e-6) guards the longitude divisor, but nothing clamps maxlat + dlat.
6. An empty geocoder result raises IndexError — rg.search(...)[0] is indexed unconditionally in both get_bbox_label and get_best_semantic_label, so the failure surfaces as IndexError rather than click.BadParameter.

One small extra: get_bbox_label is annotated -> str but returns the raw geocoder dict. The tests assert the current dict behaviour rather than the annotation — happy to flip that if the annotation is the intended contract.

I used xfail(strict=True) so each case flips to a failure the moment it's fixed, rather than silently passing. If you'd rather these were skip, or split from the passing tests into a separate PR, I'm glad to restructure.

I think this needs maintainer approval for workflow runs on a first-time fork contribution. I'll rebase onto latest main whenever that's useful.

If a regression test around #491 would help, I'm happy to look at that next.

[iferencik]

@eklavya072 thank you for submitting this PR. You are spot on with those 6 use cases, all of them are valid. The BboxParamType is very primitive (expects only a sequence of 4 values) . It would be nice to enhance the validation with your edge cases

[eklavya072]

thanks for looking at it. pushed the validation.

convert() now rejects non-finite ordinates, lon/lat out of range, and min >= max,
which handles the inverted and zero-area cases in one rule. nan had to be checked
before the range check since it compares False against everything and was slipping
straight through.

also clamped latitude in buffer_bbox so a bbox already at a pole doesn't get pushed
past 90, and guarded the rg.search(...)[0] indexing in both label functions. left
longitude unclamped since a bbox crossing the antimeridian looks like a legitimate
AOI, and mgrsconv wraps longitude elsewhere anyway. used the same -180/180, -90/90
bounds cmask.py already uses.

heads up this isn't test-only anymore like my earlier comment said, it touches
rapida/util/bbox_param_type.py now. the 12 xfail cases are plain tests since they
pass. 38 passing locally, and the documented example in admin.py still parses.

two things i wasn't so sure on:

rejecting a zero-area bbox (5,5,5,5) is the only change that's a behaviour call
rather than a straight bug fix. easy to drop if a degenerate bbox means something
somewhere.

the label functions raise BadParameter, which fits the CLI path but is a bit loose
when bbox comes from gdf.total_bounds in connectivity. happy to switch it if you'd
rather something else there.

renamed the PR since "test:" doesn't really fit now.