Make a flat form fillable: create its fields from Iris's HTML

[bbertucc]

Closes #13.

A PDF with no form fields (a flat or scanned form) now gets them from Iris's HTML. Each <input>, <select> and <textarea> is placed on the blank next to its label. The blank is found in the rendered page: an underline, a box, a table cell (split at faint or dotted row lines), or a check box or radio circle. A field with no blank found is left out rather than put in the wrong place, and is warned field_not_placed. Fields made are listed in form.created, and --values fills them by name.

• Scans use Tesseract's word boxes, and line tracing allows for skew.
• Scans whose words are only placed approximately (no Tesseract) get no fields. Each control is warned instead.
• Fix: table header ids repeated across a page's tables, so a later table's cells pointed at the first table's headers.

Not PDF/UA: typed text uses base-14 Helvetica, which is not embedded. An output with created fields warns font_not_embedded and does not claim PDF/UA-1, the same as source AcroForms. I tried an embedded font in /DR, but mupdf's own appearance streams still use unembedded Helvetica, and viewers handle Type0 fonts in /DA poorly.

Tests: new fixtures form-lines (drawn), form-scan and form-scan-skewed (rendered at 300 dpi; the last rotated 1.5°). They check names, types and positions, filling, /TU names, field_not_placed, and the approximate case. A real two-page application (kept local) gets 73 of 79 fields with none misplaced. Of the 6 missed, 3 have labels the OCR misread and 3 have HTML labels that are not printed on the page.

🤖 Generated with Claude Code

[claude]

Blocking

Re-tagging this tool's own output drops a created check box from the structure tree. src/form/create.ts:96:

obj.put("AP", { N: { Yes: this.stream(cross(w, h), w, h), Off: this.stream("", w, h) } });
...
f.value = undefined; // its on-state is Yes, whatever the HTML's value

The on-state is forced to Yes, but on the next run (the output now has fields, so maker is off) emitField matches widgets with onStates(w.widget).includes(f.value) using the HTML's value. Reproduced: form-lines with <input type="checkbox" name="news" value="yes">, tag, then tag the output again with the same pages JSON → warnings field_on_other_page: news and field_not_in_html: news; the widget has no Form element. An untagged widget annotation fails PDF/UA and hides the control's label from a screen reader. value="yes"/"on" is ordinary HTML, and re-tagging is supported (retagged). Fix: use f.value || "Yes" as the on-state (and keep f.value), and add a retag test. The same mismatch hits names changed by unique() (a second name="date" becomes date-2, a dotted name gets -): on retag both HTML controls look for date and the second is field_not_in_pdf.

Non-blocking notes

• src/form/create.ts:136 const clean = base.replace(/\./g, "-") and the -2 suffixes mean a field is not always named by its HTML name, as README.md ("A field is named by its name") says. --values {"a.b": …} for <input name="a.b"> is refused no_acroform_field; only form.created shows the real name. Say so in the README.
• src/form/create.ts:59 radio groups are keyed per page() call, so a group whose buttons sit on two pages becomes two fields (card, card-2). This is reachable only for a group that spans a page break.

Accessibility impact: flat forms become fillable, with each field tagged and named, but a created check box whose HTML has a value loses its Form tag when the output is tagged again.

[claude]

Earlier blocking finding (a made check box's on-state forced to Yes, and names changed by unique(), dropped from the structure tree on retag): fixed in 0ad6a2e. src/form/create.ts:90 now uses f.value || "Yes" as the on-state, src/tag.ts:354-360 tries fieldNames() on retag, and test/forms.test.ts:189 checks this with 9 Form elements and state yes.

Non-blocking notes

• README naming (periods, -2): fixed.
• src/form/create.ts:59 a radio group whose buttons sit on two pages still becomes two fields (card, card-2). Unchanged since e26cdd3. Only a group that spans a page break reaches it.
• src/form/create.ts:90 (f.value || "Yes").replace(/^Off$/, "off"): an HTML value="Off" becomes on-state off, but on retag emitField still matches onStates(...).includes("Off"), so that check box loses its tag again. This is rare; matching on the same mapped value would close it.
• src/tag.ts:356 the fieldNames() fallback also runs for source AcroForms (when maker is off). An HTML control named date can bind to an existing PDF field date-2 once date is used. This is usually right, but it is not limited to fields this tool made.

Accessibility impact: a flat form's created fields, check boxes with HTML values and renamed fields included, keep their Form tags and names when the output is tagged again.