tonus 0.1.8 → 0.6.1

package/CHANGELOG.md

@@ -2,6 +2,669 @@
 
 All notable changes to tonus. Newest first.
 
+## Unreleased
+
+## 0.6.0 — 2026-08-16
+
+The zodiac learns what it means, and the hand stops sitting on its own floor.
+
+### Added
+
+- **`ZODIACA` — the doctrine of the twelve signs.** `caelum` answers WHERE a
+  body is; this answers what a medieval reader brought to that position: the
+  element and its Galenic humor, the quality, the ruling and exalted planets,
+  and the member of the zodiac man the sign governs. `ZODIACA[body.zodiac]` is
+  the join, so a sky and its meaning are one lookup apart.
+
+  Deliberately no dates. When the Sun enters a sign is the ephemeris's business
+  and moves with precession — a table carrying "March 21" would be wrong for
+  most of the period this library models, and wrong differently every century.
+
+  The exaltation degrees are omitted (the sign is the resolution anything here
+  reads) and so are the lunar nodes, which are not tonus bodies. Five signs
+  exalt nobody, and that silence is the tradition's.
+
+### Changed — breaking
+
+- **`SIGNS` and `SIGNA` are no longer exported.** `ZODIACA` carries each sign's
+  own `sign` and `signum` beside its doctrine, so it answers both what a sign
+  is called and what it means. The bare name arrays go back to being internal
+  to the engine that computes a longitude, which is the only place they were a
+  lookup rather than a listing.
+
+  ```js
+  // before
+  import tonus, { SIGNS, SIGNA } from "tonus";
+  SIGNS[body.zodiac];              // "Capricorn"
+  SIGNA[body.zodiac];              // "Capricornus"
+
+  // after
+  import tonus, { ZODIACA } from "tonus";
+  ZODIACA[body.zodiac].sign;       // "Capricorn"
+  ZODIACA[body.zodiac].signum;     // "Capricornus"
+  ```
+
+### Fixed
+
+- **The Manus line reads the note that was clicked.** The subheader found its
+  row by pitch, and `find(r => r.spn === spn)` returns the FIRST note of the
+  chant at that pitch — so clicking the second E of a porrectus flexus
+  described an E in a torculus eleven notes earlier, with that note's neume,
+  joint and Guidonian name.
+
+### Site
+
+- **The chant is lifted onto the hand automatically, by whole octaves.** Almost
+  nothing fits the gamut as printed: measured over 400 chants, 32 sit inside
+  their mode's gamut and 346 sit below it, so a piece read straight lands on
+  the bottom joint and stays there. The lift is computed from the chant's own
+  lowest note, which takes notes landing on a joint from 68.8% to 99.9%. The
+  `octava` button was the manual version of this and is gone.
+- **The three sung hours become one Officium segment**, with all eight hours in
+  its popover — Matins and Vespers on, the little hours and Compline there to
+  be asked for.
+- **The address bar keeps `cantus` and nothing else.** Eleven parameters were
+  mirrored there, each written the moment it moved off its default, which made
+  the bar a running transcript of a session.
+- All three analysis tracks open together; the Similes subheader reports the
+  company a melody keeps rather than describing the list; the temperament
+  slider steps a tenth of a cent, the precision its readout prints.
+
+## 0.5.0 — 2026-08-12
+
+Rendering, mostly — a day of looking at real chant on a real page and fixing
+what the page showed.
+
+### Added
+
+- **A third analysis track: `"prosodia"`** — how the melody treats the word.
+  One tent per word (the hairpin's top edge), its apex over the accented
+  syllable, the landing at the peak in the liturgical red — filled for arsic
+  (struck), an open ring for thetic (deferred). Below it a fence on a rail:
+  a stem per spoken syllable, a flat dash for a syllable recited on the tenor,
+  and a block per melisma — as wide as its real extent, as tall as its note
+  count, connected melismas joining into one ridge. It rides first in the
+  stack, directly under the lyric line it reads, and the site's vestigia strip
+  and tracks key carry it. Shaped across eight lab rounds
+  (`working/review/diagram-word-track-01…08.html`).
+- **The rubrica reservation, amended.** Ruled 2026-07-29 as the mode line's
+  alone, the liturgical red now belongs to the claims: the tonarium's mode
+  line and the prosodia's accent dots. The precedent is the score's own text
+  apparatus — the dropcap and annotations were always rubricated, and in the
+  books red is the word's colour.
+
+### Changed — breaking
+
+- **The layout options are `width` and `scale`.** `padding`, `noteScale`,
+  `systemGap`, and `custos` are gone. Nothing ever set them — not the docs site,
+  not the 28 lab plates, not the 13 stress pieces — so they were surface without
+  use. They are now constants chosen to look right at every scale, and the
+  custos appears whenever a system wraps, which is what a chant book does.
+
+  `scale` replaces `staffHeight`: `"small"`, `"normal"` (default), `"large"`, or
+  a staff height in px for fitting a known column. A caller decides how big the
+  chant should be, not how tall its staff is in pixels.
+
+  ```js
+  inscriptio(score, { width: 900, scale: "large" })
+  ```
+
+  The page margin deliberately does NOT scale with it — a margin belongs to the
+  page, not the notation, and scaling it gave a large chant *less* usable width
+  than a small one (89% of a 900px canvas against 93%). The air between systems
+  does scale, since flat 24px held the system pitch at 135px whether the staff
+  was 30 or 56.
+
+- **`inscriptio`'s look options are one `theme` object.** `fonts`, `noteColor`,
+  `staffLineColor`, and `rubricaColor` are replaced by
+  `theme: { fonts, colors }`. These travel
+  together — a caller setting a lyric face is usually setting a whole look — and
+  a house style is worth naming once and passing everywhere.
+
+  ```js
+  // before
+  inscriptio(score, { fonts: { lyric: "Junicode" }, rubricaColor: "#801", staffHeight: 48 })
+  // after
+  inscriptio(score, { scale: 48, theme: {
+    fonts: { lyric: "Junicode" }, colors: { rubrica: "#801" },
+  } })
+  ```
+
+  `fonts` keeps its four roles unchanged, `dropcap` among them: a book's
+  initial is very often not its lyric face, and the two stay separate.
+
+### Added
+
+- **A note knows the diphthong it belongs to.** `Context.diphthong` and
+  `ChantTabulaRow.diphthong` report the pair a sung vowel is part of — `ae` ·
+  `oe` · `au`, and `ui` in the cui/hui stems — or null. `vowel` is unchanged: it
+  stays the NUCLEUS, the single vowel a singer sustains, because that is what
+  the analysis passes key on. The pair rides beside it for anyone rendering the
+  off-glide.
+
+  It has to come from here, because **only the syllabifier can tell a diphthong
+  from a hiatus: the accent decides.** `cae` is one syllable — nucleus `a`,
+  gliding toward `e`; `sa-é` is two, the same letters with the accent on the
+  second and no glide at all. Downstream, given only the lyric, the two are
+  indistinguishable. A scan for the letter pair also misreads `quae`, where `qu`
+  is a consonantal glide: the nucleus is `u`, and the `ae` is not the sung pair.
+  Over the Graduale that is 71 distinct syllables — 1,526 sung notes — a naive
+  reading gets wrong, and 192 syllables over the whole corpus. The detection
+  reuses the syllabifier's own `DIPHTHONGS` set, so the word means the same
+  thing here as it does when a word is split.
+
+- **The ink is themable from CSS.** Colours now reach the SVG as custom
+  properties with the render's own value as the fallback —
+  `fill="var(--tonus-note, #111)"` — so a host stylesheet can retheme a drawn
+  chant without re-rendering it, while a file opened on its own still shows the
+  ink it was drawn with. Three properties: `--tonus-note`, `--tonus-staff-line`,
+  `--tonus-rubrica`. The emitter already carried semantic classes (`note`,
+  `lyric`, `dropcap`, `custos`, `episema`, `divisio`, `clef`, `mora`, `ictus`)
+  but an inline `fill` beats any stylesheet rule, so none of them could be
+  styled. `scale` deliberately stays outside the theme for the same reason it
+  cannot be a CSS property: line breaking consumes it long before a stylesheet
+  sees the output.
+
+### Fixed
+
+- **An accidental was emitted as a sounding note.** GABC's `fx` means "F is flat
+  from here" — a mark drawn on the staff, not a pitch to sing. The parser set the
+  state correctly and then fell through and pushed a note anyway, so `A(fxfg)`
+  returned three notes for two and duplicated the pitch. 1337 markers across 426
+  Graduale chants: 0.97% of every note in the book was a phantom, at a wrong
+  pitch, inventing a unison before each one and inflating every count, interval
+  and analysis downstream.
+
+  The SIGN still draws, and now on the note that follows it — where the books
+  print it. That needed a new field: `accidentalSign` is what to draw, separate
+  from `accidental`, which is the note's own alteration. In `fe(jx)cit(ih)` the
+  flat is printed before the I while it governs J; the emitter had been reading
+  the alteration and so drew nothing once the phantom was gone.
+
+- **A written flat could come back spelled as a sharp.** Pitch spelling was
+  derived from the pitch class alone, against a preferred-flat set that omitted
+  pc 1 and pc 6 — so a GABC `x` (a flat) landing on those degrees returned
+  `C#`/`F#` rather than `Db`/`Gb`, and reported `accidental: +1` for a source
+  that wrote a flat. The pitch was always right; only the spelling and its sign
+  were wrong. `toPitch` now takes an optional spelling preference and `notatio`
+  passes what the source wrote. The Graduale writes only the bmolle (1821 B-flats
+  and no sharps at all), so the corpus never reached the gap.
+
+- **Moderna honoured no note colour at all.** It hardcoded `#111` in seventeen
+  places while quadrata threaded the option, so a caller theming the ink saw one
+  species change and the other not.
+
+- **`staffHeight` means the same thing in both species.** Moderna's staff was
+  a hardcoded constant, so the option moved quadrata and did nothing here: a
+  request of 30, 40 or 60 left moderna at 7.4px every time. Every geometric
+  constant now derives from it through one metrics factory.
+- **The canvas is the width you asked for.** It was `max(systemMaxX)` —
+  whatever the widest system happened to reach — so a requested 900 came back
+  915, 986, 1074, 1203 by chant, and a host applying `max-width` shrank each
+  differently. That is why the same page showed one chant's notation a third
+  smaller than the next's.
+- **Lyrics count toward a system's extent.** A syllable is centred on its
+  note, so half of a final wide syllable always hung past the canvas and was
+  cut off. The layout already tracked the lyric's right edge for its collision
+  check; the width calculation never asked.
+- **The custos is a custos.** It was drawn as a shrunken punctum on the
+  authority of a comment saying no custos glyph was baked — Bravura's have
+  been baked all along. The real glyph is a hooked note whose stem points
+  toward the pitch it announces. It also floated 41px past the last note,
+  having been placed after the divisio's trailing air; and it is now
+  suppressed after a full stop, where drawing it put two marks in one place
+  and read as a heavy double barline.
+- **Analysis tracks scale with the staff in moderna**, and a track that SPANS
+  notes now reaches their ink rather than their anchors — the emitters had
+  measured the ink all along and the mapping dropped it.
+- **Capitals are measured as capitals.** The lyric width estimate was a flat
+  per-character average, and chant sets its opening word in capitals; "CAn"
+  was estimated at 22.9px against a real 27.9, so opening syllables collided
+  with what followed.
+
+### Changed
+
+- **More air.** Syllable and word spacing widened in both species, the
+  staff-to-lyric gap from 21px to 28, and moderna's note advance retuned so a
+  chant takes about one system more than quadrata rather than twice as many.
+
+## 0.4.5 — 2026-08-04
+
+The rubric true-up, the era view, and a long pass of making the library say
+what it means. The 2026-07-27 review found the Kyriale selection chain leaking
+past its own gates; the fixes landed with regression tests, the corpus pipeline
+was corrected and regenerated behind them, and the calendar's and the corpus's
+`before` arguments were reconciled into one composable view. The shelf widened,
+the office settled on one cursus, the cadence catalogue was re-mined over the
+sung corpus, and a new verb — `census` — measures every chant against the
+corpus that holds it.
+
+Then the tidying, most of it prompted by things that turned out to be measured
+rather than assumed. The salicus took Cardine's correction and, in the process,
+stopped being confused with the Solesmes ictus — two nearly disjoint sets under
+one name. The cadence catalogue learned its own denominator and joined itself
+once instead of per consumer. The appendix roughly doubled and was renamed to
+follow the register rule. The corpus ledger stopped needing three numbers to
+say how much chant there is. And the Apel formula matcher came out: what
+shipped could not use the catalogue it shipped with, and saying so was better
+than leaving a public field that answered `null`.
+
+`docs/` is a published site now, which is where 0.5 is headed.
+
+### Removed
+
+- **BREAKING — `Score.formulas` and the Apel centonization matcher are out.**
+  The catalogue of standard phrases, the per-phrase matcher, `FormulaMatch`,
+  `Formula`, `FormulaSlot`, and `formulaeFor` are gone from the library. What
+  shipped was one mode's alphabet (mode-5 Graduals) behind a matcher whose
+  per-phrase window structurally could not fire on 40 of its 53 formulae —
+  Apel tabulates a chant as a sequence across a whole respond, and matching at
+  that scale is a different algorithm that was never built. The honest state
+  was a public field that answered `null` for nearly every chant, and an
+  appendix of research the library could not yet use.
+  The census keeps its `formulas` field group: those four floats are baked
+  upstream in tonus-corpus and describe the corpus as measured, whatever tonus
+  does with them. The transcriptions, photographs, tooling and verification
+  reports are kept in the workshop, not deleted — the reading of Apel stands on
+  its own, and the pipeline is there if the matcher is ever built properly.
+
+### Changed
+
+- **BREAKING — the shelf reports ONE number, and the Kyriale leaves it.**
+  `corpus()` reported `count` (book listings, a chant in two books counted
+  twice) beside `distinct` (chants), which made the headline figure depend on
+  how many books happened to print the same melody — a reader had to know that
+  to read either. Now `count` is **how many chants tonus holds**, each counted
+  once, and `listings` is the shelf's length; `listings - count` is the
+  overlap. `genera` and `modes` sum to `count`, describing the same population
+  the headline does.
+  The Kyriale is no longer an eleventh book. There is no Kyriale in GregoBase:
+  `ky` and `gr` are the same source, partitioned by the extractor so the
+  ordinary can be routed to per-ordinary codes, so listing it beside its parent
+  counted the Graduale twice and its `total: null` read as "unmeasured" when
+  there was no separate book to measure. `ky` is not a `ChantSource` and not a
+  row in the shelf — which is 10 books, 2,187 chants, 2,767 listings.
+  Nothing became unreachable. `ordinarium({ feast })` is untouched, and the
+  ordinary is now queryable directly: **`cantus({ ordinary: "ky" })`** returns
+  all 31 Kyries, composes with the other filters, and the Kyriale's
+  bibliographic record still rides every one of its chants. A plain search does
+  not sweep the ordinary in — you ask for a Kyrie rather than stumbling onto
+  one.
+- **BEHAVIOUR — the salicus takes Cardine's correction, and stops being
+  confused with the Solesmes ictus.** Two changes, each ruled and each
+  measured separately over the sung corpus (2,887 chants, 28,498 phrases).
+  The second — redefining what counts as a salicus at all — was ratified on
+  its own terms, since it decides what the word denotes everywhere in tonus,
+  not just how one note is weighted.
+  **Detection.** A salicus is "at least three ascending notes in which the
+  next-to-last is an oriscus" [biblio: cardine-semiology, ch. 16]. tonus had
+  been calling any ascending run with the editorial ictus a salicus — 2,795
+  groups, of which **36 (1.3%) actually carried an oriscus** — while missing
+  188 of the corpus's 226 real salici, which classified as `scandicus`. Two
+  nearly disjoint sets under one name. Now: `salicus` 2,795 → **226**, exactly
+  the chants that meet Cardine's definition; `scandicus` 498 → 2,501, and the
+  4-and-5-note ictus ascents (566) join `compound`. An ictus-marked ascent
+  with no oriscus is a scandicus that Solesmes marked for rhythm, and the mark
+  is still readable on `context.ictus`.
+  **Weighting.** The printed editions lengthen the oriscus; the manuscripts
+  show the principal note is the one immediately following it. The 1.3×
+  prolongation moves from the next-to-last note to the **summit** — which is
+  the last note at any length, since the oriscus is next-to-last by
+  definition, so the 4- and 5-note forms need no separate rule. The salicus
+  stays arsic: tension toward the summit is the arsic gesture, so Cardine
+  strengthens that rule rather than contradicting it.
+  **Downstream:** 208 of 28,498 phrases (0.73%) change `rhythmicType`,
+  overwhelmingly toward VIII. This is the one point where tonus's rhythmic
+  layer departs from Mocquereau and Suñol, and `ir.ts` now says so at its
+  header.
+  **The limit, stated plainly:** tonus sees only what the transcription marks.
+  Bevenot's own example — the mode-6 _Requiem_ introit's fa-sol-la — carries
+  the ictus and no oriscus in GregoBase, so it still reads as a scandicus. He
+  is reading the manuscripts; the corpus is a printed edition that resolved
+  the oriscus away. Recovering those wants the sources, not a looser rule.
+- **The tonarium's cadence label says how typical a close is, not which
+  family it belongs to.** The bracket read `"2,0,-2 @0"` — the family's name,
+  which a reader could not weigh. It now reads `"×2.1"`: the family's share
+  within this chant's mode over its share of the corpus at large. Measured
+  across four books the figure spans ×0.46 to ×10.56, median ×2.22; a lift
+  below 1.0 prints too, since an atypical close is information. A mode-less
+  chant, or a family with fewer than ten occurrences in the chant's mode,
+  falls back to the plain corpus share. The key is not lost — each cadence
+  now draws inside a group carrying `data-cadentia`, the join back to
+  `CADENTIAE` and the provenance a margin gloss can print.
+- **BREAKING — five appendix tables renamed to match the register rule.**
+  `TEMPUS_NAME` → `TEMPORA`, `GRADE_NAME` → `GRADUS`, and the internal
+  `OFFICE_LABELS` / `ORDINARY_LABELS` / `MODE_LABELS` → `OFFICIA` /
+  `ORDINARIA` / `MODI`, the last three now public. The house rule is that
+  Latin names carry Latin content and English names carry codes or English;
+  the appendix had three spellings of "code → display string" and five tables
+  whose Latin values sat under English names. Names now follow the register of
+  their values, so `SEASON_LABEL` ("Advent") and `TEMPORA` ("Tempus
+  Adventus") are distinguishable by name rather than by memory. Only the two
+  `cal` tables were public before; the rest is new surface.
+- **`officium` throws on an unrecognised `hora`.** It returned `[]`, which
+  read as "no chants at this hour" rather than "there is no such hour" — the
+  same silent-nothing the unknown-query-key guard beside it already refused.
+
+### Added
+
+- **Every shelved book reports what it HOLDS.** The Nocturnale was the last
+  one answering `full: null`, which read as "not yet measured" — but the tally
+  had existed all along in its own extract. It holds **1,564** chants against
+  the 470 tonus ships, and it shares nothing: `unique` is all of them. That is
+  a measurement, not a gap. The crosswalk pairing a nocturnale chant with a
+  GregoBase twin is enrichment, a route to metadata, not a claim that two books
+  print the same chant, so counting those as shared would invent a
+  relationship. The shelf now holds 10,156 chants before the cut.
+- **Three genera that are reported but not shipped now have names.**
+  `Improperia` (the Good Friday Reproaches), `Antiphona Mariana` (Marian
+  antiphons outside the office cycle) and `Supplicatio` (litanies) appear in a
+  book's pre-cut `full` tally, where they had been printing as bare codes —
+  `genus: "su"` — while every other row carried a Latin genus. Being outside
+  the cut is not a reason to be nameless in a table tonus publishes.
+- **The cadence catalogue carries its own denominator, and joins itself.**
+  Every `CadentiaFamilia` gains `share` — its occurrences over ALL 28,481
+  phrase-ends, not over the 58.7% that cleared the table's floor, which would
+  have flattered every family in it. `CADENTIAE_POPULATION` ships beside it
+  with the same total per mode digit, so a family's **lift** in a mode is one
+  division: `(modes[m] / byMode[m]) / share`. The ratio itself is not baked —
+  export the vocabulary, not the arithmetic.
+  A `Cadence` now carries `finality`, the share of ITS family's corpus
+  occurrences that land at a final close, joined once in `notatio` instead of
+  by every caller rebuilding the index. It rides the cadence while `familia`
+  still does not, because the two differ in kind: the signature already IS the
+  family's name, but how often that family closes cannot be read off it — of
+  the 55 families landing on the final, 31 do not close, and their finality
+  spans 0.054 to 1.000. Detection stays pure; the corpus table meets detected
+  data in the score builder, where `MODES` already does.
+  Both cadence catalogues are now documented as what they are — _tradita_
+  (the treatises' figures, final cadences only) and _inventa_ (the corpus
+  tally, any target, and so the only account of medial closes) — with the
+  measured coverage of each and a worked lift example that was run before it
+  was printed.
+- **The appendix widened, so callers stop transcribing the library's own
+  vocabulary.** `HORAE` (the eight canonical hours, Matins first — the order
+  is the content, and `officium`'s validation reads the same list, so the two
+  cannot drift), `OFFICIA`, `ORDINARIA`, `MODI`, `SOURCES` (the book codes
+  `cantus({ source })` takes, with their bibliographic records), and
+  `CENSUS_GROUPS` / `CENSUS_ORDER` — the census field groups and the block
+  index, so asking whether a chant is censused stops needing a `try/catch`.
+  A table is admitted when a caller would otherwise type it out, because a
+  transcribed copy drifts and fails as wrong answers rather than as an error.
+  The appendix is now grouped by engine.
+- **The census distance rule is documented as a contract.** Similarity is
+  cosine per field group, never over the flat 225; `by: "all"` is the
+  equal-weight mean. Grouping is userland, so a caller pooling blocks is
+  computing a distance and must reproduce the rule or silently disagree with
+  `census()`. `census.md` now states it for callers with the three ways to get
+  a plausible wrong answer, and a worked example — pooling the 178 Communions
+  — that was run before it was printed, and reproduces `census()` exactly.
+- **The analysis tracks ship with `inscriptio`.** `tracks: ["chironomia"]`
+  draws the conducting hand's wave — arsic crests, thetic
+  troughs, pick-up loops, Pierik letters. `tracks: ["tonarium"]`
+  draws the melodic lane: the four maneriae rails, the melody
+  compressed to its ambitus, the mode line in rubrica (solid inflection,
+  dashed transposition). A cadence is the melody's own ending re-inked
+  at full strength — the same curve at the same width — landing on a
+  terminal node (filled when the family's measured finality in CADENTIAE
+  closes, open when it suspends), labelled beneath by its signature on a
+  light end-ticked bracket spanning the figure. One governing ink system
+  runs through both tracks: one black graded by stratum opacity (rubrica
+  is the mode line's alone), and one nib law — velocity as stroke
+  width — for every pressure-bearing line. Either track rides either
+  species, and both may ride one score: they stack in a fixed order, the
+  chironomia above the tonarium, whichever order they are asked for. The
+  two-register principle — the rhythmic band under the square notation,
+  the melodic band under the transcription — is the house default, not a
+  rule the renderer enforces. Confidence is
+  opacity; claims under 0.45 draw nothing. The tracks consume the same
+  note anchors the geometry contract exports — the score's own ink is
+  untouched, and the geometry is unchanged with tracks on.
+- **The duae species share one lyric setting and one opening.**
+  Quadrata's staff→lyric gap now equals moderna's (21px at the default
+  staff height, scaling with it), and both species default the lyric
+  weight to 518; an explicit `fonts.lyric.weight` still overrides.
+  Moderna now honours the front matter (`title`, `rubric`,
+  `annotation: "auto"`) as quadrata does; the official opening of a
+  tonus score is the incipit as `title` plus the auto genus/mode mark,
+  with no dropcap.
+- **The era view composes end to end.** `festum({ date, before })` stamps
+  the view year on the returned `Feast` (`feast.before`), and every day
+  verb — `proprium`, `ordinarium`, `officium` — serves the
+  same view without being told the year twice; an own `before`
+  overrides. Previously the day verbs' types promised `before`/`century`/
+  `cursus` (they extend `CantusQuery`) while the implementations diverged
+  three ways: `proprium` threw "unknown query key," `officium` and
+  `ordinarium` silently ignored them.
+- One admissibility rule for every door: `engines/chant/attest.ts`, a leaf
+  module shared by `cantus`, the day verbs, and the census. One spelling
+  at the door — `before`, a year.
+- Under a view, `ordinarium` **re-picks** over the admissible pool (the
+  Kyriale offers ranked alternatives by design); `proprium` and `officium`
+  degrade to silence — the corpus cut's evidence law.
+- **The shelf widens to eleven books.** Four more marked Solesmes office
+  books join — Antiphonale Monasticum Solesmense (`ams`), Psalterium
+  Monasticum (`psm`), Cantus selecti (`cse`), Chants of the Church
+  (`cot`) — admitted by the stated rule: Solesmes and rhythmically marked.
+  The shipped corpus is 2,887 rows over 2,187 distinct chants.
+- **`tonus.census({ id })`** — one chant measured against the corpus that
+  holds it: a per-group profile with typicality, `balance` (distance and
+  deviant groups), and nearest neighbours by per-group cosine (`k`, `by`,
+  `before`). One block of 225 float32s per shipped chant; see
+  `docs/census.md`.
+- **`corpus()` with no argument returns the whole shelf** — the rollup
+  plus every book's ledger — and every book carries `full`, the pre-cut
+  tally: the ledger of the cut, auditable rather than asserted.
+- **`CADENTIAE` joins the appendix** (with `CadentiaFamilia`): the
+  corpus-grounded cadence catalogue, mined from every phrase-end in the
+  sung corpus — the same population the census counts — keyed by shape
+  and arrival.
+- Test suites: `tests/ordinarium.test.mjs` (the selection chain end to
+  end) and `tests/era-view.test.mjs` (the view's whole contract).
+
+### Removed
+
+- **The `rite` option.** tonus assembles one cursus, the Benedictine: the
+  Roman office table carried no chant at all on 63.4% of its days, and its
+  little-hours psalmody had no other consumer, so `rite: "romanum"`
+  returned a cursus nobody sang. A stale `rite` key now throws
+  (`officium` rejects unknown keys) rather than being silently ignored.
+- **`matutinum` as a separate verb.** Matins is an hour of `officium`
+  (`hora: "matutinum"`), returned flat like every other hour; the
+  three-nocturn ordo shape remains unmodelled.
+- **`century`.** It was `before: N * 100` in different clothes; the two
+  spellings converged on the one that is a year.
+
+### Fixed
+
+- **"Requiescant in pace" was the dismissal on every green feria** (~20
+  days/year): `isRequiem` missed the bare "Requiescant" incipit AND the
+  last-resort appendix return was ungated. The same hole handed ferias an
+  ad libitum Gloria. The appendix now never reaches a day whose rubric
+  does not admit it; the Requiem stays reachable via
+  `ordinarium({ mass: 102 })`.
+- **The Gloria follows the day's rank rubric, not its season**: ferias
+  print none; a I-class feast inside Advent (Immaculate Conception) keeps
+  hers. The "Benedicamus Domino as in Mass II" borrow the book directs is
+  now real — ferias and penitential Sundays sing Mass II's dismissal.
+- **The paschal rubric is a time, not a rank**: removed from the solemn
+  set (an Eastertide Tuesday no longer leads with an appendix Kyrie;
+  Easter keeps Lux et origo every year), and BVM outranks paschal in
+  `rubricForDay`, so Cum jubilo is reachable in Eastertide.
+- **Credo V was never sung**: the off-year rotation's parity was coupled
+  to the two-year bias (2 divides 6). The off-years now advance their own
+  cycle — all six credos are heard across twelve years.
+- The leftover masses-1–9 preference on high feasts (precedence measuring
+  what the rubric should) — deleted; it defeated the appendix solemnity
+  boost on the class feasts it was built for and pinned high BVM feasts to
+  mass IX.
+- `centuryOf` off-by-one at exact century multiples: `before: 1100` now
+  admits the closed 11th century, as its own doc always claimed.
+- "Kyrie XVII C" refiled from a synthetic mass 100 to mass 17 (corpus-side
+  incipit parsing read the variant letter C as roman 100) — it now rotates
+  as Mass XVII's sibling printing.
+
+### Data
+
+- Regenerated from the corrected pipeline: `office-roman` (canticle
+  antiphons split positionally — the last SURVIVOR of matching is no
+  longer promoted to the Benedictus/Magnificat slot), `office-monastic`
+  (464 days; rubric-year variant files excluded from the merge; stub
+  redirects followed to their actual targets), `office-ferial` (honestly
+  monastic rows only; the tier-then-book tiebreak; Matins 94 → 98 of 128),
+  `commune-office` (rubric-variant sections no longer blended; the Introit
+  genre-cap hole closed), `kyriale`.
+
+## 0.3.0
+
+The vox-ectomy. The voice engine (`vox`, `chorus`, the formant tables,
+personae, consortia, liquescentia coda tables, spectrum, accordatio) leaves
+the public library for orreliquum, the private instrument it was always
+being built toward. tonus returns to what it is: theory, calendar, chant,
+notation, and the heavens — the algorithms, publishable and citable.
+
+### Added
+
+- **`accent` on `ChantTabulaRow`** — a boolean marking whether a note's
+  syllable bears the Latin tonic word-accent. The value already existed inside
+  the parse (`detectVowelAccent` drives articulation stress); it is now surfaced
+  as data on the tabula, so downstream analysis can read accent placement
+  without re-deriving it. Threaded `ParsedNote → Context → ChantTabulaRow`.
+
+### Removed
+
+- **`tonus.vox()` and `tonus.chorus()`** and every voice type
+  (`Vox`, `VoxInput`, `VoxParams`, `Vowel`, `Formant`, `Locus`, `Latinitas`,
+  `Persona`, `PersonaName`, `Coda`, `TuningLike`, `Chorus`, `ChorusOpts`,
+  `Dispersio`, `ConsortiumName`, `Voces`). The namespace holds fourteen
+  verbs again. `docs/voice.md` emigrates with the code.
+- Voice test suites (`voice-vox`, `voice-chorus`, `voice-features`,
+  `voice-spectrum`) move with their subject.
+
+### Unchanged
+
+- `harmonia`'s planetary voices (`VoicedPitch`, `VoicedBody`, the Greek
+  vowel tables) are its own and remain.
+- Every other engine, table, and contract.
+
+## 0.2.0
+
+Rendering becomes a standalone engine (`inscriptio`), the per-chant analysis
+chart is completed, a voice engine is added, and the emitter surface settles on
+one format (SVG).
+
+### Added
+
+- **`inscriptio(score, opts?)`** — the standalone SVG renderer. Draws a `Score`
+  and returns `{ svg, geometry }`. Two notation species, each with its own
+  spacing pass: `"quadrata"` (square-note, SMuFL glyphs baked inline) and
+  `"moderna"` (modern round-note transcription — treble-8 clef, engraved slurs,
+  after the Lomer practice). A multi-system **layout engine** (`width`,
+  `systemGap`, `custos`) and **front matter** (`title`, `rubric` /
+  `annotation: "auto"`, `dropcap`, `rubricaColor`).
+- **The geometry contract** — `geometry: NoteGeometry[]`, one entry per note in
+  tabula order (system, x, y, systemY). A public API: downstream analysis tracks
+  build on it instead of scraping the SVG.
+- **The intonation channel** — `accidentals: "standard" | "heji" | "cents"`, a
+  **moderna** overlay. HEJI comma accidentals are baselined on the Pythagorean
+  chain (tonus's default tuning), so a Pythagorean chant renders clean and
+  syntonic-comma arrows bloom only under a just preset; meantone is not just, so
+  `heji` throws under it. `cents` labels signed deviations against `"pythagorean"`
+  (default) or `"et"`. Square notation (`quadrata`) carries only the accidentals
+  GABC itself expresses (flat / natural / sharp) — the HEJI and cents overlays are
+  modern analytical marks, so quadrata throws when asked for them.
+- **Score metrics** on `prosody`: interval statistics (histogram, maxLeap,
+  leapRate, step/skip/leap `motus`), `tessitura`, the melodic `arcus`
+  (initial/peak/final + arch index), `melismaCadential`; and conveniences
+  `phrase.noteCount` / `phrase.syllableCount` / `syllable.melisma`.
+- **`vox` and `chorus`** — a singing voice modelled as formant and spectrum
+  data, and seeded ensembles. `formantes` tunes to a temperament directly:
+  `tenor.formantes("a", temper, vis?)` — `vis` weights the pull, 0 (phonetic
+  truth) to 1 (fully tuned, the default).
+- **The appendix.** The export law is settled: verbs live on the namespace,
+  return values are plain data, and the named exports are canonical constant
+  tables only — `SEASON_LABEL`, `TEMPUS_NAME`, `GRADE_ORDER`, `GRADE_NAME`,
+  `MODES`, `TONES` (types `PsalmTone`/`Differentia` ride with their table).
+  The grade helper functions (`gradeOrder`, `compareGrade`, `ritusToGrade`)
+  retire from the surface: `gradeOrder(g)` is `GRADE_ORDER.indexOf(g)`.
+- **`docs/` ships in the package** — the documentation renders from the
+  installed tarball, pinned to the version it describes.
+- **GABC lyric markup decodes.** The angle-bracket text tags — excluded since
+  the MIDI-only days — are parsed at intake: `<sp>` specials become real
+  characters (℣ ℟ † ǽ œ, the raised *), style tags (`<i>`, `<b>`, `<sc>`,
+  `<c>` rubric color, `<e>` elision) survive as styled `runs` on `Syllable`
+  and the tabula row, and both notation species draw them as `<tspan>`s.
+  Layout hints (`<clear>`, `<nlba>`, centering braces) and `\pageref`
+  cross-references vanish. Before this, quadrata printed `<sp>V/</sp>` as
+  literal lyric text and syllable widths were measured tags-and-all.
+- **Per-role text faces** — `inscriptio`'s `fonts` option assigns a face to
+  the `dropcap`, `title`, `annotation`, and `lyric` roles (family, optional
+  weight, optional size scale). By default the SVG carries references and
+  the host page supplies the face; a slot may instead `embed` the caller's
+  own font bytes (base64) into the SVG's `<style>` for a self-contained
+  file. tonus bundles no font files either way.
+
+### Changed
+
+- **License: PolyForm Noncommercial 1.0.0** (was MIT through 0.1.x, which
+  remain MIT). Free for any noncommercial purpose; commercial use by
+  arrangement with the author.
+- **Corpus double-escape fixed.** Every `gabc` field stored its non-ASCII as a
+  literal `\uXXXX` escape, which had silently disabled accent detection across
+  the whole corpus — so note weights, prosody, rhythm, and imprint were computed
+  accent-blind. The extractor now decodes correctly; **computed accent weights
+  shift corpus-wide** as a result. A guard test asserts no gabc carries an escape.
+- **`cantus({})` throws.** An empty or unknown-key chant query is a caller bug,
+  not an empty result; it throws with guidance (matching the `festum` contract).
+- NABC pipes stripped from the corpus (`(notes|nabc)` → `(notes)`), with a
+  corpus-wide guard test.
+
+### Fixed (pre-release review)
+
+- **The Pythagorean chain is the medieval dodecachord (E♭–G♯).** The chain had
+  stacked twelve fifths ascending from C, spelling F as E♯ — a 521.5¢ wolf
+  ut–fa — and b molle as A♯, so `temperamentum` disagreed with `harmonia`'s own
+  pure F. The naturals now sit F–B with b molle and E♭ on the flat side, the
+  ficta sharps on the sharp side; the heji/cents baseline derives from the same
+  chain constant, so a flatted chant under the default tuning renders clean
+  (it had thrown).
+- **Quadrata figure grouping is phrase-aware.** A single-syllable phrase
+  followed by another phrase merged figures across the divisio, silently
+  dropping the second lyric and the divisio itself. An accidental inflecting
+  any note of a ligature now prints before the whole figure (Solesmes
+  practice) instead of vanishing on non-initial notes.
+- **Canticles by name resolve their own texts.** `magnificat` and
+  `nunc dimittis` pointed at the wrong psalter rows (one returned the Symbolum
+  Athanasium); `te deum` is withdrawn — the Te Deum is not psalmody.
+- **Formula step-skeletons count Apel's degrees.** Octave-aware relative to the
+  final's register: the mode-5 tenor reads +4, where the old ±half-octave fold
+  read −3. (Cadence figures keep their deliberate local-contour fold.)
+- **The register doctrine at the seams.** `Matins.nomen` / `Matins.ritus`
+  (were `name` / `rank`), `PsalmTone.nomen`, and `rubricaColor` (was `rubrica`,
+  a Latin key that took a CSS color string).
+- **The error contract enforced across the surface.** `notatio`, `festum`,
+  `caelum`, `matutinum`, `harmonia`, `proprium`, `temperamentum`, and `vox`
+  throw with guidance on junk input instead of raw TypeErrors or silent
+  fallbacks; a bare `chorus()` sings (the schola is the default); `psalmus`
+  no longer stamps the wall-clock year into `source.year` (determinism);
+  `iter` honours the `latinitas` shift; a user-GABC `office-part` header
+  normalizes to its OfficeCode.
+
+### Removed
+
+- **The MusicXML and MIDI emitters** (`score.musicxml()`, `score.midi()`).
+  tonus emits one format now: SVG. Microtuning still lives on the tabula
+  (`bend`/`hz`/`offset`) for a Web-Audio player to read directly — microtonally
+  exact, which MIDI never was.
+
+## 0.1.8
+
+- **Roman Matins.** `matutinum({ feast })` assembles the structured Roman night
+  office — the nocturns with their great responsories — from the community
+  *Nocturnale Romanum* (new `nr` chant source, 1,564 chants). Sanctorale feasts
+  draw Matins from their commune; coverage is the sanctorale and Advent today
+  (see COVERAGE.md). A separate accessor; the flat `officium` path is unchanged.
+
 ## 0.1.7
 
 - **`corpus(code)`** — metadata and analytics for a corpus book. Returns its
@@ -44,7 +707,7 @@ All notable changes to tonus. Newest first.
 
 - **Modulation detection** (`score.modulations`) — where the tonal centre leans
   away from the home mode, calibrated against Suñol.
-- **Richer modal affinity** — degree-, ictus-, and cadence-weighted, with ranked
+- **Modal affinity** — degree-, ictus-, and cadence-weighted, with ranked
   initials after Rockstro.
 - **`modus()`** tunes its finalis, tenor, and ambitus through the temperamentum.