[aenawi]

Four defects the web can write and the other two formats cannot: an order imposed by <bdo>, a run left un-isolated where <bdi> belonged, an inset measured from a physical edge, and a layout reversed by hand to fake right to left. Closes PLAN §5.2.

They are rules in mirsam-core, not in mirsam-html

That is the design question this item had to answer. §5.2 called them "HTML's own", and putting them in the adapter would have followed that description straight past the architecture: a rule living in a format adapter is one the next format re-implements or quietly lacks. So the shared vocabulary grew and the formats that cannot speak it say so. Alignment::Distributed is the precedent, and four more refusals now sit beside it in the committed list.

The model gained a part of a paragraph

Whether a range is isolated from its neighbours, and whether its order was imposed rather than resolved, are properties of a range, and neither is recoverable from the characters — a run boundary is something only the document knows. TextUnit::spans is that, as byte offsets into the text a report shows, the same coordinates tatweel-padding already gives its offenders. Empty for both OOXML adapters, and empty means nothing was said rather than one plain run.

The four rules

rule	severity	proved by
bidi-override	error / warning	the imposed order against what the run resolves to under the direction the override itself names
isolation-missing	warning	the paragraph resolved twice, once with the run isolated, isolates stripped before comparing
inset-physical	warning	an asymmetric physical inset under right-to-left text
order-reversed	warning	flex-direction: row-reverse on a flex container

<bdo dir="rtl"> and an embedded U+202E give the same instruction and both replace the algorithm instead of informing it, which is invariant 4 from the other end — so bidi-override and bidi-control are one defect in two spellings. The isolates isolation-missing uses are computed and never written: invariant 4 forbids a document holding a bidi control, and asking what the algorithm would have produced had one been there is what turns "this looks risky" into a finding with two orders attached.

isolation-missing is a warning and not from doubt. Nothing is mis-rendered; the algorithm did what it is specified to do. What is wrong is that the order of text outside the run is a function of text inside it, so the line is laid out correctly for this owner's name and undefined for the next.

Two arguments worth reading before review

inset-physical reports a value stated elsewhere, and alignment-incoherent does not. An absent alignment is a real state a template fills in, which is why the alignment rules split into a written tier and an inherited one. An absent inset is not a state anything fills in: a box with no inset is simply not indented. A physical left inset under Arabic therefore exists because somebody wrote it, and the only open question is where — which the evidence answers by naming the declaration.

order-reversed's provable form is narrower than its name. DOM order reversed by hand, stating nothing, leaves no trace any tool can read; a container of Arabic in some order carries no evidence of which order was intended, and claiming otherwise would be the assertion ADR 0004 forbids. The declarative form is readable and is also the common one. This limit is written into AGENTS.md rather than left implied.

Neither of the last two is fixable — a claim about repairs, not severity. Both repairs are edits to markup or to a stylesheet, not new values in the shared Fix vocabulary, and a variant nothing lowers would be inferred capability.

What the adapter had to be rebuilt for

Whitespace used to be collapsed at the end, over the finished string, and a span recorded against the raw characters names a range of a string nobody sees once three spaces became one. Text is now collapsed as it is gathered. The first version put a run's start one byte early, on the space a collapsing box had not yet written, and a <pre> of nothing but newlines crossed its own trim offsets — the adapter's own tests caught both.

Checks

• make verify green: fmt, clippy -D warnings, full suite, release build.
• 34 conformance cases (was 29). Four new refusals: a run that is a bidi boundary, an inset from a named edge in either spelling, and a reversed layout.
• tests/fixtures/quarterly-page.html carries one instance of each new defect beside the <bdi> line that proves the silence; quarterly-page-correct.html is still reported clean.
• 17 rules registered, up from 13.