colitur.git, branch main

Merge branch 'of-litcal-readings'

2026-09-03T19:13:38Z

The OF's readings had never been validated against anything but the lineage they arrived on. Pointing a second lineage at them found ten corrupt references on the first run.

chore: 1.2.1

2026-09-03T19:13:33Z

Ten wrong scripture references reached users; nothing else about the engine changed and no output format moved. A patch bump. README's status line and test count follow the release, as they have to: the count is 917 under the exhaustive sweep, not 916.

fix(of): correct ten reading references, regenerated from lectio

2026-09-02T10:54:07Z

Nine ended at verse 22 where the OLM ends them at 20; the tenth read "Matthew 17:9a,10-135", a verse Matthew 17 does not have. Every one was in a Year II entry, or in of-advent-2-saturday, which has no year cycle. Not one Year I entry was affected -- that distribution is what made this a transcription fault rather than ten independent slips. Corrected upstream in lectio's own internal/caldata/of-lectionary.ini (its suite green) and this file regenerated from it, per the header's own instruction not to hand-edit a generated artifact. The diff is exactly the source SHA-256 pin plus the ten references; both pins in test_lectionary_of.ml move with it. The Ordinary Time invariant added in the previous commit now asserts outright rather than pinning six exceptions: in Ordinary Time the OF weekday Gospel does not depend on the year of the cycle, so a -i/-ii pair whose Gospel differs is a defect, and the expected violation set is empty. Mutation-tested against the Year II entry alone -- an earlier attempt changed both halves of the pair and passed for the wrong reason, which is worth knowing about this shape of test. Measured: the 20-vs-22 divergence class against litcal goes 26 -> 0 over the twelve witnessed years. 917 tests, every year 1583-9999. Also, L6/L7 (the Sacred Heart / John the Baptist tie at an equal Tabula entry) gains the citation its note said did not exist. Tabula I.3 enumerates "Sollemnitates Domini, beatae Mariae Virginis, et Sanctorum" -- an order, not a bare category, and read that way it settles the tie against colitur's current alphabetical answer. Recorded with its counter-argument (entry 4 marks its own internal order with explicit lettering; entry 2's list cannot be an ordering at all) and with what would settle it. Verdict stays open, behaviour unchanged.

feat(cli): --help is one screen, not the whole manual

2026-09-02T10:30:56Z

It printed 374 lines: every output format, the RFC 4180 quoting rule, the template scope hazard, the publish manifest semantics. All of that is real documentation and all of it is already in colitur(1), installed beside the binary. A help text nobody reads at the terminal is not more helpful than a short one, it is less -- the reader stops looking. Now 42 lines: the commands, the flags most runs need, the year domain, the exit status, and where the rest lives. Everything removed is in the man pages, which is checked rather than assumed -- each dropped section has a counterpart section there. One thing moved rather than went. cli.t pinned `rubrics`' documented field list against what `rubrics` actually prints, because that field count drifted once already and stayed wrong for a whole review round. Deleting the prose would have quietly retired the check, so it now reads the worked example out of man/colitur.1 instead, and test/dune's cram stanza gains the man page as a dep -- a sandbox holds only what that stanza names. Mutation-tested: changing the documented example reddens the suite. The tagline also stopped saying "Roman rite, 1962" -- the OF has shipped since v1.1.0.

test(of): pin the six corrupt Year II Gospels in Ordinary Time

2026-08-28T14:15:13Z

Built the litcal readings comparison the OF provenance notes had left as future work, and it found a defect class the existing layers structurally could not see: the litcal layer compares season, grade and identity, and stops there, so every reading colitur serves has been validated by nothing except the lineage it arrived on. Ten references are corrupt. Nine end at verse 22 where the OLM ends them at 20; the tenth reads "Matthew 17:9a,10-135", a verse Matthew 17 does not have. Every one lives in a Year II entry or in of-advent-2-saturday, which has no year cycle -- not one Year I entry is affected, which is what makes this a transcription fault rather than ten independent slips. This commit adds only the detection, and only for the six the invariant can reach. In Ordinary Time the OF weekday Gospel does not depend on the year of the cycle: OLM n. 69 point 4 alternates the FIRST reading and leaves the Gospel a single one-year series. So of-ordinary-time---i and its -ii twin must carry the same Gospel string, and six pairs do not. Scoped to Ordinary Time deliberately. The unscoped sweep reports 17 further disagreeing pairs in of-advent-3/4-* and of-christmas-*, all of them correct: those weekdays are keyed to the calendar DATE (OLM n. 69 point 3), so their -i/-ii entries describe two different dates rather than two years of one cycle. An unscoped test would report seventeen false positives and would, rightly, be switched off by whoever read it next. Pinned rather than green, because the fix is not ours to make here: data/of/lectionary.sexp's own header forbids hand-editing a generated artifact, and lectio's internal/caldata/of-lectionary.ini carries all ten verbatim. The exact six are listed, so the test fails if a seventh appears and equally if one is fixed without updating the list -- both are changes a reader must see. Mutation-tested in both directions: introducing a seventh violation reddens it, and correcting Mark 1:14-22 reddens it too. Evidence for the corrections, in docs/research/of/olm-1981-ocr.txt, the Latin OLM itself rather than a witness: "Mc 1, 14-20", "Mc 4, 1-20", "Mc 5, 1-20", "Mt 7, 15-20", "Mt 17, 14-20", "Mt 18, 15-20", "1 Sam 1, 9-20", "1 Sam 3, 1-10. 19-20", "Mt 17, 10-13". litcal independently confirms three of the six Gospels and all four references the invariant cannot reach. James 5:13-22 is unlocatable in the OCR but impossible on its face: James ends at 5:20.

Merge branch 'cli-pretty'

2026-08-28T08:03:37Z

--pretty, and the --month/--date/--today flags that make it usable: a box spans seven lines, so the flag that improved reading had removed grepping.

chore: 1.2.0

2026-08-28T08:03:32Z

Two presentation flags and no change to what is computed. Nothing was removed and no default output moved, so a minor bump. The CHANGELOG entry for 1.1.0 is backfilled here: that release was tagged without one, which would have left this file jumping 1.0.0 -> 1.2.0. Its content is taken from the v1.1.0 tag message rather than reconstructed. README's status section claimed the OF module lived on an unreleased branch and counted 887 tests, and its build block advertised the fast suite at ~8s and the exhaustive sweep at ~80s. Every one of those was true when written and none is now: 909 fast (~40s), 916 under the sweep (~8min). Stale suite timings are a defect this project has corrected once before, in 0.10.0.

feat(cli): --month, --date and --today narrow a report

2026-08-28T08:03:27Z

--pretty made a report readable and, in the same stroke, ungreppable: a box spans seven lines, so `grep 2026-03-` returns fragments of March rather than March. The nearest working equivalent is a paragraph-mode awk incantation that depends on the boxes happening to be blank-line separated -- knowledge no reader should need to select a month. The three flags are ALTERNATIVES, not a stack. Naming two exits 2 rather than letting one quietly win, the same discipline the positional year and --year already follow, and the same reason this program refuses a flag it cannot honour instead of accepting and ignoring it. --date and --today CARRY a year, so the year becomes optional on those two and required only after the window has had its say: `colitur day --today` is a complete command while `colitur day` still is not. A year named alongside them must agree. The disagreement message names the flag rather than printing two bare numbers, because with --today the second year appears nowhere on the command line for the reader to trace. temporal refuses --year yet accepts --date/--today including as its year source. That is not a back door to the refused spelling: --year is a second spelling of the positional year, which temporal deliberately does not offer, whereas --date selects a DAY and merely happens to determine which year contains it. The window is a predicate over an ISO date STRING rather than over a Date.t, because the two report shapes reach it differently -- resolved days carry a Date.t, temporal_report carries a flat Record whose date is already text -- and one shared test is one implementation rather than two that can drift. Accepted by day, readings, rubrics and temporal; refused by easter, emit, table, render, publish, check, convert, lang, config and new-overlay. --pretty had shipped undocumented in both --help and colitur(1). Both now carry it, alongside the new narrowing section; the man page's awk example is shown precisely so the flag that replaces it is justified. Default output is byte-identical to 1.1.0 across seven years, five commands and both rites, and across all five emit formats and the shipped templates.

feat(cli): --pretty draws each day as an ASCII box

2026-08-27T13:45:49Z

Aligned columns made the fields legible but the days ran together -- with 365 of them the eye had nothing to catch on. Each day now gets its own box: a heading carrying the date and the liturgical colour, then the celebration, its rank and season, and any commemorations, each on its own line. The box art is pure ASCII, only + - and |, never Unicode box-drawing. That is the point rather than a limitation: this format exists to be pasted or piped into a document, a mail or a plain-text ordo, and U+2500 and its relatives survive that only when every stage agrees about encoding and font. +---+ has never failed to render anywhere. readings gets a labelled block, so a citation says what it is instead of being the second of three bar-separated fields; the OF second reading simply omits its row on the days without one. rubrics becomes a label/value list. temporal is the day box minus the sanctoral it does not have. A blank line separates consecutive boxes -- without it the bottom rule of one day and the top rule of the next sit adjacent and read as a single doubled line, which is the same "not distinct enough" this change set out to fix. Alignment counts UTF-8 code points, not bytes, so "Pen~afort" and "Fremiot" still line the right edge up at 76 columns; a byte-counting pad shears the box by one per multi-byte character. Over-long values are truncated with a ~ rather than allowed to overflow, since a box whose right edge does not line up is worse than a clipped name that the default output still carries in full. Verified: no ANSI escape reaches a pipe on any of the four commands, and the default output is byte-identical to installed 1.1.0 across all five.

feat(cli): --pretty, for reading in a terminal

2026-08-27T12:12:43Z

The row commands print for awk: single-space fields, slugs, a variable tail. That is right for the default and wrong for a person, who mostly wants to know what today is. --pretty gives the same four commands -- day, readings, rubrics, temporal -- aligned columns, the day's liturgical colour as a swatch, and commemorations on their own indented line rather than lengthening the row. The colour was already computed and simply thrown away on a terminal. Colour is written only when stdout is a terminal, so piping or redirecting yields plain aligned text: the alignment survives, the escapes do not, and the swatch degrades to the colour's initial so the information is not lost with them. NO_COLOR is honoured on PRESENCE whatever its value, which is the convention's own rule -- treating it as a boolean is the usual way to get it wrong. Every other command refuses the flag rather than accepting it and doing nothing: emit, table, render and publish already choose their shape through --format and --template, and easter prints six key/value lines, not a grid. An intermediate version accepted it everywhere and silently ignored it on five commands, which is the failure mode this program refuses everywhere else. Two things the layout had to learn. Column widths are a minimum, not a maximum: the Latin season names run past them ("Tempus per annum ante Septuagesimam" is 35 against 34), so pad always leaves a separator or the next field fuses onto it -- that is how "Septuagesimam 1S. Hilarii" happened. And the commemoration indent is measured from the row actually printed rather than computed from the column constants, or it sits under the wrong column on exactly the rows that have something to indent. Presentation only: bin/pretty.ml decides nothing about what a day is, and nothing reads it. Default output is byte-identical to v1.1.0 -- verified across day, readings, rubrics, temporal and easter for 2026, 1583 and 9999.