Merge branch 'ef-rebootstrap': re-bootstrap the sanctoral, refresh the differential

Re-bootstraps data/ef/sanctoral.sexp from lectio's corrected generator
(322 -> 327 entries), propagating fixes colitur had inherited wholesale:
15 entries wrongly marked commemoration-only that are really III-class
feasts, four entries missing entirely, six commemorations an electronic
transcription had hidden, and two class corrections.

38639 days change across 1583-9999, in three shapes with no remainder:
2245 where a III-class feast now correctly beats a class-4 feria, 35174
where one of five restored commemorations appears, and 1220 on 13
January. All five restored entries were verified against a photographic
scan -- real, correctly ranked, and confined to their own dates.

13 January is NOT fixed by this. It moves from one wrong answer to a
different one, because RG 112 and a Holy Family office are both unbuilt.
Recorded as its own allow-list entry with a citation and a count, rather
than left inside a season-boundary entry's cardinality pin, and
deliberately not pinned by a golden test.

Then refreshes the lectio differential fixture, which predated the seven
EF defects fixed in that project. Nine allow-list entries close -- each
was a real divergence adjudicated in colitur's favour, and each closes
because lectio was fixed, with its RG citation preserved. The list goes
from twelve entries to four; the remaining 540 differences are fully
attributed with none unexplained.

1 files changed, 113 insertions, 57 deletions

diff --git a/test/test_differential.ml b/test/test_differential.ml
index e12cd9a..89e6873 100644
--- a/test/test_differential.ml
+++ b/test/test_differential.ml
@@ -51,9 +51,20 @@
 
    *** THE THREE-LAYER DESIGN (controller ruling, Task 15 dispatch) ***
 
-   Of 16801 day-pairs (2005-2050), 11206 already match on the seven columns.
-   Of the 5595 that don't, exactly 25 distinct (field-diff) signatures cover
-   all of them (task-15-class-summary.txt). They resolve into three strictly
+   REFRESHED (2026-08-12, task 2026-08-12-colitur-rg16a, branch ef-rebootstrap):
+   the fixture was regenerated from lectio's CURRENT HEAD (3b32c00), which has
+   fixed seven EF calendar defects since the fixture was first pinned (commit
+   2386a45) -- see the fixture's own provenance note for the exact commits and
+   defect list. Of 16801 day-pairs (2005-2050), 11826 now already match on the
+   raw seven columns (up from roughly 11201 against the stale fixture); of the
+   4975 that don't, only 10 distinct RAW (field-diff, including the
+   never-compared week column) signatures cover all of them, independently
+   recomputed against this fixture and colitur's current output, not copied
+   from the original task-15-class-summary.txt (which described the OLD
+   fixture and is stale). After Layers A and B below strip the week column and
+   the vocabulary/numbering artifacts, only 540 of those 4975 remain genuine
+   differences -- all 540 explained by Layer C (368 + 138 + 31 + 3, see each
+   entry's own citation), 0 unexplained. They resolve into three strictly
    separate layers:
 
    - Layer A (this file's [norm_season]/[norm_slug]): vocabulary. A
@@ -67,7 +78,7 @@
      done even where it would shorten the table.
 
    - Layer B ([strip_epiphany_index]): numbering. The ONE case in the whole
-     5595 where a slug's embedded index genuinely cannot be reconciled by a
+     4975 where a slug's embedded index genuinely cannot be reconciled by a
      literal table (Time-after-Epiphany's non-constant offset, limit 3
      above) -- both sides' embedded week digit is stripped to a common
      family+weekday form before comparing, while rank and colour (which
@@ -76,23 +87,48 @@
      family would still be caught by everything except the digit itself.
 
    - Layer C (data/ef/expected-divergences.sexp, matched by
-     [layer_c_reason] below): the CITED allow-list. Eleven genuine liturgical
-     disagreements (C11 added by Task 16's oracle work, below), each citing
-     its RG paragraph and stating which engine is right (always colitur,
-     verified against the Missal/register, never against lectio's own
-     behaviour -- "lectio does it differently" is not itself a justification
-     anywhere in this file). This is the ONLY layer that may cover a
-     difference in rank, colour, or which celebration is observed; A and B
-     never do (enforced structurally below: A/B only ever
-     touch the season/slug fields, and Layer C's predicates each require an
-     exact, narrow field-diff SET, not "anything goes").
+     [layer_c_reason] below): the CITED allow-list. Each entry cites its RG
+     paragraph and states which engine is right (always colitur, verified
+     against the Missal/register, never against lectio's own behaviour --
+     "lectio does it differently" is not itself a justification anywhere in
+     this file). This is the ONLY layer that may cover a difference in rank,
+     colour, or which celebration is observed; A and B never do (enforced
+     structurally below: A/B only ever touch the season/slug fields, and
+     Layer C's predicates each require an exact, narrow field-diff SET, not
+     "anything goes").
+
+   REFRESHED (2026-08-12, ef-rebootstrap fixture regeneration): lectio's own
+   fix wave (2386a45 -> 3b32c00, provenance note has the full commit list)
+   independently fixed the exact defects nine of the thirteen entries this
+   file used to carry were about -- C2, C3, C4, C5, C7, C10, C11, C12 and C13
+   all now match ZERO rows against the refreshed fixture and were CLOSED
+   (removed from data/ef/expected-divergences.sexp and from
+   [layer_c_reason] below), each with its citation preserved and the closure
+   recorded in docs/research/rules-register.md, not silently deleted -- see
+   the report for the full before/after account. C9's old shape ("St Joseph
+   observed ON a Lent Sunday", lectio's own defect 4) also closed the same
+   way, but a NARROWER, different divergence involving the same saint
+   survived the refresh and was re-cited as a new entry, C14 (below), not
+   folded back into the old C9 id (the two are different rules; conflating
+   them would be exactly the "count proving cardinality where identity was
+   required" vacuity flavour this project's harnesses are specifically
+   checked against). Three entries survive unchanged in kind: C1 (368 rows,
+   up from 361 -- the 7 rows C13 used to carve out fold back into C1's own
+   shape now that lectio's slug matches there too, see C1's own note), C6
+   (138 rows, unchanged count but a narrower shape -- lectio fixed the RANK
+   this entry used to also cover, leaving only the SLUG naming difference,
+   see C6's own note) and C8 (31 rows, wholly unchanged -- lectio still
+   computes no Rogation days at all).
 
    13 January (register §6's long-open "Baptism of the Lord" item) is
-   EMPIRICALLY CONFIRMED FIXED, not allow-listed: colitur's slug/rank/colour
-   for 13 January already equal lectio's exactly, in all 46 years (the
-   sanctoral wiring landed in Task 11's "13 Jan now resolves to the Baptism
-   of the Lord" review note). The only residual difference there is season
-   (covered by Layer C's C1, the Jan 6-13 boundary) -- see the report. *)
+   EMPIRICALLY CONFIRMED FIXED, not separately allow-listed, in ALL 46 years
+   again as of this refresh: lectio's own fix wave independently fixed its
+   equivalent of the same RG16(a) defect the now-closed C13 used to allow-
+   list (its slug now matches colitur's `commemoration-of-the-baptism-of-
+   the-lord` on all 7 of the Sunday years too, not only the 39 non-Sunday
+   ones), so the residual difference on every one of the 46 rows is season
+   alone -- covered by C1, the Jan 6-13 boundary, with no separate entry
+   needed any more. *)
 
 module Cal = Colitur_kernel.Calendar
 module Layer = Colitur_kernel.Layer
@@ -122,7 +158,7 @@ let allow_list_path = "../data/ef/expected-divergences.sexp"
    deliberately, after re-running the exact command the provenance note
    names -- never by copying the actual value back in to make a mismatch
    pass, which would defeat the point of pinning it at all. *)
-let fixture_sha256 = "2ca3eeeda4e7a0406c4d004c1b2003fc0df671aca9af18a1b506543a721c8bac"
+let fixture_sha256 = "fc4832f1bc1bdfd322e808ae4929487628981792c23f959c1764c8eaedbfaf7f"
 
 (* Same technique tools/bootstrap_sanctoral.ml already uses for this exact
    purpose (that file's own comment: shelling out to the system's
@@ -382,15 +418,21 @@ let diff_fields (l : row) (c : row) =
 let subset xs ys = List.for_all (fun x -> List.mem x ys) xs
 let day_of_date date = int_of_string (String.sub date 8 2)
 
-let advent_feria_slug slug =
-  List.exists
-    (fun wk -> List.exists (fun wd -> String.equal slug (Printf.sprintf "ef-advent-%d-%s" wk wd)) weekdays)
-    [ 3; 4 ]
-
-let sunday_iclass_slugs =
-  [ "ef-advent-sunday-2"; "ef-advent-sunday-4"; "ef-lent-sunday-1"; "ef-lent-sunday-2"; "ef-lent-sunday-3" ]
-
-let rose_sunday_slugs = [ "ef-advent-sunday-3"; "ef-lent-sunday-4" ]
+(* ef-rebootstrap fixture refresh (2026-08-12): C2, C3, C4, C5, C7, C10, C11,
+   C12 and C13's own predicates/helper bindings (sunday_iclass_slugs,
+   rose_sunday_slugs, advent_feria_slug, fixed_iii_class_reclassified_slugs,
+   jan13_lord_sunday_dates_2005_2050, and the inline C7/C10/C11 date/slug
+   checks) are REMOVED here, not left dead -- each matched zero rows against
+   the refreshed fixture (lectio's own fix wave, 2386a45 -> 3b32c00,
+   independently fixed the exact defects these entries were about), so their
+   citations moved to docs/research/rules-register.md (closure record, with
+   the RG paragraph preserved) and their sexp rows to nothing --
+   data/ef/expected-divergences.sexp no longer declares those ids at all.
+   Removing a predicate strengthens this harness the same way removing its
+   sexp row does: if any of these nine shapes ever reappears (a regression in
+   colitur, or lectio moving again), it now surfaces as UNEXPLAINED instead
+   of being silently re-absorbed by a citation whose own divergence no
+   longer exists. *)
 
 (* Fix round 1, finding 1: C1 and C6 (below) originally gated on calendar
    date alone, with no slug/slug-family check -- unlike every other entry
@@ -413,6 +455,22 @@ let jan_6_13_slug slug =
 let nativity_octave_day_slugs =
   [ "ef-nativity-octave-day-5"; "ef-nativity-octave-day-6"; "ef-nativity-octave-day-7" ]
 
+(* C14 (ef-rebootstrap fixture refresh, 2026-08-12; replaces the closed C9,
+   see data/ef/expected-divergences.sexp's own C14 note for the full RG
+   citation): the exact 3 civil days, across the whole 2005-2050 window,
+   where St Joseph's (19 March, I class) RG 96 transfer walk is congested
+   enough by Passiontide/Holy Week/the Easter octave to cross Easter AND
+   collide with the Annunciation's own transferred "sedes propria" (RG 96
+   Attamen (a), Monday after Low Sunday) -- independently re-derived from
+   this fixture and colitur's own output (`grep joseph-spouse` both sides,
+   filtered to rows with a real post-Layer-A/B diff), not copied from the
+   register's prior "2008, 2035, 2046" prose. A LITERAL date list, not a
+   slug predicate, per the SAME discipline C12/C13 already used (a plain
+   "involves this slug" predicate is exactly what let the old C9 silently
+   cover two textually-unrelated divergences at once -- see C14's own sexp
+   note for the full account of why that was wrong). *)
+let joseph_annunciation_collision_dates_2005_2050 = [ "2008-04-01"; "2035-04-03"; "2046-04-03" ]
+
 (* [layer_c_reason l c diffs] returns the [data/ef/expected-divergences.sexp]
    [id] this row-pair's remaining (post Layer A/B) diff set belongs to, or
    [None] if nothing here explains it (a genuine, uncovered failure). *)
@@ -424,42 +482,40 @@ let layer_c_reason (l : row) (c : row) diffs =
     && subset diffs [ Season; Colour_f; Slug_f ]
     && (not (List.mem Slug_f diffs) || jan_6_13_slug c.slug)
   then Some "C1"
-  else if List.mem c.slug sunday_iclass_slugs && diffs = [ Rank ] then Some "C2"
-  else if List.mem c.slug rose_sunday_slugs && subset diffs [ Rank; Colour_f ] then Some "C3"
-  else if advent_feria_slug c.slug && diffs = [ Rank ] then Some "C4"
-  else if starts_with ~prefix:"ef-lent-ember-" c.slug && subset diffs [ Slug_f; Rank ] then Some "C5"
   else if
     m = 12
     && (d = 29 || d = 30 || d = 31)
-    && subset diffs [ Slug_f; Rank ]
-    && (not (List.mem Slug_f diffs) || List.mem c.slug nativity_octave_day_slugs)
+    (* NARROWED, fixture-refresh review: this was [subset diffs [ Slug_f; Rank ]].
+       Since lectio's own Christmas-octave rank was corrected (3b32c00) all 138
+       rows diff on [Slug_f] ALONE, and the note below says so -- but leaving
+       [Rank] in the accepted set did not merely over-permit, it made the entry
+       BLIND: every row already carries a Slug_f diff, so an added Rank diff
+       changed neither the diff-set membership nor the count. Demonstrated by
+       the reviewer: dropping the RG 91 entry 17 elevation in temporal_ef
+       (Class2 -> Class4 on octave days 5-7) left BOTH differential tests green
+       with C6 still reporting 138. The slug guard below already pins identity;
+       [Rank] only removed the ability to notice a rank regression. *)
+    && diffs = [ Slug_f ]
+    && List.mem c.slug nativity_octave_day_slugs
   then Some "C6"
-  else if (String.equal c.slug "matthew" || String.equal c.slug "thomas") && subset diffs [ Slug_f; Colour_f ]
-  then Some "C7"
   else if
     (String.equal c.slug "ef-rogation-monday" || String.equal c.slug "ef-rogation-tuesday")
     && subset diffs [ Season; Slug_f; Colour_f ]
   then Some "C8"
   else if
-    (String.equal l.slug "joseph-spouse-of-the-bl-virgin-mary"
-    || String.equal c.slug "joseph-spouse-of-the-bl-virgin-mary")
-    && subset diffs [ Season; Slug_f; Rank; Colour_f ]
-  then Some "C9"
-  else if
-    (String.equal l.date "2011-07-02" || String.equal l.date "2011-07-04")
-    && subset diffs [ Slug_f; Rank; Colour_f ]
-  then Some "C10"
-  else if String.equal c.slug "ef-passiontide-2-thursday" && diffs = [ Colour_f ] then Some "C11"
-  (* RG16(a) task, fix round 1 (CRITICAL finding): a C12 entry briefly lived
-     here, allow-listing 2 February on a Sunday against lectio, when
-     `data/ef/adjustments.sexp` retagged the Purification `Bvm`. REVERTED --
-     the user has ruled: follow the oracle (register §6.0 has the full
-     account). The Purification is `Lord` again, colitur observes it
-     outright on those six 2-February Sundays, matching lectio's own
-     `class = lord` exactly -- no divergence to allow-list. Left
-     unimplemented here deliberately, not merely absent by oversight: this
-     comment is the record that C12 existed and was removed, should a
-     future reader wonder why the numbering skips it. *)
+    List.mem l.date joseph_annunciation_collision_dates_2005_2050
+    && subset diffs [ Slug_f; Rank ]
+    (* Colitur-slug guard, added by the fixture-refresh review. Fix round 1's
+       own finding 1 established that a Layer C predicate must pin WHICH
+       celebration the citation is about, not merely the date -- C1, C6 and C8
+       all carry that guard; C14 was written after that ruling and omitted it.
+       Demonstrated live: with [Calendar.compare_deferred] switched to RG 98
+       origin order, the observed celebration on all three dates flips to the
+       Annunciation, and C14 still matched exactly 3 rows with its count pin
+       green. The suite only reddened because the SIBLING dates fell outside
+       every predicate -- geometry, not this pin. *)
+    && String.equal c.slug "joseph-spouse-of-the-bl-virgin-mary"
+  then Some "C14"
   else None
 
 (* ---------------------------------------------------------------------- *)
@@ -539,7 +595,7 @@ let test_dates_align () =
   in
   Alcotest.(check (list (pair string string))) "no misaligned dates" [] mismatched
 
-(* The core assertion: every one of the 5595 raw differences is either
+(* The core assertion: every one of the 4975 raw differences is either
    normalised away (Layers A/B) or named in the cited allow-list (Layer C).
    Nothing else is permitted to pass silently. *)
 let test_no_unexplained_differences () =