[mkiani12]

Addresses issue #46.

The three reported bugs are already fixed on main

Verified before writing anything, by running the reproducers:

	state	check
SpectralProfileCalculator.histogram_export()	fixed — no ax.figure(...), uses ax.set_xlabel/set_ylabel	writes a real PNG; xaxis/yaxis both length 6
PCACalculator.histogram_export()	fixed — if self._output is None	works on a fresh instance without execute() first
MagDirCalculator ties	fixed, and vectorized	identical dates → magnitude all zero, direction all 0

grep confirms no other hasattr(self, "_output") survives, as you suggested checking.

So this PR is about your fourth section

That was the part worth acting on. 89% coverage, and all three defects lived in covered code — because the tests asserted that the plumbing got called rather than that the code ran. Your diagnosis of each was exact: the MagicMock() as ax accepts ax.figure(...) happily; the PCA test called process() on the line before the broken guard; the MagDir fixture used values where no difference is ever zero.

tests/smoke_test.py is the counterweight, built roughly as you described — real synthetic georeferenced GeoTIFFs on disk, every calculator constructed through its public API, execute() / histogram_export() / to_raster() called for real, assert a file appears. Nothing patched.

It came out a bit larger than the ~20 lines you estimated, mostly because the 37 constructors are irregular (channel, exportType, level, time, operation, selecte, kernel sizes, tif_paths). Keeping it table-driven contains that, and a guard test fails if a newly exported calculator has no smoke case, so the coverage can't quietly rot. GLCM gets an 8×8 raster of its own so its per-pixel loop stays in milliseconds.

Notably SVMCalculator is in there — it's smoke-testable at all only because the non-interactive training_samples path from #39 exists.

It immediately found four more bugs

Which is the argument for the approach, so worth listing. Two are regressions from the nrows/ncols forwarding added for the docs/API issue:

MosaicCalculator.execute() was broken outright. BaseTool.execute() forwards nrows/ncols to _export_file(), but MosaicCalculator overrides _export_file and never got those parameters — every call raised TypeError: takes from 2 to 11 positional arguments but 13 were given. Fixed.

PCA wrote files called None_<uuid>.png. PCACalculator._export_file() bypasses BaseTool._export_file, so it repeated neither the output-directory creation nor the tool-name fallback for filename_prefix. Since that argument now defaults to None, it was interpolated literally — into a directory that didn't exist, so it raised FileNotFoundError too. Fixed.

GeoeyeCalculator on a single-band raster raised IndexError: tuple index out of range from shape[2], rather than saying it needs a multispectral stack. Now a clear ValueError. (My first smoke case fed it the wrong input — but the error should say so.)

MosaicCalculator.process() never assigns self._output. The merged raster is written directly by _export_file, which sets _output to a file path rather than an array, so to_raster() can't work for this tool and the documented BaseTool lifecycle doesn't hold. Left alone deliberately — fixing it changes public behaviour and three existing tests assert the current semantics. It's documented in the smoke test's skip reason; happy to take it as a separate issue if you think it should change.

Proving the tests actually bite

The complaint in your issue is that tests can look like they cover something without measuring it, so it seemed wrong to just assert these are better. I reintroduced each of the three original defects and confirmed the corresponding tests fail:

mutated: spectral profile ax.figure(...) restored
mutated: PCA hasattr guard restored
mutated: MagDir zero-case now inherits class 1

FAILED test_histogram_export_writes_a_png[SpectralProfileCalculator]
FAILED test_spectral_profile_histogram_export_runs
FAILED test_histogram_export_writes_a_png[PCACalculator]
FAILED test_pca_histogram_export_works_without_execute_first
FAILED test_magdir_handles_pixels_that_did_not_change

Both the generic harness and the targeted tests fire on each. Then reverted.

Results

521 passed on Python 3.12; 520 on 3.11 and 3.13; all runs headless with DISPLAY/WAYLAND_DISPLAY unset. The smoke suite adds ~7 seconds. The 25 skips are the 24 calculators without a histogram_export, plus the documented Mosaic to_raster case.

Not included

GLCM's per-pixel cost, which you mention alongside the MagDir vectorization. MagDir is done, but GLCM is a genuine windowed co-occurrence computation rather than something four np.where calls replace, and it got ~4× more expensive when angle-averaging was added for #38. It deserves its own PR with before/after timings and numerical-equivalence tests, rather than being folded into a correctness change.