Smaller text: space glyph in the base-14 subsets, one text object per line (#288)

[fschutt]

Fixes #288.

HTML text in a base-14 family was written as two text objects per word: the word, then its space in a system fallback font, each with its own Tf BT /Span <</ActualText …>> BDC Tm TJ EMC ET. Three changes, measured on 20 OCR'd book pages (~1000 words each) rendered with from_html:

	0.12.8	+ text-object merge	+ whole-unit kerns	+ space glyph
positioned words (absolutely placed spans)	612 KiB	608	436	257 KiB
flowing paragraphs	461 KiB	457	370	184 KiB

Per page, the compressed content stream goes from 45 KB to 17 KB (positioned) and from 32 KB to 12 KB (flow). A Helvetica line is now one BT, Tf, Tm and TJ. The extracted text of the rendered pages is unchanged.

1. defaultfonts/: the space (54b4493)

None of the 14 subsets had a glyph for U+0020 / U+00A0, so every space fell back to whatever system font azul found (HelveticaNeue on macOS). That ended the glyph run at each word and embedded a second, machine-dependent font. A space has no outline, so scripts/add_space_glyphs.py adds an empty space glyph with the standard base-14 advance, mapped from U+0020 and U+00A0. It's appended so existing glyph ids are unchanged, and the script is reproducible and idempotent. Helvetica's exact coverage goes 213 → 215.

Still missing and still falling back: ! (U+0021), ” (U+201D), soft hyphen, and Ž ž € in Times-Italic / Courier-BoldOblique. These need outlines, so the subsets would have to be re-cut from their sources.

2. src/text_merge.rs: op pass under optimize (c2ab8d8)

• Text objects: consecutive ones merge when only text state (Tf, colour, Tc/Tw/Tz/TL/Ts/Tr) lies between them. This only happens when the next object sets Tm before showing or moving anything, since BT resets the text matrix.
• State: a font or fill colour that is already in effect is dropped. The tracked state is forgotten at Q and gs.
• /ActualText: spans that follow each other on the same baseline become one span with the joined text. Spans on different baselines stay separate, so line breaks survive copy-paste.

3. Glyph-run merger: whole-unit kerns, one string per unkerned sequence (c2ab8d8)

The merger wrote a kern after almost every glyph ([<0030> -0.16767642<0041> -0.15256259<0052> …]) for the sub-unit difference between the layout's positions and the integer /W widths. Kerns are now rounded to whole thousandths of an em. The exact pen is carried along, so the error stays below half a unit instead of accumulating, and glyphs without a kern between them share one hex string. The glyph_run_merge tests use exact integer kerns and pass unchanged.

Tests

tests/text_run_merge.rs (7):

• same-baseline runs merge into one text object / one span per line
• optimize: false writes the ops as given
• a text object relying on BT's matrix is left alone
• drawing between text objects keeps them apart
• Tf is set again after Q
• a Helvetica HTML line is one glyph run with a single embedded font
• extracted text is unchanged by the pass

All 207 tests pass.

🤖 Generated with Claude Code