[aenawi]

The first item of M5, and the third format. PLAN §5.1.

mirsam audit page.html now works. HTML is the first adapter that is not an OOXML package — which is the point of doing it before XLSX, whose adapter would have reused everything mirsam-ooxml already has and proved nothing about the abstraction.

The direction is usually not in the document

dir="rtl" is the version an attribute linter sees. body { direction: rtl } and .rtl { direction: rtl; text-align: right } are where a great many Arabic sites actually state it, and an adapter that read attributes alone would report direction-unset on every one of those pages. That is a finding on formatting the author chose — invariant 2, reached through the adapter rather than through a rule, and the failure mode this project treats as worse than a miss.

So css.rs runs a real, if small, cascade: type / * / #id / .class / [attr] / [attr=value] / :root, descendant and child combinators, specificity, !important, and the properties CSS says are inherited. It reads <style> elements, the style attribute, and stylesheets linked by a relative path on this machine.

At-rules are skipped whole. A declaration inside @media applies under a viewport this tool does not have, and applying one would be claiming to know the reader's screen.

Every omission can only cost a finding; none of them can manufacture one. That is the test each shortcut was held to. A selector this crate cannot match is dropped — that selector, never the whole rule — and a declaration nobody applied leaves the property where it already was.

dir is a hint the stylesheet outranks

In a browser dir="rtl" is a rule in the user agent's stylesheet, so an author's direction: ltr beats it. Reading dir as the last word would report the wrong direction for a page whose CSS disagrees with its markup, and then propose a repair that changes nothing a reader sees. CascadeOrigin models the three origins and the adapter reports what the reader gets.

dir="auto" is Unset, and it stops inheriting

auto does not state a direction. It asks the browser to take the first strong character, which is precisely the sentence Resolved::Unset already carries — nothing anywhere; the renderer picks. So Arabic under it is direction-unset: a warning, the fragile tier.

That is the honest reading rather than a harsh one. First-strong gets مرحبا 2026 right and 2026 مرحبا wrong — the second begins with a number, resolves left to right, and puts the year on the wrong side. A paragraph is correct under auto when today's text happens to agree, which is not a decision about the element.

It also replaces the inherited direction rather than deferring to it. Arabic marked auto under an ltr body renders right to left, and reporting it as contradicted-by-the-chain would be a finding on text the browser already gets right.

One font stack, so complex-font-missing cannot fire here

OOXML gives a run a Latin slot and a complex-script slot, and that rule reports the document that filled the first and left the second empty. CSS's font-family answers for every script on the element, so the stack lowers into both slots and the rule's precondition is unreachable. A defect a web author cannot write, exactly as a hard left edge is one a Word author cannot write — recorded as a refusal in the conformance suite, never a skip.

The mirror of it: CSS's left and right are physical, so alignment-incoherent is live on HTML in a way it is not on Word.

The tree is html5ever's, the DOM is ours

<p>a<p>b is two siblings; a <div> inside a <p> closes it; text between <table> and its first <tr> is foster-parented out of the table. Each of those moves a node's ancestors, which is exactly what dir is inherited along — a hand-rolled nesting stack would resolve direction from a chain no browser has.

markup5ever_rcdom is not taken: it describes itself as unsupported, unfuzzed and not for production. dom.rs is the ~200 lines that avoid depending on that claim.

The one core change — ADR 0009

A package holds everything that decides its own text. A page can link a stylesheet on a server this tool will never call, because mirsam performs no network I/O — and the rules in it are rules nobody applied, so a direction-unset finding on such a page may be answered by CSS the tool never saw.

Standing rule 4 says that has to be sayable. DocumentReader gains unread_sources(), defaulting to nothing, and every report carries:

"sources": { "unread": ["https://cdn.example.test/site.css"] }

beside fonts: {checked: ...}. It reads {"unread": []} on every .pptx and .docx, which for a package is the whole truth. It returns opaque strings the engine never parses — the core did not learn what a stylesheet is.

The alternative was for the CLI to downcast to HtmlDocument, which would put a format's name in the driving adapter and leave a library caller unable to ask.

The conformance suite moved

It asks whether the adapters agree, so it has to see all of them. Inside mirsam-ooxml it would have made that crate depend on its peers — the hexagon leaking through the test tree. It now lives in mirsam-conformance, a crate with no library code in it, and its vocabulary trait writes a document rather than a package, so a single-file format needs no exception to join.

Acceptance

All 29 conformance cases pass with HTML among the formats, 24 of them without a line changing — the same situation comes back as the same finding whichever of the three applications wrote the file.

The five that moved:

• two new refusals, both properties of CSS rather than of the adapter — text-align has no distributed value, and one font-family cannot fill a Latin slot and leave the complex-script one empty;
• a_latin_font_with_an_empty_arabic_slot_… now asserts the two formats with two slots report it and the one with a single stack cannot state it;
• a_hard_left_edge_… now asserts both formats that can state it report it identically, where it used to assert only one format could;
• the committed refusal list grew by those two entries.

Corpus: quarterly-page.html carries one instance of each defect the adapter can see — including the unread stylesheet and the exit code repair gives a readable format with no writer — and quarterly-page-correct.html is the page the tool must leave completely alone. Unlike the decks beside them they are not generated: a page is its own source.

make verify is green.