docs: make the orientation true again, for the next instance to inherit

The same pass this repository ran on 2026-07-25, and needed again: the last six
commits moved things the documentation had not caught up with. Everything below
was checked against a command's output, not against memory.

**The specification did not know about `words:`** — nought occurrences. § 7.12
described a calendar whose own vocabulary was English and unchangeable, which
stopped being true two commits ago. It now has a *Sprache des Blattes* section
covering `words:`, `font:` on a document generator, and the Latin-1 boundary that
makes the second necessary. This project's rule is to update the spec in the same
breath as the code; that did not happen, and this is the repair.

**§ 15's question 2 changed shape and now says so.** "Whether a font must be
shipped" was theoretical while a missing glyph in a generator's text went
unnoticed and a document could name no font. Both are fixed, so stage 2 is
reachable everywhere and the question is now whether it is *reasonable to ask* —
a decision for use, but one the user is finally shown rather than handed a box.

**Decision 53** records the three changes as the one decision they were, in the
order that is the actual lesson: refusing the silent failure before supplying the
remedy left a window in which a Polish calendar was less usable than before.

**Three stale numbers in `docs/CLAUDE.md`**, each verified against the tree:
the specification is ~2800 lines and not ~2330; it covers thirteen generators and
not ten; there are fifty-three implementation decisions and not fifty-one. Also
checked and correct: 1231 tests, 19 presets (twelve generators, seven `lines`
papers — the eighth `generator: lines` is a *section* inside `notebook-a4`),
18 examples, 45 modules, two tags, `0.11.1`.

The Done table gains the row for everything after the audit, the module table
names all three functions both page paths now share, and the release paragraph
says two releases rather than one first tag.

The habit about probes is rewritten with its real count: five times over these
sessions I concluded something was unchecked, and five times the probe was wrong
and the code was right — including twice while writing *this* commit's checks.

1231 tests green, ruff clean, every committed example byte-identical.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: DocAtPrompt

docs/CLAUDE.md

@@ -126,6 +126,7 @@ edge, so a tick now carries its position *and* its value separately. `unit` says
 | post-M9 `notebook` | the **second document generator**, and the first that composes blades (§ 7.13, decision 50): sections, each filled by an ordinary generator, with a linked contents page, opt-in dividers and a cover. The seam: a `DocumentPage` may carry a `Fill` (generator name + validated config) instead of marks, and **the handle calls the blade** — `pages.document_page_marks` is the one function that says what is on a page, used by the writer, the capability pre-flight and the media check. A section is a definition in miniature, validated by that blade's own `config_model` with the loader's context (so `px`/`%w` resolve there too). Document bands are now laid out **per page**, so `{page}` counts and a per-page `{section}` names the section; `calendar-a4` stayed byte-identical. `page_layout.Page` extracted from `calendar_layout` and shared. Preset `notebook-a4`, example `15-notebook` |
 | post-M9 band colour | a header/footer `Band` takes `background` (a full-width strip, band height only) and `text_color`, both default off — drawn in `layout_band` so both the blade and document paths get them; resolves the title-page contrast (§ 8.9) |
 | pulled forward into M1 | format table, presets, `check`, overwrite protection, placeholders — the M1 acceptance criteria needed them |
+| 0.11.x — after the audit | **`tests/test_geometry_readback.py`**: every blade's geometry measured out of a finished PDF against numbers taken from the definition, never from the code — a hexagon's every edge exactly `size`, ring radii against the cumulative series, a mandala mapped onto itself by a twelfth of a turn, a perspective fan concurrent within a millimetre, `form`'s 8 mm ruling identical on A4 and A5, log positions against `math.log10`. **`check_page_furniture`**: the frame is checked on both page paths, not only drawn on both. **`check_text_glyphs`**: a generator's text is checked for missing glyphs (§ 12 point 13), which no code had ever done. **`words:` and `font:`** on the document generators, so every word on the sheet comes from the definition and there is a way out of Latin-1 for it (decision 53). **`labels: none`** accepted on `grid` and `polar`, the spelling § 7.10 documents |
 | 0.10.0 release-readiness | the audit above, in five commits: three degenerate values that hung or crashed instead of refusing; **`page_furniture`**, so the document path carries the page model (duplex, border, hole marks, ruler, stamp, background) instead of dropping it silently, plus three keys refused by name and a fourth (a multi-sheet blade in a notebook section) refused for now (decision 52); five messages made actionable and `--version` added; a ruler tick that ran off the edge, a ruler font that was never opened, an unreachable bookmark guard; and twenty-six stale comments |
 
 ### Not done
@@ -146,8 +147,9 @@ edge, so a tick now carries its position *and* its value separately. `unit` says
 
 **Two more things wait, and neither of them is code:**
 
-- ~~**PyPI.**~~ **Published on 2026-07-26 — `ctrlgrid 0.11.0`, MIT, Python ≥ 3.11,
-  wheel and sdist, under the repository's first tag `v0.11.0`.** M1's first
+- ~~**PyPI.**~~ **Published on 2026-07-26 — `ctrlgrid`, MIT, Python ≥ 3.11, wheel
+  and sdist. Two releases so far: `v0.11.0`, the repository's first tag ever, and
+  `v0.11.1` for the metadata it should have carried.** M1's first
   acceptance criterion is met at last, and measured rather than assumed: `uvx
   ctrlgrid millimeter-a4 --pages 3 -o out.pdf` installs from the network and
   produces a sheet whose MediaBox reads 210.000 × 297.000 mm with a 1.000 mm line
@@ -234,7 +236,8 @@ of the five serious findings were there. That is where to look first.
 ## Read this before writing anything
 
 **The specification is the source of truth, and it is unusually complete** —
-~2330 lines covering the architecture, all ten generators, the page model,
+~2800 lines covering the architecture, all thirteen generators — eleven blades
+and two documents — the page model,
 validation, milestones and the decisions that were explicitly rejected and why.
 
 Most questions that come up while implementing are already answered there, with
@@ -251,7 +254,7 @@ work around it.
 
 **Where the specification was genuinely silent**, the resolution is recorded in
 [`implementation-decisions.md`](implementation-decisions.md) —
-fifty-one of them so far, each with the section it belongs to and the
+fifty-three of them so far, each with the section it belongs to and the
 reasoning. Read it before changing a default; several look arbitrary and are not.
 
 ## Language split
@@ -278,7 +281,7 @@ and knows nothing about margins.
 | `axes.py` | `AxisPeriod` — what the handle needs from a blade for § 8.3/§ 8.5 |
 | `model.py` | pydantic sections; `Section` base with `extra="forbid"` + `deferred` |
 | `loader.py` | YAML → `Document`; formats, presets, devices, name lists |
-| `pages.py` | `Geometry`, page loop, placeholders, `preflight`, `build`, `snap: pixel`, and **`page_furniture`** — the one function that says what the handle draws around a page's own marks, used by the blade path *and* the document path |
+| `pages.py` | `Geometry`, page loop, placeholders, `preflight`, `build`, `snap: pixel`, and the three functions both page paths share: **`page_furniture`** (what the handle draws around a page's own marks), **`check_page_furniture`** (whether it fits) and **`check_text_glyphs`** (whether the font can draw it). Each exists because the blade path had it and the document path did not |
 | `frame.py` | header/footer layout, border, background, hole marks, the edge ruler's marks and its fit check, stamp |
 | `labels.py` | counting patterns `n`/`a`/`A`, explicit lists (§ 7.10) |
 | `clip.py` | Liang–Barsky in exact rationals, shared by `perspective` and slanted `lines` |
@@ -356,7 +359,10 @@ the pre-flight, never while pages are being written (§ 12 point 13).
 - **One coherent block per commit**, message explaining the decisions rather
   than the diff. `git log` is where the open-question resolutions live.
 - Verify against a real PDF, not just unit tests — `tests/pdfread.py` reads
-  geometry back out.
+  geometry back out (lines, dash arrays, whole subpaths and fitted circles), and
+  `tests/test_dimensional.py` plus `tests/test_geometry_readback.py` are where
+  that is done for every blade. **Every expected number in them comes from the
+  definition or from a sentence of the specification — never from the code.**
 
 ## Where to start
 
@@ -394,8 +400,13 @@ this project runs on — claims are backed by a command's output, never asserted
 
 ## What to do next
 
-**Nothing in the specification is unbuilt, and nothing from the 2026-07-25
-handover is left.** That handover
+**Nothing in the specification is unbuilt.** As of 2026-07-26 the tool is also
+*published* — the sessions since 0.9.0 audited it as a stranger meets it, read
+every blade's geometry back out of a finished PDF, and put it on PyPI. The user
+has said there is more to do and will say what; nothing below is a queue they
+have asked for, it is the state to start from.
+
+Nothing from the 2026-07-25 handover is left either. That handover
 ([`docs/superpowers/plans/2026-07-25-next-steps-handover.md`](superpowers/plans/2026-07-25-next-steps-handover.md))
 carries a *done* note on each of its phases saying what was built and what was
 decided against; the designs and plans behind the larger ones live beside it in
@@ -410,7 +421,7 @@ supply:
 |---|---|
 | a device | the empty `quirks` (decision 31), and the rM2 figures nobody has measured on the device — the Paper Pro is owner-checked, the rM2 is not |
 | a pair of scissors | print `examples/16-net-tray.yaml` and `box-tuck-a4` at 100 %, cut, fold. The tray checks the geometry; the tuck-top also checks the thickness rule. **No test can do this**, and until someone does, the tests only agree with themselves |
-| use | the three § 15 questions below, and 1.0.0 itself. 0.11.0 is *published*, so the DSL can now meet someone other than the test suite — which is the one thing 1.0.0 has always been waiting for |
+| use | the three § 15 questions below, and 1.0.0 itself. The DSL can now meet someone other than the test suite, which is the one thing 1.0.0 has always been waiting for |
 
 If a new feature *is* wanted, the recipe has not changed since M1: a design
 settled with the user first when there is a real fork in it, then a plan, then
@@ -431,11 +442,15 @@ Three habits earned their place the hard way and are worth keeping:
   the same centre. The third quantity, the grid, was the only one that could
   have contradicted it, and it was not looked at until the user did.
 - **A probe that does not fire proves nothing either — suspect the probe first.**
-  During the release audit I twice concluded a check was missing, and twice the
-  check was there and my test was wrong: a name list without `{name}` in any band
-  draws no text, so the glyph check has nothing to refuse; and a pipeline's `$?`
-  is `head`'s exit code, not the tool's. Both would have been reported as bugs.
-  Before writing "X is not checked", make the probe fail on purpose.
+  This happened *five* times over the release sessions, and not once was the code
+  at fault: a name list without `{name}` in any band draws no text, so the glyph
+  check had nothing to refuse; a pipeline's `$?` is `head`'s exit code, not the
+  tool's; a `--seed` that never reached the loop made the maze booklet look like
+  a regression; Vera has 256 glyphs and no Polish, so "a font file fixes it"
+  failed for the fixture's reason and not the check's; and a grep for a claim in
+  `docs/CLAUDE.md` missed it because the sentence wrapped across two lines.
+  Every one of them was about to be written up as a bug. **Before writing "X is
+  not checked", make the probe fail on purpose.**
 
 A specific paper aeroplane was considered and **refused**: it is a drawing, not
 a structure, and § 2 rules out a drawing language.
@@ -465,3 +480,12 @@ than on a design answer — **q4 was put to the user in July 2026 while building
 the slanted families and deliberately left out**: the angle is what unlocks
 calligraphy guides and origami pre-creasing, the unit would be a convenience on
 top of it (§ 7.1).
+
+**q2 changed shape on 2026-07-26 and is worth re-reading in § 15.** It used to be
+theoretical, because a missing glyph in a generator's text was never noticed: the
+check ran over the bands only, a Polish month name printed as a box, and a
+document generator took no font at all — so the documented remedy did not exist
+for it. Both are fixed (decision 53). Stage 2 is now *reachable* everywhere, and
+the question is no longer "does it work" but "is it reasonable to ask, or should
+a broad OFL font travel with the tool?". Still for use to answer — but now the
+user is shown the choice instead of a box.

docs/implementation-decisions.md

@@ -1038,3 +1038,43 @@ redrew every maze because unrelated furniture shifted `page.index`. § 7.13 says
 nothing about what a per-section sheet plan should mean, so it is refused until
 it does rather than answered wrongly. Honouring it per section is a real design
 question, not an oversight, and is left open.
+
+
+## 53. The language of the sheet: the tool receives words, it does not know any (§ 7.12, § 7.8, § 10.3)
+
+Three changes on 2026-07-26 that are one decision, and they came out of a
+question the user asked rather than out of the audit: *can any Latin-script user
+set this up freely now?* The answer was no, in three places.
+
+**A generator's text was never checked for missing glyphs.** § 12 point 13 asks
+the pre-flight to measure "Kopf-, Fuß- **und Beschriftungstext**" and then
+whether every glyph is present; `missing_glyphs` was called from exactly two
+sites, both of them the bands. So `months: [styczeń, …]` passed `check`,
+reported a successful run, and put `stycze` and a box on the paper — the silent
+almost-right of § 5.1, in the one place the tool had a query built to prevent it.
+`check_text_glyphs` collects **characters as a set per font** over the marks —
+page 0 for a blade, every page for a document (decision 49's reason: a month name
+only appears on its own month's pages) — so a 386-page planner with tens of
+thousands of glyphs is checked against a few dozen distinct ones and the run
+still takes a second.
+
+**The tool wrote its own words on the user's sheet.** § 7.12 had applied § 7.8 to
+the *names* and not to the vocabulary, so a German calendar came out with
+`Jänner` under `Index Year Month Notes`. `words:` names the nine, defaulting to
+today's English so no definition moves. Rejected, and worth recording: shipping
+month/weekday tables for common locales. It would be a translation layer by
+another name (§ 3.4), it would need maintaining, and the very first user whose
+language was not in it would be worse off than today.
+
+**A document generator could not name a font**, which turned the first change
+from a fix into a trap: refusing a character with no way to draw it is worse
+advice than it looks. `CalendarConfig` and `NotebookConfig` take a `font:` like
+every blade, reaching the pages through `page_layout.Page`, which now carries the
+family token — so all seventeen call sites in `calendar_layout` and the
+notebook's three follow without knowing they do. **One font per document**: there
+is nothing on these pages that wants two, and the alternative (a font per view)
+would be a typographic system nobody asked for.
+
+The order these were done in is the lesson, not the code. Refusing the silent
+failure first and supplying the remedy second left a window in which a Polish
+calendar was *less* usable than before. Both belong in one step.

docs/pflichtenheft-vorlagengenerator.md

@@ -1230,6 +1230,35 @@ auf dem Inhaltsverzeichnis, wird mitgezählt, wenn geprüft wird, ob die Seite
 passt, und bleibt leer, wenn keine angegeben ist — der Kalender erfindet keine
 Bedeutungen. Damit ist § 7.12 vollständig.
 
+#### Die Sprache des Blattes (ergänzt 2026-07-26)
+
+§ 7.8 entscheidet den Grundsatz — die Beschriftungen gehören zur Sprache des
+Formulars, also nimmt das Werkzeug sie entgegen und erfindet sie nicht — und
+§ 7.12 hatte ihn auf die **Namen** angewandt (`months`, `weekdays`, `label`)
+und auf das eigene Vokabular nicht. Ein deutscher Kalender kam deshalb gemischt
+heraus: `Jänner` unter einer englischen Navigationsleiste, daneben `Contents`,
+`Full-year overview` und `Half-year 1`. Das sind Wörter des *Werkzeugs* auf dem
+Blatt des Nutzers, und genau die schließt § 7.8 aus.
+
+**`words:`** benennt sie: `index`, `year`, `month`, `week`, `notes` (die
+Navigationsleiste), `contents`, `full_year_overview`, `half_year` und
+`full_year`. Jede Vorgabe ist das Englisch, das bisher gedruckt wurde, also
+ändert der Block keine bestehende Definition. Beim Notizbuch heißt das eine Wort
+`contents_title`, und der Rücklink auf einem Trennblatt folgt ihm.
+
+Das ist **keine Lokalisierung** (§ 3.4) und es wird weiterhin keine
+Übersetzungstabelle mitgeliefert: Das Werkzeug nimmt Wörter entgegen und kennt
+keine.
+
+**`font:` an einem Dokument-Generator.** Die Grenze dahinter ist die Glyphen-
+abdeckung (§ 10.3): Die Standardschriften reichen bis Latin-1, decken also
+Deutsch, Französisch, Spanisch, Italienisch, Portugiesisch und die nordischen
+Sprachen ab — aber nicht Polnisch, Tschechisch, Ungarisch, Türkisch, Rumänisch
+oder Kroatisch. Deshalb nimmt ein Dokument-Generator wie jede Klinge eine eigene
+Schrift entgegen, **eine für das ganze Dokument**; ohne sie lehnt die
+Vorabprüfung ein Zeichen ab, das die Schrift nicht zeichnen kann (§ 12 Punkt 13),
+statt ein Kästchen zu drucken.
+
 ### 7.13 `notebook` — verlinktes Notizbuch aus Abschnitten (Dokument-Generator)
 
 Ein PDF, das mehrere Papiere trägt: vierzig gepunktete Seiten zum Journalen,
@@ -2666,6 +2695,17 @@ Bewusst noch offen:
    Latin-1 ab. Zeigt sich, dass Namen mit `ł`, `ğ` oder `ő` häufig genug sind,
    um Stufe 2 zur Pflicht zu machen, wäre eine mitgelieferte OFL-Schrift der
    bequemere Weg. Erst nach Erfahrung entscheiden, nicht auf Verdacht.
+
+   **Der Boden dafür ist seit 2026-07-26 gelegt, die Frage bleibt offen.** Bis
+   dahin war sie theoretisch, weil ein fehlendes Zeichen in Generatortext gar
+   nicht auffiel: `missing_glyphs` lief nur über die Bänder, ein polnischer
+   Monatsname wurde still als Kästchen gedruckt, und ein Dokument-Generator nahm
+   überhaupt keine Schrift entgegen — die dokumentierte Abhilfe existierte für
+   ihn also nicht. Beides ist behoben (§ 7.12, § 12 Punkt 13). Damit ist Stufe 2
+   für jeden Fall *erreichbar*, und die Frage lautet jetzt sauber: ist sie
+   **zumutbar**, oder soll eine breite OFL-Schrift mitreisen? Das entscheidet
+   Gebrauch, nicht Verdacht — aber es entscheidet ihn jetzt an einer Stelle, wo
+   der Nutzer die Wahl überhaupt gezeigt bekommt.
 3. ~~**`reportlab`-Lizenz** am Repository gegenprüfen (BSD-3-Clause erwartet).~~
    **Geklärt (2026-07-24):** reportlab 5.0.0 steht unter **BSD-3-Clause** — der
    Lizenztext (`reportlab-5.0.0.dist-info/licenses/LICENSE`) führt exakt die drei