Four checks that reported a result on something they had never looked at
A check that fails is useful. A check that passes without measuring is worse than no check, because you stop looking. Four of Orbit's shipped with that shape and all four are fixed here. orbit_check_email_auth graded a domain that does not exist in DNS, warning that no DKIM selector was found "among the 27 common defaults that answered" — one fact about a missing zone, replayed 27 times and totalled as evidence. orbit_dark_mode_check returned pass on a composite whose logo it had never inspected, because img is self-closing and both of its colour walkers stepped straight over it. orbit_accessibility_lint caught a heading skipping forward and was blind to one dropping backward — absent from the issues, the passes and the not-measured list alike. And orbit_score_subject_line gave the README's best line and two deliberately generic ones the same 96/sharp.
What shipped
•Bug fix — orbit_check_email_auth no longer grades a domain that has no DNS zone. A non-existent domain NXDOMAINs every name under it, and each of those counted as a selector that answered, which produced a graded warning about the email authentication of a domain nobody has ever registered. A lookup that cannot resolve now reports not-measured with the reason and tells you to check the spelling and pass the root domain rather than a hostname. The apex read it uses is the one SPF needed anyway, so this costs no extra lookup.
•Bug fix — orbit_dark_mode_check reads images. It measured colour pairs and bare white text and walked past every img tag, so a header with a real logo on a resolvable light ground returned pass with zero invert risks, whatever the image actually was. It now flags an image sitting on a ground that inverts with no dark-mode adaptation declared — either a picture ancestor carrying a prefers-color-scheme source, or the data-ogsc/data-ogsb hook Apple Mail respects. It makes only the claim markup can support: whether an arbitrary PNG's pixels survive inversion needs a rendered client, and an image whose background cannot be resolved is reported as not measured rather than scored as a pass.
•Bug fix — orbit_accessibility_lint catches a heading level dropping to one the outline never opened. It tracks the heading sequence as a nesting stack now, the same model a screen-reader user builds navigating by level: h3 back to an h2 that is still open is an ordinary end-of-subsection and stays silent, while h5 straight to an h1 that was never opened is flagged, because that is where the reader loses their place. It deliberately does not flag every decrease. A lint that fires on every well-structured email is a lint people turn off.
•Behaviour change — orbit_score_subject_line is wired to Orbit's slop detector. It scored three lines of very different quality identically because its own dictionaries are literal patterns and marketing jargon has endless variants. The detector's structural heuristics are dropped on the way in — anaphora, paragraph rhythm, fragment ratio and hedge density read a shape a subject line does not have, and letting them fire would penalise good copy for being short. What survives is the phrase and skeleton layer, scoped to the first 300 characters.