tepbc-ross-retyped/CONTINUATION.md

CONTINUATION — Ross 1988 re-typeset

State (end of session 5, 2026-09-15)

Edition complete and verified: 455 output pages, make verify clean — 431/431, 0 overfull, 0 underfull font warnings, reprocheck 0 tokens short. 84 underfull boxes remain, the documented cost of ragged right without hyphenation. Git is the history; the PDF is not tracked (make remakes it).

Changed this session:

• Text face is XCharter, replacing TeX Gyre ScholaX (Sammay). Chosen on metrics: ScholaX had the smallest lowercase relative to its capitals of the faces tested, x-height/cap 0.645 vs 0.717, so it read as decorative. Bengali needed no manual scale — Scale=MatchLowercase re-derived Tiro Bangla and the measured Bengali-to-Latin ratio held (2.12 → 2.08). Page count 470 → 455; underfull boxes 123 → 84. The measure is now ~73 characters a line against the ~70 that 5.95in was chosen for — an open decision, see below.
• Small caps removed entirely. A typescript cannot set them, so the source defines none.
• Clickable regions marked: 0.4pt hairline rule, RGB(95,125,165), on Contents/List-of-Plates page numbers and resolved cross-references. Endnote superscripts and whole-line entries stay unmarked. polish.py emits \xref for the marked kind and \hyperlink for the unmarked.
• Nine diacritic corrections, all verified against the page image at 500–900 dpi. Full record and the ~30 still unchecked are in QUESTIONS.md.
• The edition's one editorial [sic] was removed: it now interpolates nothing into the author's prose.
• Plates inside quotations are full-measure again (\platefullmeasure); plate 1's \plx entry is in build_maps(), so pl. 1 links at last.

Next, in the order I would take them

1. Link audit. CLAUDE.md requires re-auditing the cross-reference rule after any change to it, and polish.py changed twice this session (\xref, and \plx added to build_maps). Print every link with its context and read them. Never done; nothing else validates that a link lands on the right page.
2. The measure — hold ~70 characters (about 5.7in) or accept 73. One line in geometry.
3. Diacritic sweep remainder, ~30 occurrences listed by page in QUESTIONS.md. Last five batches found nothing, so the yield looks low; unchecked is not cleared.
4. Quotation slips — 10 checkable in a library, 5 in archives nobody outside them can reach.

Earlier state

State (end of session 4, 2026-09-14)

• All 431 PDF pages transcribed. No \unsure readings outstanding.
• Chapter 5 Vincent Figgins and Richard Watts closed at orig. p.211 (PDF 213, notes 1–93, plates 89–96).
• Chapter 6 European Typefoundries orig. pp.214–240 (PDF 216–242, notes 1–81, plates 97–105).
• Section C Indigenous Ventures title at orig. p.241 (PDF 243).
• Chapter 7 The Girīśa Vidyāratna Press and later foundry types. orig. pp.242–263 (PDF 244–265, notes 1–54, plates 106–117).
• Part II Mechanical Typefounding and Composition of Bengali title at orig. p.266 (PDF 268).
• Chapter 8 Linotype Composition orig. pp.267–308 (PDF 269–310, notes 1–124, plates 118–133).
• Chapter 9 Monotype Composition orig. pp.309–337 (PDF 311–339, notes 1–54, plates 134–148).
• Part III Photocomposition of the Bengali Script title at orig. p.338 (PDF 340).
• Chapter 10 Filmsetting orig. pp.339–363 (PDF 341–365, notes 1–74, plates 149–160).
• Chapter 11 Digital Photocomposition orig. pp.364–404 (PDF 366–406, notes 1–80, plates 161–174).
• Conclusion orig. pp.405–410 (PDF 407–412, notes 1–19).
• Epilogue orig. pp.411–418 (PDF 413–420, notes 1–14, plates 175–178).
• Bibliography orig. pp.419–429 (PDF 421–431), set with the new biblist/bibgroup/bibhead macros: a bold group heading, then one entry per paragraph with a 1.6em hanging indent, set solid.

Earlier state (end of session 3, 2026-09-13)

• Building clean: 195 edition pages, 199/431 PDF pages done. No \unsure readings outstanding.

• Done (PDF pages): 1 title, 2 colophon, 3 dropped, 4–16 front matter, 17–30 Introduction (orig. pp. 15–28, notes 1–26, plates 1–6), 31 Part I title, 32 Section A title, 33–74 chapter 1 Charles Wilkins (orig. pp. 31–72, notes 1–131, plates 7–25).

• Errata section at the end; seven entries. \erratum + src/errata.tex + make check keep text and list in step. PLAN.md's Editorial row records the policy; its Type row records the specimen rule.

• Inline type specimens: 214 cut so far into plates/inline/, set with \ig at native size. Workflow per page: NB work/lines.py's line N is usually the line after the one counted visually — if a sheet comes back with the wrong line, try the band one earlier. work/lines.py N lists the text lines → work/glyphs.py N y0,y1 … makes a numbered contact sheet of the ink runs on chosen lines → tools/crop_inline.py N name=INDEX[:left|right|first|last|mid|trimright|trimleft] … cuts them. When the space between a word and the glyph is too noisy to register as a gap, measure the run's column profile and use the INDEX+PX form (e.g. 48+63).

• Watch for scanner bars in plate crops. The --rule guard that stops engraved rules splitting a plate also let thin bars survive; crop_plate.py now sweeps any very dark thin run at an edge after autocropping (strip_edge_bars). Nine plates were re-cropped for this. If a plate looks like it has a black line down one side, that is the cause — re-run make plate with its recorded recipe.

• Bengali policy (Sammay, 2026-09-13): modern script → Tiro Bangla text; a historical typeface → cut from the scan and set with \ig. Slips inside quoted matter are never corrected — mark them \qslip{} (logged to qslips.log) and list them in QUESTIONS.md for checking against the source.

Typography and plates — reworked 2026-09-13/14 at Sammay's request

• Type: 11pt TeX Gyre ScholaX (Century Schoolbook, from texlive-fonts-extra, added to the Dockerfile for it) on a 5.2in measure ≈70 characters a line; leading 1.22; ragged right; hanging punctuation; \frenchspacing; widows/orphans forbidden; heads a step larger; endnotes smaller and tighter (\notesection); running head names the current chapter.
• tools/polish.py runs first in build.sh: src/pages (faithful transcription, never hand-edit the output) → work/pages (what is actually typeset). It applies ties, en-dashes for ranges, small caps for institutional acronyms, \mbox round specimens, and cross-reference links. See CLAUDE.md for the full contract, especially the internal-vs-external rule for p. N references and the instruction to re-audit the links after changing it.
• Contents and List of Plates entries are whole-line links to page.N.
• Plates: crop_plate.py deskews an out-of-square page (rule-centre measurement, profile-variance fallback) and sweeps the grey gutter shadow. Deskew is the only step that resamples. All 89 plates were re-cropped; a sample was checked and none lost content. If a plate looks tilted or has a dark band down one side, re-run make plate N=… ARGS="<recipe from the manifest>".

Known cosmetic issue

Three plate files (p0130, p0131, p0218) report an identical Overfull \hbox (57.8pt); the rendered pages show no visible damage and the plates sit inside the measure. Cause not yet identified — not the image width (the images differ) and not the margin note (moving it outside the minipage changed nothing). The three share one trait: their images are wide enough that width=\textwidth (not height=0.74\textheight) is the binding constraint, so the minipage content is exactly the measure and anything else on the line tips it over.

Open with Sammay — see QUESTIONS.md

Nothing outstanding. The quotation-slip table in QUESTIONS.md lists the 15 suspected slips inside quoted matter, left as printed for checking against the quoted sources; it is a courtesy list, not a blocker.

Next

1. Transcription is complete. What remains is finishing work, not conversion: read the built PDF end to end for damage the per-batch spot-checks could miss; settle the two open QUESTIONS items (p.295 compliments, p.348 a dropped to); move the QUESTIONS "adopted corrections" list into the colophon.
2. Expect more inline specimens — budget for the lines.py → glyphs.py → crop_inline.py loop per page.
3. Later polish: move the QUESTIONS "adopted corrections" list into the colophon; a plate-only page's running head shows the previous text page too ("orig. pp. 16–38") — acceptable, refine if wanted.

Never kill a build mid-run without clearing work/main.aux

A full build is ~20 min at 420 pages, so the temptation is to stop one that is carrying a change you are about to supersede. Don't, unless you then rm -f work/main.aux work/main.toc work/main.out. build.sh does not clear them, so the next run reads back a .aux truncated mid-\enotez@note and dies at \begin{document} with File ended while scanning use of \enotez@note / Runaway argument? — an error that looks like an unbalanced brace in a page file and is not. Check tail -c 200 work/main.aux first: a clean block-boundary size (e.g. exactly 172032 bytes) ending mid-entry is the tell.

Hyphenation — settled, do not re-open

The original never breaks a word at a line end. Checked over all 431 leaves: 59 line-ending hyphens, every one a compound the author typed or OCR noise off a plate image (work/lineend-hyphens.json holds the list). Output hyphenation is therefore off (\lefthyphenmin=62, \hyphenpenalty=10000), with \exhyphenpenalty left low so a typed hyphen may still end a line as it does in the original.

Unnumbered in-text figures — \fig

Chapters 7 and 8 set illustrations in the run of the text with no plate number and no caption (the sentence before them carries the reference): orig. p.245's two cut/penned ব forms, orig. p.259's tier diagram. \fig{plates/fig-pNNNN.png} (added 2026-09-14) centres one at native size, shrinking it only if it is wider than the measure. Cut these by explicit pixel box with PIL — crop_plate.py is for whole-page plates — and record the box in the manifest notes column.

A printed page number inside the plate's own column range

Plate 119 (PDF 272) has the thesis's printed page number sitting in the same columns as the plate's lower figure, so no --box can exclude it. Crop such a plate with PIL instead: paint the page number's rectangle white, then take the ink bounding box. Record both steps in the manifest notes column.

Framed plates need --raw

Chapter 6's plates are reproductions of whole framed pages (plates 97–103). The frame's own border rules sit at the very edge of the crop, where strip_edge_bars reads them as scanner bars and sweeps them off — taking content with them on p0222, where two lines were lost. crop_plate.py --raw (added 2026-09-14) takes an explicit --box exactly: no shadow sweep, no deskew, no bar sweep, no autocrop. Measure the box first by finding the frame's rules (columns/rows whose ink mean exceeds 0.5), then pass --raw --box L,T,R,B as fractions of the raw page — --box is fractional, not pixels, and pixel values raise a PIL DecompressionBombError. Every plate recipe is in the manifest notes column.

Measuring the original's leading — autocorrelate, don't count bands

Counting gaps between ink bands gives nonsense: a line's ascender band and its x-height band are often detected separately, halving the apparent pitch, and a page with a plate or a displayed extract poisons the median. Pages came back at 40, 53, 70, 94 px for the same document. Autocorrelating the row-ink profile over the text block and taking the peak in 35-140 px is robust: every ordinary page then returns 94-95 px at 300 dpi, and so does every specimen page. The original is leaded 22.9 pt uniformly — the thing to reproduce. Its specimens reach 16.9 pt at the 90th percentile, 74% of the line, and only 0.7% crowd it. Our build now measures 23.1 pt on a plain page and 23.1 pt on a specimen page (r=0.85 / 0.41; the low r is a seven-line block, not unevenness).

First-line indent and the ragged2e trap

The original never indents: it separates paragraphs by a blank line, and being justified, that blank line sits under a flush line and reads clearly. Reflowed and ragged right the cue is much weaker, so this edition indents 1.6em and keeps \parskip small (Sammay, 2026-09-14). Setting \parindent alone is not enough — ragged2e's \RaggedRight assigns \parindent from \RaggedRightParindent (0pt by default) every time it is invoked, so set both, and remember that any \RaggedRight inside a macro re-zeroes the indent locally.

Swallowed digits: a TeX integer assignment eats the text after it

Reported 2026-09-14: the Halhed shelfmarks on orig. p.420 came out as - 5593, and 5596, A and 5660B, E, F, A, B. The transcription in src/pages/p0422.tex was correct and so was work/pages/p0422.tex; the loss was in the typesetting. biblist set \everypar{\hangindent=1.6em\hangafter=1}, and \hangafter is an integer parameter: TeX goes on scanning digits after the 1, so a paragraph beginning 5590 - 5593 assigned \hangafter=15590 and the number was gone. Fixed by terminating the assignment with \relax. Any numeric assignment in an \everypar, a hook, or a macro that can be followed by body text needs \relax.

One build at a time in work/ — now enforced by a lock

Every pass runs in work/, so two concurrent builds share main.aux: the second reads the first's half-written copy and dies at \begin{document}. It surfaces as "Undefined control sequence" or "File ended while scanning use of \enotez@note", and the reported line belongs to main.aux, not to the file the message appears to name — chasing it through src/main.tex or the page files wastes an hour. The tell is the log's file stack: the last file opened before the error is (work/main.aux.

build.sh now takes an atomic mkdir work/.buildlock and refuses to start rather than corrupting a running build. Diagnosed wrongly twice before that (2026-09-14/15): first as a truncated aux left by an interrupted run, then as an orphaned container. Neither. The cause both times was simply starting make again while an earlier build was still going — a build takes several minutes and prints nothing until it ends, so a running one is easy to forget. Check docker ps first.

A group that ends before its own \par loses its leading

Twice now: the plate captions, and the last paragraph of src/colophon.tex. \setstretch (like \small) is a paragraph-shape parameter read at the \par that ends the paragraph, so {\setstretch{1.2} text} or \begingroup\setstretch{1.2} ... \endgroup sets the final paragraph at the outer leading. End the paragraph explicitly — \par before the closing brace or \endgroup.

tools/reprocheck.py — the reproduction check that found it

Bag comparison, not sequence: every token in src/pages/*.tex must reach the built PDF at least as often as it appears in the source. Order-insensitive, because endnotes and the Contents move text about. Run it after any build that changes the preamble:

python3 tools/reprocheck.py [pdf]

It de-TeXes the page files (macro table in the script says which argument of each macro carries copy) and compares against pdftotext -layout. Two extraction artifacts are reported separately as "merged into a neighbour", not as losses: poppler glues an endnote superscript onto the word before it (em.+55), and it drops the thin space between the two Bengali forms in a Scheme-of-Transliteration cell. -layout is required — without it poppler rejoins a compound broken at its own hyphen (Staats-Druckerei), which this text does constantly, and ~80 false positives appear. A clean run prints 0 tokens short.

Environment

Runs from this directory with Claude Code; tools via make (Docker). CLAUDE.md has the rules and macro table. The Makefile exports UID/GID (zsh does not). make check and make build must both pass at the end of every batch.

Layout

ross-thesis/
  PLAN.md  CONTINUATION.md  QUESTIONS.md  manifest.tsv  build.sh
  src/preamble.tex  src/colophon.tex  src/errata.tex  src/main.tex (generated)  src/pages/pNNNN.tex
  tools/prep.sh  tools/ocr2tex.py  tools/crop_plate.py  tools/crop_inline.py  tools/check.py  tools/setrow.py  tools/reprocheck.py
  work/lines.py  work/glyphs.py (scratch helpers for the specimen workflow)
  fonts/TiroBangla-*.ttf   plates/pNNNN.png  plates/inline/pNNNN-*.png  plates/logo-proquest.png
  scans/ work/ (scratch; excluded from tarball)   older-versions/   ross-1988-retypeset.pdf