Figures are set inline, so true figure geometry costs pages: two measured experiments

[tannevaled]

Two changes that make a figure's size more correct both make the page count worse, and for the same reason: this engine sets a figure where it stands. \figure is

\long\def\figure{\par\bigskip\begingroup\centering\def\@captype{figure}\@discardopt}

so a tall graphic pushes the text down and leaves the rest of the page empty, where LaTeX would float it to the top of a page or onto a float page and let the text continue. The measurements below are the evidence, kept here so neither experiment has to be run twice.

Experiment 1 — wiring the PDF rasteriser into the CLI

RasterizePDF is a seam the engine exposes and its comment says "the CLI and loom inject one"; the CLI does not. go-tex/pdfrender v0.1.0 plugs straight in (3 lines). A vector .pdf is the commonest figure on arXiv: 673 of the 1694 graphics the 200-paper corpus includes, ahead of PNG (567), JPEG (235) and EPS (146).

Measured on the 71 papers that include a .pdf figure (300 s timeout, 63 of them have a tectonic reference):

	pages	page error	exact	0-page
placeholder (today)	1435	270	11	0
rasterised	1518	370	7	4

Four documents produce nothing at all: the renderer needs ~17 s on a figure-heavy paper and they run out of clock. The rasteriser itself is not at fault — asked for 200 dpi on a 720×864 bp MediaBox it returns exactly 2000×2400 px.

Experiment 2 — reading the size the figure file declares

A figure this engine cannot draw still states how big it is: an EPS in its %%BoundingBox, a PDF in its /MediaBox. Reading it is ~90 lines of parsing and no rendering, and it makes the reserved box the right SHAPE — today the placeholder is the requested width by a fixed 90 pt, whatever the figure.

	pages	page error	bias
90 pt default	1435	270	−158
declared box	1570	320	−66

Note the two columns disagree: the absolute error rises while the systematic bias more than halves. That is the signature of a compensating error — a placeholder that is too short has been hiding the space a floated figure would not have taken.

What would have to be true first

Real float placement: figure/table deferred to the top or bottom of a page, several to a page, and a float page when they cannot fit — \topfraction, \textfraction, \@dbltopnumber and the rest. With that in place both experiments should be re-run; without it, neither change can be judged by page counts.

Branches kept for reference: wire-pdf-figures, figure-boundingbox.

[tannevaled]

Second experiment rejected, for the record: expanding a maths source through the gullet.

The maths retry path substitutes a macro's body into the source as text. When that body is TeX rather than maths — a class's \@for loop, its \expandafter, its \number — the maths layer chokes on what arrives. The obvious move is to let the engine flatten it: run the whole source through expandList, as \edef does, and retry.

Measured over 200 arXiv papers and 200 beamer talks:

	pages	glyph paths	page error
main	4399	10283909	655
gullet retry	4416	10229675	651

The page error improves and 54234 glyphs leave the page: one paper falls from 69952 glyphs to 15669, keeping 22% of its content. A class's \ifx inside a formula opens a conditional the source never closes, and the skip ate the document from there.

That hole is now closed (#174 — an isolated expansion stops at its sentinel), and with it the same experiment costs nothing and gains nothing: same 651 page error as the guard alone, and one beamer talk loses 423 glyphs. So the gullet retry stays out; what it was reaching for — class internals inside a formula — needs the material identified before it is expanded, not a bigger hammer.

[tannevaled]

The float placer measured, and the geometry re-run on top of it

The two experiments in this issue were re-run over the 157-paper arXiv corpus
after fixing two defects in the gated placer (#196): a captured float body was
executed with no group of its own, so an assignment inside one figure left \hsize
at 0pt for the remaining 230 pages of a real paper; and a float could not ride the
page it was written on, where \@addtocurcol (latex.ltx:15636) tests it against the
room left in the column.

All rows are GOTEX_FLOATS=1 except the first, against tectonic page counts:

	Σ error	median	bias	exact	≤1 page	≤2 pages
inline (today's default)	632	3	−88	20	59	77
float placer	620	2	−106	21	60	79
placer + GOTEX_PDFASPECT	650	3	−16	25	56	77
placer + declared box (EPS %%BoundingBox / PDF /MediaBox)	637	3	−39	24	57	78
placer + declared box + PDFASPECT	650	3	−16	25	56	77

Two things are now settled.

The placer earns its place. With the two fixes it beats the inline path on every
column that matters — the median document is 2 pages out instead of 3, and 79 papers
are within 2 pages instead of 77.

Real figure geometry still costs pages, and the reason is no longer placement.
Knowing a figure's true size nearly removes the systematic shortfall (bias −106 →
−16) and gives the best exact count seen so far (25 of 157), but it moves the
typical document the wrong way: the median error goes 2 → 3 and the ≤1-page count
60 → 56. So the compensating error this issue named has not gone away, it has moved:
the placeholder was hiding space that the placer, given the true heights, still does
not use the way LaTeX does — shrinking an over-tall float to \topfraction, or
sending it to a float page rather than letting it displace text.

The last row is the useful negative: the declared box adds nothing once PDFASPECT
is on. They recover the same PDF page box; EPS is the rarer case (146 of 1694
graphics).

Nothing from the geometry side ships on this evidence. What ships is the placer.

[tannevaled]

Turning the placer on by default is blocked by #195, not by the placer

Measured over the same 157 papers, page counts against tectonic:

	Σ error	median	bias	exact	≤1 page	≤2 pages
inline (default today)	632	3	−88	20	59	77
placer, GOTEX_FLOATS=1, before #195	620	2	−106	21	60	79
placer as the default, i.e. after #195	731	3	+247	16	45	73

GOTEX_FLOATS now gates two independent things: this placer, and #195's faithful
evaluation of a one-column figure's relative width (0.9\textwidth as a real
dimension instead of the 0.9pt a raw-text scan produced). Separately each is
defensible; together they overshoot — the bias goes from 106 pages SHORT to 247
pages LONG, and the papers within one page fall from 60 to 45.

That is the same compensating error this issue is about, seen from the third
side: a figure that reserves its true width also reserves its true height, and
the placer still does not do what LaTeX does with a float that big — shrink it
against \topfraction, or move it to a float page instead of letting it push
text.

So the placer stays gated. What it needs before it can be the default is not more
placement logic but the float-size rules around it, and those have to be measured
against the widths of #195 rather than against the old sub-point placeholder.

One unrelated defect fell out of the attempt: \listoffigures and
\listoftables were rendering a heading with nothing under it, because a class
asks for its list by file name (\@starttoc{lof}) and entries are recorded under
the caption type (figure). Fixed in #197.

[tannevaled]

Where the faithful widths actually go wrong: two-column papers set in one column

Following the previous comment, the +247-page overshoot was traced rather than
guessed. Three measurements, all with GOTEX_FLOATS=1 over the 157 papers:

	Σ error	median	bias	exact	≤2 pages
inline	632	3	−88	20	77
placer + #195 widths	731	3	+247	16	73
+ \@largefloatcheck (height clamped to \textheight)	731	3	+247	16	73
+ declared box (EPS %%BoundingBox / PDF /MediaBox)	670	3	+156	18	78

\@largefloatcheck changes nothing: not one float in 200 papers is taller
than the text height, so float SIZE clamping is not the missing rule.

The declared box is worth having — 38 papers closer, 2 further — because
without an intrinsic size a width-only placeholder comes out SQUARE: measured on
2408.08037, every figure reserved 476.55 × 476.55 pt, nearly a whole page each,
21 times.

But the real cause is the column measure. 2408.08037 is
\documentclass[reprint,twocolumn,…]{revtex4-2}: its figures ask for
width=\linewidth, which in the real document is one COLUMN (~246pt) and here is
the full text width (476pt) — so each figure is twice as wide and twice as tall,
and the paper runs 29 pages against a reference of 12. Of the 57 papers that gain
3 or more pages from the faithful widths, 26 are two-column classes (revtex
reprint, acmart, IEEEtran).

So the order of work is settled by measurement, not by preference: the faithful
one-column widths of #195 cannot be turned on before two-column support
(GOTEX_TWOCOLUMN, the W2′ workstream), because half their damage is one-column
typesetting of documents that asked for two. Until then the placer is better off
with the old sub-point widths — which is exactly the state GOTEX_FLOATS gates
today, and why it stays gated.