# coreclr-fit · PDF export

Read this only when someone has the report and asks for a PDF.

**Check you have all of it first.** The last line of this file reads **End of pdf.md · `coreclr-fit v1`**.
If what you are holding does not end there, you have a summary of it, and the part you are missing is the
stylesheet — which means the page you build will be unstyled, which is worse than no PDF. The file is at
`skills.horus.com.uy/coreclr-fit/pdf.md`, and a copy saved to disk with `curl` is never summarised. It is not part of the analysis, and
nothing here changes what the report says.

**The PDF is the same report, typeset.** Same sections, same numbers, same wording. You are not
rewriting, shortening or re-ordering anything. If you find yourself editing the content while you
build the page, stop: the markdown file is the source and it has already been handed over.

There is one output, the whole report. The summary alone is four paragraphs and does not make a
document.

## 1 · Build the HTML

Write `coreclr-fit-{app}.html` with the report's own `{app}`, **in the same folder the PDF goes to**
(§ 5: their Downloads), never beside the report. The skill's standing promise is that the only file it
puts in their repository is the report, and an HTML left at the solution root quietly breaks it — it is
the second file § 5 argues against, 50 KB of it, with no line in any `.gitignore` expecting it.

It is self-contained, everything inline and no external requests, so it works from anywhere: the person
is going to open this on a plane or forward it to someone behind a firewall.

**Write the markup yourself. Do not run a converter.** You have the report — you either wrote it a
moment ago or you just read it — so emit the HTML the way you would write any page.

This is one path on purpose, and an earlier draft of this file offered two. It let you convert the file
with `pandoc` when the report came from somewhere else, and that turned out to be a fork nobody can
stand on: to know which branch you are on you have to read the report, and once you have read it you
have the content, so both branches are always open. Two agents took different ones and produced visibly
different documents. Every rule in §§ 1 to 3 — what comes out of the body, how a card is built, what a
blockquote becomes — is written for someone emitting markup. Run a converter and they all turn into
surgery on generated HTML, which is a second procedure nobody wrote down.

No converter also means no dependency. `pandoc` is not installed on a stock machine and neither is any
Python markdown library, and discovering that halfway through is a worse place to be than simply writing
the tags.

```html
<!doctype html><html lang="{the report's language}"><meta charset="utf-8">
<title>CoreCLR fit · {app}</title>
<style>{the whole stylesheet from § 4, inline}</style>
<body>
  <div class="head">
    <p class="eyebrow">CoreCLR fit</p>
    <h1>{app}<span class="tfm">{the TFMs, exactly as the report's own title line writes them}
      &nbsp;→&nbsp; .NET 11</span></h1>
  </div>
  <div class="summary">
    <h2>{the § Summary heading, in the report's language}</h2>
    {the four paragraphs of § Summary, each in its own <p>}
  </div>
  {the card, § 2}
  {the rest of the report, as HTML, starting at the bold paragraph that sits under the
   ASCII card — the one that says why the verdict holds — and running to the end}
  <div class="foot">
    <pre class="glyph">   ▄██▀▀██▄
 ▄█▀   ██▀  ▄
██  ▄▄▄   ▄███
██ ██▀██▄██ ██
███▀   ▀▀▀  ██
 ▀  ▄██   ▄█▀
   ▀██▄▄██▀</pre>
    <div class="txt">
      <p class="mark">horus <span>studio.</span></p>
      <p class="lead">CoreCLR fit, numbers from an upgrade we ran ourselves on .NET 11,
      measured on two phones.</p>
      <p><a href="https://skills.horus.com.uy/coreclr-fit">skills.horus.com.uy/coreclr-fit</a></p>
      <p>Need a hand with this? <a href="https://horus.com.uy/contact">horus.com.uy/contact</a></p>
    </div>
  </div>
</body></html>
```

**Four things come out of the body and do not get repeated.** The `# CoreCLR fit: …` title, because
the header carries it. The `## Summary` heading and its paragraphs, because they are already in
`.summary`. The ASCII sign-off block at the end, because the footer replaces it — a logo drawn in
characters makes sense in a terminal and nowhere else. And the `---` immediately above that sign-off,
which exists to separate it in a text file; left in, it draws a rule directly above a footer that
already has a border.

**Ordinary markdown is ordinary HTML**, and this file only writes down the transformations that are
not obvious. `*emphasis*`, `**strong**`, `` `code` ``, links, lists and fenced blocks become `<em>`,
`<strong>`, `<code>`, `<a>`, `<ul>` and `<pre><code>` with nothing special about them. Everything this
file does spell out is a case where the obvious answer is wrong.

**A table whose header row is entirely empty has no header.** Several tables in the report are
two-column key/value lists written as `| | |`, which is the markdown way of saying "no headings here".
Give those no `<thead>` at all and start them at `<tbody>`; the stylesheet puts a grey band behind
`thead th`, and a band with nothing in it reads as a rendering fault. **This is not editing the report**:
the content of an empty row is nothing, and nothing is what you render.

**A header with some empty cells is still a header, and it keeps its band.** `| | Finding | file:line |
Scope |` leads with a blank because the first column holds a severity mark, which needs no title; the
other three are doing their job. Same for `| | Measured | |`. Render those as headers. The test is
whether anything is in the row, not whether everything is.

**`{app}` is the app's name as the report writes it, spaces and all.** It goes into the `<h1>` and the
`<title>` exactly like that. Only the PDF filename in § 5 needs it flattened.

**The mark is the same ASCII glyph the report signs off with**, copied character for character out of
the sign-off block you are removing from the body. Nothing is downloaded and nothing is embedded: it is
seven lines of text in a `<pre>`, coloured with CSS.

That is also why it is the right mark here. This document came out of a terminal, and carrying the same
glyph the console and the markdown report carry makes the three obviously one thing. The half-block
characters need a line-height near 1.0 or they stretch, which is the one thing to get right.

**Nothing else in the page reaches the network either.** No fonts, no stylesheets, no analytics. The
HTML and the PDF stay on their machine.

## 2 · The card becomes real bars

The ASCII card does not survive being typeset — it is a monospace drawing that exists because a
terminal has nothing else. Rebuild it as markup, reading the values off the card you already wrote.

**Every number below belongs to an app called Northwind that does not exist.** They are here to show
the shape, and not one of them is a value you should keep. `Northwind`, `58`, `52`, `31/50`, `14/20` and
`13/30` appear in no real report, so if any of them survives into your page you copied the example
instead of reading your own card.

```html
<div class="card">
  <div class="axis"><span class="lbl">Readiness</span>
    <span class="bar"><i style="width:58%"></i></span>
    <span class="val">58/100</span><span class="cap">worth the prep</span></div>
  <div class="axis"><span class="lbl">Upside</span>
    <span class="bar"><i style="width:52%"></i></span>
    <span class="val">MEDIUM</span><span class="cap">up to 44% less time on order lists</span></div>

  <div class="verdict">{the → line from the card, as written, however many sentences it runs to,
       with the ASCII line wrapping removed}</div>

  <div class="dims">
    <div class="axis"><span class="lbl">Build &amp; TFM</span>
      <span class="bar"><i style="width:62%"></i></span>
      <span class="val">31/50</span><span class="cap">net9.0 · 32-bit RIDs</span></div>
    {one per dimension}
  </div>

  <div class="meta"><b>Breaks the build: {the count, or the word the card uses when there are
       none}</b> &nbsp;·&nbsp; After the fix list: {a} → {b}<br>
  <!-- the ✖ stays on the ASCII card and does not come here: the severity legend is further down
       the page and a lone glyph above it reads as decoration -->
  {the card's last lines: candidates verified · SDK · TFM · file counts · the version string.
   Drop the leading ── rule, which only exists to draw a line inside a box, and drop the skill
   URL, which the footer already carries}</div>
</div>
```

**The bar width is a percentage of the maximum, not of 100.** Readiness uses its own score. A dimension
uses `points / maximum × 100`, so Northwind's `31/50` renders as `62%` and its `13/30` as `43%`. Getting this wrong
is the one error a reader spots instantly, because the bar and the number sit side by side.

**Upside is the exception, because its score is not on the card.** The card shows only the band — `LOW`,
`MEDIUM`, `HIGH` — and that is deliberate: a number invites "percent of what?". The score itself is in
the heading of the report's Upside section — Northwind's reads `— 52 / 100 · MEDIUM`. Take the width from
there,
and keep the band as the value beside the bar. Do not estimate it by counting blocks in the ASCII bar;
that bar is already rounded to twenty steps.

**Both axis bars are the same colour.** Red on a `LOW` bar reads as an error when it is a finding.

**The skill's URL comes off that last line.** It is on the ASCII card because a terminal has nowhere
else to put it. Here the footer carries it, and twice on one page is once too many.

## 3 · Blockquotes become callouts

Every `>` block in the report is a note the reader should not skim past, and in a wall of justified
text they disappear. Two kinds:

- **`IMPORTANT`**, red, for anything describing a way things fail. The `Full` trimming note is always
  this one.
- **`NOTE`**, grey, for everything else, including the baseline paragraph above the gains table.

```html
<div class="callout important"><span class="kicker">Important</span>{the block's content}</div>
```

Anything else: if the block describes how something goes wrong, it is `IMPORTANT`; otherwise `NOTE`.
When it is genuinely unclear, use `NOTE`. The red one earns its colour by being rare.

The severity legend line gets `class="legend"` rather than being a paragraph like any other.

## 4 · The stylesheet

Paste it whole. It is tuned for A4 and for the shapes this report actually contains; editing it to
taste is how two reports stop looking like the same document.

```css
:root{
  --ink:#111; --muted:#6b6b6b; --line:#e4e4e4; --accent:#ff3131;
  --bar:#1a1a1a; --bar-bg:#ececec;
}
@page{ size:A4; margin:19mm 17mm 20mm; }
*{ box-sizing:border-box; }
body{
  font:10.5pt/1.62 -apple-system,"Segoe UI",Inter,Helvetica,Arial,
       "Apple Symbols","Segoe UI Symbol",sans-serif;
  color:var(--ink); margin:0; -webkit-print-color-adjust:exact; print-color-adjust:exact;
}
h1{ font-size:21pt; line-height:1.15; margin:0 0 2mm; letter-spacing:-.01em; }
h1 .tfm{ display:block; font-size:10pt; font-weight:400; color:var(--muted);
         margin-top:2mm; letter-spacing:0; }
h2{ font-size:13pt; margin:10mm 0 4mm; padding-bottom:1.5mm;
    border-bottom:1px solid var(--line); letter-spacing:-.005em; break-after:avoid; }
h3{ font-size:11pt; margin:6mm 0 2mm; break-after:avoid; }
p{ margin:0 0 3.5mm; }
strong{ font-weight:650; }
a{ color:var(--ink); }
code{ font:9.5pt/1.4 ui-monospace,"SF Mono",Menlo,Consolas,monospace;
      background:#f5f5f5; padding:.5mm 1.2mm; border-radius:2px; }
pre{ background:#fafafa; border:1px solid var(--line); border-left:2px solid var(--accent);
     padding:3mm 4mm; overflow:hidden; break-inside:avoid; }
pre code{ background:none; padding:0; font-size:8.5pt; line-height:1.45; }

table{ border-collapse:collapse; width:100%; margin:0 0 5mm; font-size:9.5pt;
       line-height:1.6; break-inside:avoid; }
thead th{ text-align:left; background:#f4f4f4; font-weight:650; color:var(--ink);
          font-size:8.5pt; text-transform:uppercase; letter-spacing:.06em;
          padding:2.2mm 3mm; border-bottom:1px solid var(--line); }
td{ text-align:left; padding:2.4mm 3mm; border-bottom:1px solid var(--line);
    vertical-align:top; }

.callout{ margin:0 0 5mm; padding:3.5mm 4.5mm; background:#f7f7f7;
          border-left:3px solid #8a8a8a; border-radius:0 3px 3px 0;
          font-size:9.5pt; line-height:1.6; break-inside:avoid; }
.callout .kicker{ display:block; font-weight:650; font-size:8pt; letter-spacing:.09em;
                  text-transform:uppercase; margin-bottom:1.5mm; color:#8a8a8a; }
.callout p:last-child{ margin:0; }
.callout.important{ background:#fff5f5; border-left-color:var(--accent); }
.callout.important .kicker{ color:var(--accent); }
.legend{ font-size:9pt; color:var(--muted); margin:0 0 5mm; }

ul,ol{ margin:0 0 3mm; padding-left:5mm; }
li{ margin-bottom:1mm; }
hr{ border:0; border-top:1px solid var(--line); margin:7mm 0; }

.head{ border-bottom:2px solid var(--ink); padding-bottom:4mm; margin-bottom:6mm; }
.eyebrow{ font-size:8pt; letter-spacing:.18em; text-transform:uppercase; color:var(--accent);
          font-weight:650; margin:0 0 2.5mm; }

.summary{ font-size:12pt; line-height:1.6; margin-bottom:7mm; }
.summary h2{ margin-top:0; }
.summary p{ margin:0 0 3.5mm; }
.summary p:first-of-type{ font-weight:600; }

.card{ border:1px solid var(--line); border-top:3px solid var(--accent); padding:5mm 5mm 4mm;
       margin:0 0 7mm; break-inside:avoid; }
.axis{ display:grid; grid-template-columns:22mm 42mm 16mm 1fr; align-items:center;
       gap:0 3mm; margin-bottom:2.5mm; }
.axis .lbl{ font-weight:650; font-size:10pt; }
.axis .val{ font-weight:650; font-size:10pt; font-variant-numeric:tabular-nums; }
.axis .cap{ font-size:9pt; color:var(--muted); line-height:1.35; }
.bar{ height:7px; background:var(--bar-bg); border-radius:1px; overflow:hidden; }
.bar > i{ display:block; height:100%; background:var(--bar); }
.verdict{ margin:4mm 0; padding:3mm 0 0; border-top:1px solid var(--line);
          font-size:10.5pt; font-weight:600; }
.dims{ margin-top:4mm; }
.dims .axis{ grid-template-columns:34mm 26mm 14mm 1fr; margin-bottom:2mm; }
.dims .lbl,.dims .val{ font-weight:500; font-size:9.5pt; }
.dims .bar{ height:5px; }
.meta{ margin-top:4mm; padding-top:3mm; border-top:1px solid var(--line);
       font-size:8.5pt; color:var(--muted); line-height:1.5; }
.meta b{ color:var(--ink); font-weight:650; }

.foot{ margin-top:11mm; padding-top:4mm; border-top:1px solid var(--line);
       display:flex; align-items:center; gap:7mm; break-inside:avoid; }
.foot .glyph{ font:7.5pt/1.0 ui-monospace,"SF Mono",Menlo,Consolas,monospace;
              color:var(--accent); margin:0; white-space:pre; flex:none;
              background:none; border:0; padding:0; }
.foot .txt{ font-size:8.5pt; line-height:1.65; color:var(--muted); }
.foot .txt p{ margin:0; }
.foot .txt .mark{ font-size:9.5pt; font-weight:650; color:var(--ink);
                  letter-spacing:.22em; margin-bottom:1.2mm; }
.foot .txt .mark span{ color:var(--accent); }
.foot .txt .lead{ color:var(--ink); }
.foot .txt a{ color:var(--muted); text-decoration:none; }
```

## 5 · Print it

**Say what you are about to do before you do it.** This skill promised to read and change nothing, and
its commands are read-only by construction. Printing a PDF is not: you write an HTML file and you launch
a browser. It is their own browser, headless, on a local file, and it installs nothing — but it is still
a program starting, and they should hear it from you rather than from a permission prompt. One sentence,
and let them say no.

**If there is nobody there to answer, say it and carry on.** A run with no one at the keyboard should
not stall on a courtesy: state what you are about to do, do it, and let the sentence stand as the record.
Waiting for a reply that cannot come is the one outcome worse than not asking.

Any Chromium browser prints a PDF from the command line, and one of these is on almost every machine.
Try them in order and use the first that exists:

```
macOS      /Applications/{Google Chrome.app/Contents/MacOS/Google Chrome,
                           Microsoft Edge.app/Contents/MacOS/Microsoft Edge,
                           Brave Browser.app/Contents/MacOS/Brave Browser,
                           Chromium.app/Contents/MacOS/Chromium}
           the same four under ~/Applications
           then the PATH, as on Linux

Windows    C:\Program Files\Google\Chrome\Application\chrome.exe
           C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe
           C:\Program Files\BraveSoftware\Brave-Browser\Application\brave.exe
           then the PATH

Linux      google-chrome · google-chrome-stable · chromium · chromium-browser
           microsoft-edge · brave-browser        (on PATH)
```

**Check the PATH on macOS too, not only the four app bundles.** A machine with Chromium from Homebrew,
or Chrome installed per-user under `~/Applications`, has a browser; a list of four hardcoded paths
reports that it does not, and then you stop for no reason.

```
"<browser>" --headless --disable-gpu --no-pdf-header-footer \
  --print-to-pdf="<output>" "file://<absolute path to the html>"
```

`--no-pdf-header-footer` matters: without it every page gets the file path and a date stamp in the
margins, which is exactly what you do not want on a document someone forwards.

**It worked when it prints `N bytes written to file <path>` and exits 0.** On macOS it also writes
`CVDisplayLinkCreateWithCGDisplay failed` and `Trying to load the allocator multiple times` to stderr.
Both are noise from a headless browser on a machine with a display, both are harmless, and neither means
the PDF is wrong. Judge the run by the written-bytes line, not by a clean stderr.

**Where it goes.** The user's Downloads folder, `~/Downloads` or `%USERPROFILE%\Downloads`, named
`coreclr-fit-{app}.pdf`, the same `{app}` as the report's file name, so `coreclr-fit-kestrel-field.md`
becomes `coreclr-fit-kestrel-field.pdf`. Not the solution root: the report is already in their repo and a second file
there is one more thing to gitignore.

**If no browser is there, say so and stop.** Leave the `.html` in Downloads, tell them the path,
and tell them to open it and print to PDF from the browser — the stylesheet already carries the print
rules, so the result is the same. Do not install anything, do not reach for pandoc or LaTeX, and do not
quietly produce something worse. A person who asked for a PDF and got a silent half-answer has to
check your work, which costs them more than the PDF was worth.

## 6 · Before you hand it over

Check these against the HTML, which you can open and read. Every one of them is answerable without a
PDF renderer:

```
[ ] Every bar width is its points over its maximum, and Upside's came from the section heading
[ ] Both ASCII blocks are gone: the card is markup now, and the footer replaced the sign-off
[ ] No table has a grey header band with nothing at all in it. A band holding one word is fine
[ ] Every number, every file:line and every sentence is the report's. Nothing from the example survived
[ ] Nothing in the page says anything the markdown report does not
```

**Do not tick anything you have not looked at.** A checklist filled in from intention rather than from
the file is worse than no checklist, because it ends the review.

**You will not be able to inspect the PDF itself**, and that is expected: rendering one needs a tool
this skill will not install. So check the HTML, which is the same document, and then say plainly what
you verified and what you did not — *"I checked the page; I have not opened the PDF"* is the honest
sentence, and it takes one line.

---

**End of pdf.md · `coreclr-fit v1`**