colitur.git/tools, branch v1.0.0

feat(ordo): wire the extraordinaryform.org preface column in as a third witness

2026-08-24T14:16:59Z

Extends the preface comparison (RG 482-499) with a third, independent publisher, extraordinaryform.org's three annual Ordo editions (2024-2025, 2025-2026, 2026-2027), previously declined for two reasons that are now resolved: - "Advent" is not a preface: RG 494(b) grants the Trinity preface on every Advent SUNDAY, RG 498 the residual Common on every Advent FERIA; this source's own "Advent" label covers both indiscriminately (confirmed live on all 11 non-impeded Advent Sundays in the corpus). Classified by the day's own civil weekday, not a static table. - The known Vigil-of-St-Lawrence gap is reconfirmed, with a nuance not present in the earlier audits: colitur's own RG33-third-trigger now omits the vigil in some years too, so the corpus's gap and colitur's answer coincide in 2025 (a full match, not a coincidence) while 2027 still shows the classic identity-wrong-but-preface-invisible shape. Re-derives the known menu-trap office filter (Requiem-skip, "and "-chain) against the live JSON and finds a second, general pattern: "Saturday of Our Lady" (RG78's BVM Saturday Office), whenever listed, is unconditionally the office -- checked exhaustively across all 38 instances in the corpus, zero counter-examples. A third, generic "absorption" rule was deliberately not built: it would resolve 2 of the 11 residual divergences but wrongly misfires on a structurally identical but semantically different shape (an optional Rogation-day Mass option). The JSON's own extraction carries a genuine classifier gap: parse_ordo.py's hardcoded PREFACES set omits "Christ the King", misfiling it into comms; recovered in the generator, not left blank. 11 of 1088 comparable rows diverge, resolving to 7 adjudicated root causes (E1-E7): two Commemoration_only Marian titles reduced to commemorations (Mt Carmel, Ransom), two genuine sanctoral data gaps cross-confirmed absent by the FIUV/LMS witnesses (Miraculous Medal, St John before the Latin Gate), two single-witnessed efdotorg outliers against FIUV/LMS agreement (St Michael's Dedication, All Saints on a Sunday), and one isolated source labelling slip. Mutation-tested: corrupting the RG495 BVM-Saturday clause reddens 9 tests across this new layer and both existing Ordo witnesses. Per-edition agreement: 358/363 (2024-2025), 358/363 (2025-2026), 361/362 (2026-2027).

feat(ordo): wire the LMS preface column into the witness suite

2026-08-23T21:01:42Z

Extends tools/extract_lms_ordo.ml's row type with a praef field (the first "Pr of ..." / "Common Pr" match in each day's universal block, same first-in-reading-order discipline already used for Gl/Cr) and regenerates all three LMS fixtures. Also brings lms-ordo-2024-2025.sexp under the generalised, probe-verified provenance template the Witnesses task previously added only for the two newer editions -- it had kept Task 6's original hand-written header. Characterisation finding: this Ordo prints OPTION LISTS ("Pr of X or Pr of Y or Common Pr"), not a single value per day, and several named options (Martyrs, All Saints and Patron Saints, the Dedication of a Church, the Most Holy Sacrament, St John the Baptist, the Angels) are genuinely absent from RG 484-497's own fourteen named prefaces. Every instance found in the universal block pairs such an extra with a genuine RG-482 answer, so test/test_lms_ordo.ml compares by membership (classify_praef_options, a 15-entry prefix table split on the standalone token "or"), not string equality. Opposite-prediction probes (Nativity/Apostles/Requiem) are hard- verified inside the extractor itself before any fixture is trusted, the same discipline already used for the Creed column. Result: 396/396, 399/399, 398/398 comparable rows agree across the three editions (Good Friday excluded per edition, no Mass that day), zero allow-list entries needed. extraordinaryform.org's own three editions were characterised but deliberately not wired in this pass -- see .superpowers/sdd/2026-08-21-colitur-celebrant-rubrics-phase1/preface-witnesses-report.md for the full account (the field is genuinely per-day, contrary to a naive header-count read, but carries its own classification anomaly during Advent and reconfirms the corpus's known St Lawrence vigil defect, and building a full extractor for it is a separate, comparably large task not needed to decide the RG 497 question this same report settles). Claude-Session: https://claude.ai/code/session_017ZBxCCRM2ojnBupp3SBxV9

fix(ef): ferias after the Ascension repeat its own Mass, not a Sunday's

2026-08-22T18:55:18Z

Ascension has no octave under the 1960 rubrics -- confirmed structurally: the Proprium de Tempore's own table of contents runs straight from "In Ascensione Domini" to "Dominica post Ascensionem" with no per-feria headings in between. Step 3 of the reading chain therefore fell through to Preceding_sunday on every unoccupied weekday in that stretch, walking each one back to the 5th or 6th Sunday after Easter's own citation. The Missal's own Proprium de Tempore carries an explicit marginal rubric immediately after the Ascension's Mass propers, before "Dominica post Ascensionem" begins: "Diebus a feria VI post Ascensionem usque ad feriam VI ante vigiliam Pentecostes inclusive, dicitur Missa de festo Ascensionis, cum Gloria et praefatione de Ascensione, sine Credo et Communicantes proprio" -- every feria without its own proper, from the Friday after the Ascension through the Friday before the Vigil of Pentecost, repeats the Ascension's own Mass. RG 490(b) corroborates independently: it assigns the Ascension preface "tamquam de Tempore" across that identical window, which only makes sense if the Mass actually repeated there is the Ascension's, not a numbered Sunday's (whose own Mass would carry no need for a preface override at all). Witnessed directly, not only inferred from the rubric: the LMS Ordo prints "Mass of the Ascension" literally on every affected day across all three of its fixture windows, in BOTH the tail of Ascension's own week and the following week -- ruling out the alternative reading (that the second week instead resumes "Dominica post Ascensionem"'s own Mass, which a first pass over the FIUV Ordo's own week-labelled titles could suggest at a glance; its own [praef] field for the same two dates independently reads "Ascensionis", resolving the apparent conflict). Implemented as data, not logic, matching this codebase's own established precedent for the identical RG 299 shape (the week after Trinity Sunday, data/ef/lectionary.sexp's [trinity_week_entries]): six new hand-authored entries, [ascension_week_entries] in tools/bootstrap_lectionary.ml, regenerated against a git-pinned copy of lectio's own source ini (SHA-256 verified unchanged) so the regeneration carries no unrelated drift. The window's two Saturdays are deliberately excluded -- RG 78's votive Office of Our Lady on Saturday always wins an otherwise-unoccupied Class4 Saturday first, before step 2 (this fix's own entry point) is ever reached, so an entry for either slug would be dead data. Blast radius measured in-process across the full 1583-9999 domain (throwaway probe, not committed): 16 350 days total across the six slugs, every one resolved via step 2 (Own_slug), zero anomalies, and the window's two Saturdays confirmed to contribute zero hits via this cause across the whole domain, corroborating the RG78-interception reasoning by direct sweep. Closed L2 in data/ef/expected-divergences-lms.sexp. New cited entries: C41 in data/ef/expected-divergences.sexp (95 of 16801 rows, 2005-2050 -- lectio has no equivalent rubric and still walks these ferias back); M31 in data/ef/expected-divergences-missalemeum.sexp (3 rows in the 2026-2027 window, plus the same shape independently reconfirmed in the 2035 and 2038 live-capture windows) -- missalemeum, Divinum Officium's lineage, still gives the same wrong answer colitur itself used to give, which is why these rows used to agree with M18's shape and now escape it (M18 373 -> 370, the same mechanism M23-M27 established when the citation axis was first built). Golden pin: 2031-05-23, the exact date the bug report was filed against. Mutation-tested: reverting data/ef/lectionary.sexp alone (a git-diff capture/restore, not a source-code revert, since the fix is data) reddens 19 tests across the differential, both oracle fixtures, all three LMS Ordo windows, the golden pin, the lectionary unit test and every Render golden template; restored and reconfirmed green (681 tests; 683 under COLITUR_EXHAUSTIVE_SWEEP=1, ~175s).

fix(ordo): correct L5/F3 misadjudication -- Cum Sanctissima, not a gap

2026-08-22T16:46:27Z

data/ef/expected-divergences-lms.sexp's L5 (16 dates, LMS Ordo Gloria=true vs colitur's false) and data/ef/expected-divergences-fiuv.sexp's F3 (6 dates, FIUV Te Deum=true vs colitur's false) were both misadjudicated as colitur's own gap under an RG 302(b)/431(b) hypothesis. Both are wrong. These are occurrence differences from the 2020 decree Cum Sanctissima, which colitur excludes from its 1962 core by binding design (CLAUDE.md decision 2). RG 303(b) ("dici potest tantum si occurrit dies liturgicus IV classis") already prohibits 302(b) on the Class3 privileged ferias both entries cover -- the original text quoted 303(b) but never followed it to its conclusion. Decisively, the FIUV Ordo prints BOTH readings on the six shared dates, separated by a literal "-VEL-" token: the primary, strict- 1962 reading (feria wins, sine Gloria, no Te Deum) matches colitur exactly; the second, Cum-Sanctissima reading (saint kept, Gloria, Te Deum) matches the LMS Ordo exactly. F3 itself was never a real Te Deum finding: tools/extract_fiuv_ordo.ml's extract_te_deum scanned each block unbounded for "Ad Mat.", so on these six days -- whose primary office states no "Ad Mat." of its own -- it silently read the second office's own Te Deum instead. Fixed with a new split_at_alt_office, cutting every block at "-VEL-" before any field extraction runs; the alternative is captured, not discarded, in a new alt_te_deum field. A hard characterisation probe asserts the split fires on exactly the six known dates. Auditing for the same trap elsewhere (this is the third such manufactured divergence in this project) found a related but narrower marker, "Vel (sec. decretum <>) Missa ...", offering an alternate Mass formulary within a single office. It never corrupted Gloria/Credo (always stated before the marker on all 20 real instances, confirmed by diff), only the uncompared praef field; hardened defensively anyway, before it could ever fire live. The LMS Ordo (all three editions) carries no equivalent dual-print mechanism at all. expected-divergences-missalemeum.sexp's M27 is the same genus but was already correctly adjudicated, not touched. Both allow-list entries are rewritten, not deleted: the original SHAPE/ CITATION/WHY-NOT-FIXED-HERE reasoning is preserved verbatim as a recorded, superseded hypothesis, matching this project's own correction convention. test_lms_ordo.ml's matching predicate is unchanged (the 16-date population still fires, and always will, because it reflects a standing exclusion, not a bug); only its citation comment is corrected. test_fiuv_ordo.ml's F3-specific predicate/count assertion is removed, matching this file's own prior F2/F4 closure discipline. lib/ is untouched. dune test: 676/676. make check (COLITUR_EXHAUSTIVE_SWEEP=1, full 1583-9999 domain): 678/678.

feat(ef): the Gloria in excelsis, RG 431-432, deferring to Breviary 237-238

2026-08-22T15:56:42Z

Phase 2 of celebrant-rubrics-phase1: colitur rubrics gains a fifth column, whether the Gloria is said. Follows the Creed's own seam exactly -- Rite.t.gloria, Liturgical_day.t.gloria, wired through calendar.ml the same way. RG 431(a)/432(a) defer the Gloria to the Breviary's own Te Deum rule (nn. 237-238), so te_deum is implemented as its own named predicate, cited clause by clause, not collapsed into a colour heuristic. 431(c) (Holy Thursday, the Easter Vigil Mass) and 432(b)/(d) (violet; a Requiem) are independent overrides checked ahead of the Te Deum-derived answer. Every clause this engine has no dimension to model (votive Mass classes, the wider n.302 "Missa festiva" categories) is stated as N/A with its own reasoning, not silently dropped. Validated against the FIUV universal Ordo (Gloria and Te Deum) and all three LMS editions (Gloria). A first pass over-trusted a clean-looking 15-for-15 FIUV contradiction of 237(b)'s own Septuagesima exception and replaced it with a blanket "every Sunday" rule; the evidence was itself corrupted -- the FIUV extractor recognised only one of the source's two Te Deum negations ("non dicitur", not "sine"), so every "sine Te Deum" Sunday read wrongly true. Fixed in tools/extract_fiuv_ordo.ml, fixture re-extracted, and the literal 237(b) reading restored once the corrected data confirmed it. A second bug surfaced alongside it (Palm/ Passion Sunday wrongly reading true via Temporal_ef.named's own table membership, then Christ the King wrongly reading false from an over-broad fix) is closed with an explicit two-slug exclusion. Domain-wide 1583-9999: every violet or Rose day is gloria=false except the Easter Vigil (RG 431(c) lex specialis), every Requiem is gloria=false, both measured exhaustively, zero exceptions. Mutation- proved: disabling 431(c) reddens 8 tests including all four oracle comparisons; disabling 238(c)'s feria-I-classis exclusion reddens exactly the dedicated Ash Wednesday unit test, a genuine blind spot in both oracle layers, reported rather than hidden. Two open, cited findings, neither fixed here (out of this task's "follow creed's exact seam" scope): a privileged Lenten/Passiontide feria carrying one commemoration reads Gloria=true in the LMS Ordo but Te-Deum=true/Gloria=false in FIUV -- the two oracles disagree with each other, not merely with colitur (data/ef/expected-divergences-lms.sexp L5, expected-divergences-fiuv.sexp F3); and a pre-existing, uncited Colour.Violet bug on Rogation Monday/Tuesday in Temporal_ef.temporal, surfaced by this comparison but root-caused as a separate defect (L6). day/readings verified byte-identical against a build from the branch tip before this task (v0.10.1's own tag predates an already-landed bissextile fix that legitimately changed both, so it is not the right baseline). 671 tests green (dune test); 678 with the exhaustive sweep (COLITUR_EXHAUSTIVE_SWEEP=1 dune test --force, ~104s).

test(fiuv-ordo): add the FIUV universal Ordo as a second, independent witness

2026-08-22T13:11:43Z

Wires in the FIUV (Foederatio Internationalis Una Voce) Ordo for 2025-2026 as validation layer seven -- the first UNIVERSAL (not diocesan) source in this project, and the first compiled by someone other than the LMS editions' own Peter Day-Milne (Joseph Shaw, FIUV President). tools/extract_fiuv_ordo.ml hand-parses pdftotext's -layout dump of the Latin-language ordo (no Str/regex, frozen deps). Three real parsing traps were found and fixed by tracing actual false results against the raw source, not assumed in advance: - "Gloria" is also the first word of "Gloria Patri" (the psalm doxology), which can appear INSIDE the very Mass clause being scanned (Good Friday: "...omittuntur ps. Iudica me et Gloria Patri), Gloria, sine Credo..."). Guarded by rejecting any "Gloria" hit whose next token is "Patri". - Christmas Day's own rubric uses the PLURAL "Missae"/"Missae" ("tres Missae celebrantur..."), never the singular "Missa" every other real day anchors on -- silently produced gloria=None/credo=None on the one most doctrinally unambiguous day in the calendar until traced and fixed, guarded in turn against "Missae defunctorum" (an unrelated Requiem-prohibition notice appearing at the end of many day-blocks). - The Vespers boundary (bounding the primary Mass option before a second "Vel Missa" alternative or before Vespers information) missed the ligature spelling "VESPERÆ", the dominant form in this source, letting a whole day's trailing prose leak into the captured [praef] field on the task brief's own worked example day (19 September). A distinct second stop-marker bug: the malformed trailing entry this source shares structurally with the LMS PDFs (a stray, mislabelled duplicate of 1 January) was initially over-matched by a generic "first token is 4 digits" heuristic, which false-positived on an indulgence-decree date citation wrapped mid-sentence in October ("...20 augusti / 1885, 26 augusti 1886..."), truncating the extraction to 308 of 400 real days. Fixed by anchoring on the tail entry's own distinctive "1st Jan" opening instead. Five opposite-prediction Creed probes are re-verified by the tool itself against its own extracted rows (a failed probe is a hard exit, not a printed claim) before the fixture is written. test/test_fiuv_ordo.ml compares the Creed against colitur across all 399 comparable days (Holy Saturday excluded, no Mass that day) -- CLEAN, zero unexplained divergences, a genuine second confirmation of {!Rite_ef.Rubrics_ef.creed} independent of the LMS lineage. [praef], [te_deum] and [class_] (including the FIUV ordo's own "III cl." vs "III cl. (Priv.)" Cum Sanctissima convention, stated on its title page) are captured into the fixture but deliberately not compared -- colitur has no Gloria/preface predicate yet, and its default configuration models no Cum Sanctissima overlay to compare the class convention against. All 634 tests green (dune test), 635 with the exhaustive sweep and make check. No lib/ file touched. Claude-Session: https://claude.ai/code/session_017ZBxCCRM2ojnBupp3SBxV9

test(lms-ordo): extend the LMS witness layer from one year to three

2026-08-22T12:51:07Z

Generalises tools/extract_lms_ordo.ml's start-of-body detection: the hardcoded "End of November 2024" sentinel assumed every edition opens with an "End of November" stub section. The 2023-2024 edition has none (Advent Sunday 2023 fell on 3 December, so the compiler folded the two tail days into the "December 2023" header directly) -- the generalised detector anchors on the body's own "ORDO" title line instead, verified unique in all three known editions and distinct from the Table of Contents' lowercase "Ordo" entries. Extracts test/fixtures/lms-ordo-2023-2024.sexp (397 rows) and lms-ordo-2025-2026.sexp (399 rows), each with a provenance header carrying its own SHA-256, source citation, and a per-edition characterisation record: five opposite-prediction Creed probes, independently re-verified against this run's own extracted rows (a failed probe is now a hard die, not a printed claim). The 2025-2026 edition needed a different sharp II-class pairing than 2024-2025's Joachim/Stephen one, because Joachim (16 August) is impeded by a Sunday that year -- recorded in the fixture header, not silently swapped. test/test_lms_ordo.ml is refactored into a parameterised make_suite, instantiated once per edition; every count Task 6 hardcoded against the single 2024-2025 window is now a per-edition parameter, several of which turned out to differ once measured (BVM-Saturday population 12/14/13, Ascension-week exclusion 1/3/2). Two mapping gaps surfaced only by the new windows and fixed here (test-side, not colitur): Low Sunday reached via Preceding_sunday for the first time (absent from 2024-2025's own window), and the Ascension-week exclusion was wrongly keyed off week number alone -- Rogation Monday also resumes week 6's Sunday but falls BEFORE Ascension, not after, and the Ordo prints the ordinary Sunday text there. Re-bounded by date, against Ascension and Pentecost. Two genuine divergences surfaced in the 2023-2024 window and are allow-listed (L3, L4) rather than fixed, per this task's no-lib-changes scope: - L3 (verdict ordo): colitur has no bissextile calendar rule. The 1962 Missal's own calendarium moves St Matthias from 24 to 25 February in leap years (docs/research/LT.txt:5011); colitur keeps him on the 24th every year. Domain-wide upper bound measured with a throwaway sweep: 1,650 of 2,041 leap years in 1583-9999 currently observe him there. - L4 (verdict open): Christmas Eve falling on the 4th Sunday of Advent (2023-12-24, roughly 1 year in 7). colitur says the Creed is said, reading RG 475(a)'s Sunday clause; the Ordo says it is not. RG 30(a)'s own "locum tenet ... nulla fit commemoratio" for this exact vigil reads like RG 16(a)'s Feast-of-the-Lord-on-Sunday shape, which colitur already treats as Creed-preserving -- but RG 475(a) names "festo" specifically, and a vigil is its own liturgical-day category (RG 35), so the counter-reading is real too. Left open rather than adjudicated. All 630 tests green (dune test), 631 with the exhaustive sweep. Claude-Session: https://claude.ai/code/session_017ZBxCCRM2ojnBupp3SBxV9

test(ef): the LMS Ordo as a sixth validation layer

2026-08-22T10:58:15Z

An independent lineage -- neither Divinum Officium, missalemeum nor lectio -- covering the Mass formulary and the Creed. Characterised before use, per the discipline the extraordinaryform.org Ordo's blanket St Lawrence vigil gap taught this session: five opposite- prediction Creed probes (Sunday/feria, octave-override vs plain II-class, I-class) all confirmed against the raw extracted text before a single divergence was adjudicated, recorded in the fixture's own provenance header. tools/extract_lms_ordo.ml hand-parses pdftotext -layout output (no Str/regex, frozen deps) into test/fixtures/lms-ordo-2024-2025.sexp, 400 day-rows (2024-11-27..2025-12-31). Gl/Cr print per Mass-option, not per day; the extractor takes the first pair in reading order, which the source's own layout guarantees is the universal entry's, since a diocesan variant always follows it, never precedes it. The Ordo's I-V roman numeral ("Missae de sancta Maria in sabbato", RG 309(a)) is matched whole-line only, never by substring -- "V Mass of BVM" is a substring of "IV Mass of BVM". test/test_lms_ordo.ml compares two axes 1:1 by date: - the Creed (RG 475-476): 399 of 400 days comparable (Good Friday has no Mass at all, asserted as the only such day); one genuine divergence, adjudicated to the ORDO, not colitur -- All Souls' Day 2025-11-03 is a Requiem Mass, and colitur's own creed predicate has no notion of one, a pre-existing scope gap rubrics_ef.ml already documented in its own header before this task existed. - the BVM-Saturday seasonal Mass selection, the addendum's own "sharper check": all 14 Votive days in the window match their season's own numeral exactly, both directions, independently re-derived from RG 309(a) rather than calling colitur's private function. - the formulary override, for the three sources it can honestly discriminate (Proper 179, Preceding_sunday 66, Common 2, all clean); Own_slug (139) is excluded with evidence, not silently -- several ferial slugs carry citations byte-identical to their Sunday's own, so the Ordo's override line cannot distinguish the two representations. data/ef/expected-divergences-lms.sexp carries the one adjudicated entry (L1), cited to RG 476(f). Mutation-tested: inverting RG 475(a)'s Sunday clause reddens 10 existing Rubrics_ef unit tests AND this new layer's own comparator, independently, on a data source none of the other five layers touch. Reverted after confirming. No lib/ changes. day/readings/rubrics output re-confirmed byte- identical (1583, 2026, 2038, 9999) after this task; all three gates green (dune test, exhaustive sweep, make check).

fix(tools): close five more ways to defeat check_citations.py

2026-08-19T13:47:37Z

Round 1 hardened check_citations.py against its own self-poisoning bug; a review defeated it again. Five fixes, in the order they were found: 1. PATTERN silenced a whole comment block, not just the entry it was attached to -- a wrong citation on a DIFFERENT, unmarked entry in the same block (e.g. [season]'s own back-to-back trailing-comment style) was never checked at all. Fixed by scoping PATTERN with the identical leading/trailing pooling rule citations already use: an entry is excluded only by its own marker, never a neighbour's. 2. Explicit per-citation ranges (introduced in round 1 to replace a blanket +-2-line tolerance) had no upper bound, reintroducing the same defect at a much larger radius (LT.txt:8600-8650 passed if the text appeared anywhere in fifty lines). Capped at MAX_RANGE_WIDTH (3 lines); anything wider is reported MALFORMED, naming the entry and the width, instead of silently accepted. 3. The "pool too thin to verify" gate counted words, not rarity -- it flagged 11 genuinely correct citations (short Latin hagionyms with only one non-stopword) CANNOT VERIFY, while a match on nothing but "classis" (507 occurrences) passed freely alongside three siblings. Replaced with a frequency table over the whole LT.txt corpus: a token's evidence is 1/(times seen), an item's evidence is its single rarest matched token (not a sum -- summing would let several merely-common words add up to "enough" between them, the same shape as the self-poisoning bug). 4. "LT.txt:12,459" (a comma typo for one number) parsed as two unrelated bare citations, 12 and 459, either of which could coincidentally match while the intended line was never checked. Detected as a thousands-separator-typo shape (a 1-2 digit token immediately followed by an exactly-3-digit one -- the only way a real LT.txt line number, which never exceeds 5 digits, splits under one comma) and rejected as malformed. 5. The self-test suite overstated its own coverage: of round 1's seven fixture cases, only two actually failed against the pre-round-1 script. Every test is now labelled REGRESSION or CHARACTERISATION, each verified by direct replay against the named prior version rather than asserted -- 14 of 33 are genuine regression tests. Both of the review's own defeats (block-wide PATTERN silencing, the 50-line range) are reproduced as dedicated fixtures and confirmed caught; both are also confirmed to slip through the pre-round-2 tool unchanged. Claude-Session: https://claude.ai/code/session_017ZBxCCRM2ojnBupp3SBxV9

fix(tools): harden check_citations.py against its own self-poisoning bug

2026-08-19T13:01:18Z

Reproduced the defect: reintroducing the exact historical citation bug (pointing class-1's citation back at LT.txt:12459, the value a prior fix round corrected away from) made the tool report "147 citations checked, 0 look wrong". The mechanism was that the corrective comment documenting the old bug quotes the wrong historical value, and the checker pooled every quoted phrase from the whole surrounding comment block, so citing the wrong line matched the comment explaining why it was wrong. Four changes: 1. The word pool for a citation is now scoped to the entry(ies) it is attached to only -- never to quoted text elsewhere in the comment. This is the direct fix for the self-poisoning bug. 2. A citation whose pool has fewer than two distinctive words (Latin liturgical headings are short and stopword-heavy) cannot discriminate the right line from a wrong nearby one. Such a citation is now reported CANNOT VERIFY and fails the target, instead of silently passing. 3. The blanket +-2-line tolerance is gone. A bare "LT.txt:N" is checked at line N only; a heading that genuinely wraps must say so explicitly as "LT.txt:N-M". The allowance moves into the data, where it is visible. 4. The tool gets its own test suite, tools/test_check_citations.py, with a synthetic fixture covering: a correct citation, off-by-one and off-by-three mismatches, an explicit wrap range, a degenerate pool, a PATTERN-marked entry with no citation, and a dedicated regression test for the self-poisoning case itself. Wired into `dune test` via a new (rule (alias runtest) ...) in tools/dune (a plain (test ...) stanza cannot run a Python script), so it runs with the rest of the suite, not only as a `make` target. Added a --file/--lt-file override to check_citations.py so the tool (and its own tests) can point at a fixture without touching the real lang/la.ini or docs/research/LT.txt. Confirmed the "SKIPPED, exit 0" behaviour for a missing docs/research/LT.txt is unchanged. tools/__pycache__/ (a stray artefact of this script, previously untracked and ungitignored) is now in .gitignore. Measured against the current lang/la.ini (another task is still landing its sanctoral entries on this branch): 15 of 275 citations now look wrong and 42 more cannot be verified, both far above the 0 the unhardened tool reported. Not fixed here -- the data pass is separate, once the sanctoral entries land.