iop/watermark: only unset mipmap hash when watermark is enabled

[radialmonster]

Fixes the thumbnail-regeneration regression from #22103 for images
that don't use the watermark module: its metadata/tag signal callbacks
are connected in every darkroom session, not just when watermark is
enabled.

This checks self->enabled at the callback boundary, so a disabled
watermark skips both dt_iop_refresh_all() and
dt_history_hash_unset_mipmap(). Images without watermark enabled keep
a synced hash and stop regenerating their thumbnail on every darkroom
visit.

Built and tested against current master: no-op darkroom visits for
images without watermark remain synced and do not regenerate their
thumbnail. Images with watermark enabled still regenerate on every
visit after this change; this was flagged and accepted as a known
tradeoff in the discussion on #22103.

Prepared with AI assistance.

[TurboGit]

Hum, why still dt_iop_refresh_all()?

I would have done that:

diff --git a/src/iop/watermark.c b/src/iop/watermark.c
index 646e0c80a6..3e62b9e1b8 100644
--- a/src/iop/watermark.c
+++ b/src/iop/watermark.c
@@ -1405,13 +1405,13 @@ static void _metadata_changed(gpointer instance,
                               const int type,
                               dt_iop_module_t *self)
 {
-  if(self) _invalidate_tag_meta(self);
+  if(self && self.enabled) _invalidate_tag_meta(self);
 }
 
 static void _tag_changed(gpointer instance,
                          dt_iop_module_t *self)
 {
-  if(self) _invalidate_tag_meta(self);
+  if(self && self.enabled) _invalidate_tag_meta(self);
 }

[radialmonster]

I kept the change deliberately narrow and guarded only the operation causing the reported regression, preserving the existing dt_iop_refresh_all() behavior from #22103. But you’re right: when the module is disabled, there is no watermark output to refresh either, so checking self->enabled in the callbacks is cleaner. I’ll update it that way.

[TurboGit]

Thanks!