[matthiasschabel]

None of the four channels preserved the extreme colors, and two of them dropped the interpolation
mode as well. Verified on main before this change:

• reversed() passes only stops, name, and category, so a nearest colormap comes back linear
  with no extremes.
• with_extremes() clears anything not repeated in the call.
• __reduce__ carries only color_stops, so pickle.loads(pickle.dumps(cm)) == cm is already
  False for any colormap with bad set.
• as_dict() has no keys for interpolation or the extremes, and it is what both the pydantic
  serializer and _json_encode emit.

Following matplotlib 3.11 where it has a position:

• reversed() swaps under/over and neg_inf/pos_inf, which name the ends they extend, and
  preserves the rest.
• with_extremes() preserves anything not passed. Clearing a single color now means constructing
  a new Colormap, the same limitation matplotlib has. This is the one backward-incompatible
  change here.
• __reduce__ carries the full constructor state.
• as_dict() gains optional keys, written only when they hold non-default state, so existing
  payloads are unchanged.

Two changes that are more than preservation:

• Colormap("x_r") now swaps the catalog record's under/over, so it agrees with
  Colormap("x").reversed(). napari:HiLo is the only affected entry. An explicit
  under=/over= argument still lands on the end the caller named.
• The pydantic serializer emitted a catalog colormap's qualified name unconditionally, so
  Colormap("viridis", under="red") came back as plain viridis, and Colormap("viridis_r") came
  back unreversed because info is looked up under the stripped name. It now emits the name only
  when the name alone rebuilds the same as_dict(). __eq__ is not usable as that guard: it
  ignores name, identifier, and category, and ColorStops.__eq__ compares with np.allclose, so
  a cmap_kwargs={"start": 0.500001} cubehelix compares equal to the default one.

What I left alone, and why:

• Pickle passes color_stops as the object rather than as_dict()'s samples, so a colormap
  backed by a lut function keeps the function. Routing pickle through as_dict() shifted
  lut(17, gamma=2) of a parametrized cubehelix by 2e-5.
• A round trip through the dict form does not restore info, which is set only from a string or
  another Colormap, so an unpickled catalog colormap serializes as a dict rather than as its
  name. Fixing that means putting the qualified name in value, which would change existing
  payloads for every catalog colormap.
• identifier follows the name: reversed() and shifted() rename, so it re-derives;
  with_extremes() does not, so it is preserved. One consequence: cm.reversed().reversed()
  restores the name but not an explicitly supplied identifier. Say if you would rather it were
  carried through.
• The converters still have no representation for the finer classes.

Depends on #164 to fix CI/pydantic issues.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.90%. Comparing base (0ac9baa) to head (7d0f11e).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #155      +/-   ##
==========================================
+ Coverage   95.87%   95.90%   +0.03%     
==========================================
  Files         168      168              
  Lines        2228     2248      +20     
==========================================
+ Hits         2136     2156      +20     
  Misses         92       92              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tlambert03]

#151 is merged now! can you clean up the conflicts and merge in main?

[matthiasschabel]

OK, I think this should be merged and ready to go...

[matthiasschabel]

@tlambert03 It looks like the codecov failure is related to the pydantic serializer? I can look into it some more later...

[tlambert03]

hey @matthiasschabel, I pushed a bunch of stuff directly to your branch. here's what changed:

• merged main.
• ColormapDict: swapped the _RequiredColormapDict base-class pattern for a single TypedDict with NotRequired keys.
• serializer perf: the as_dict() comparison at serialization time was ~20-50x slower than main for catalog colormaps. now __init__ records a _catalog_name when the colormap comes from a catalog name with no other arguments, and the serializer just emits that (or falls back to as_dict()). that's back to about main's speed, and _is_identical is gone.
• __reduce__: unmodified catalog colormaps now pickle as just their name, which also restores info. everything else still pickles its full state.
• with_extremes() docstring: dropped "To clear one, construct a new Colormap", since Colormap(cm) currently drops all the extremes (a pre-existing bug, I'll follow up separately).
• separately, fix: keep the _r suffix in the name of a reversed catalog colormap #167 made Colormap("viridis_r").name come out as bids:viridis_r (it used to lose the _r).

trade-offs I'm ok with for now:

• pickled catalog colormaps now depend on the name still existing in the catalog
• Colormap(Colormap("viridis")) serializes as a dict
• any explicit argument counts as an override, even one equal to the default