[tannevaled]

The direction that did not exist

This package reads a PDF back as a document. Everything in the fleet that produces a richdoc — Word, Excel and PowerPoint through go-ooxml, ODT, ODS, ODP and ODG through go-odf, Markdown, CSV, JSON, XML, plain text, EPUB, MOBI, FB2 and email through go-richdoc — had nowhere to go afterwards.

fleetverbs --quiet --match ToPDF clones
→ every ToPDF in the fleet is in go-pdfkit/convert, and every one of them
  takes pictures, an SVG, a ZIP of pictures or an XPS package.

Seventeen capabilities were missing for want of this (go-pdfkit/app#71).

func Write(d *richdoc.Document, opt WriteOptions) ([]byte, error)

It composes rather than typesets: go-richdoc/html.Write (v0.3.0, written for this) renders the document as a page and go-pdfkit/html2pdf.Export lays it out. Neither half is new, and a typesetter written here would be a third implementation of line breaking, page breaking and table layout in one organisation. What that pair cannot do, this cannot either — the doc comment says so rather than leaving it to be found.

⛔ A conversion must not reach the network because the file said so

Fetch nil means refuse, and that is the design, not an oversight for a caller to fix.

A richdoc arrives from something somebody was handed — a .docx, an EPUB, an email — and its images carry URLs that whoever wrote the file chose. Fetching them by default would make every conversion a request to an address of the author's choosing, from the converting machine and with its network position. That is a server-side request forgery with a document as the payload.

TestAnImageURLIsNotFetchedByDefault hands it http://169.254.169.254/latest/meta-data/ and asserts the conversion still succeeds and that OnRefused was told — because a page with holes in it must not be silent. A caller that wants sub-resources supplies its own Fetcher, which is the only way the network is reached.

⛔ The round trip found a defect that is not ours

The PDF comes back through this package's own Parse — and the words have no spaces:

Chapterone
Asentencethathastoreachthepage.

run	ends at	next starts	gap	Run.Space
Chapter → one	118.7	122.6	3.9	15.06 (= the Tf size)
A → sentence	65.2	67.2	2.0	7.53 (= the Tf size)

go-pdfkit/extract reports Run.Space as the font size rather than a space's advance (~0.25 em). No gap can exceed a whole em, so no word break is ever found. It reaches pdfops text, markdown, word, excel and the workbench's what this page says the same way, silently — as prose without spaces rather than as an error.

Filed as go-pdfkit/extract#28 with the measurement this test produced.

⛔ The test here therefore asserts every word, in order, and names what it cannot assert yet, with the issue number and the condition for tightening it — rather than quietly comparing something weaker and leaving the next reader to think spacing was never checked.

Coverage

100.0 % of statements. The three step failures are reachable through seams (renderHTML, layOut, writePDF), because a branch no test can enter is a claim nobody checks — and each failure is asserted to say which step it was, since "laying the page out" and "writing the PDF" are different problems.

🤖 Generated with Claude Code