fix(of): correct subject over-match, provenance truth, and stale hazard comment

Four fixes from fix round 2 review:

classify_subject's "maria"/"b.m.v" substring test over-matched 12 saints
who merely carry "Maria" in their own name (Maximilian Mary Kolbe, John
Mary Vianney, Mary Magdalene among them) as Bvm. Replaced with an
explicit BVM_DATES table of the 13 genuinely Marian entries. Bvm 25 -> 13.

The provenance header claimed lectio is "never a source", which omitted
that slugs and English names both come from it. Corrected to state
exactly what does (slugs, en names) and does not (every substantive
field) come from lectio. A missing lectio file used to degrade silently
to a different SHA-256 with no en names and no warning; it now fails
loudly. lectio's own SHA-256 is now pinned alongside the Missal PDF's.

Four more entries carried a second pdftotext artifact distinct from the
letter-spacing already handled -- a combining diacritic rendered as a
stray spacing character (Kim Taegŏn/Chŏng Ha-sang, Đũng Lạc, Makhlūf,
Kęty). Hand-repaired the same way Peter Damian was, each corroborated by
the identical corruption recurring in the Missal's own index.

precedence_of.ml's PHASE 3 HAZARD comment about All Souls (2 November)
was asserting a gap that no longer exists now that OF sanctoral data
ships. Added test_all_souls_beats_an_ordinary_sunday, which resolves the
shipped All Souls entry against a real per-annum Sunday (2025-11-02)
through the actual Precedence.resolve pipeline, and updated the comment
to record the resolution instead of demanding it.

1 files changed, 28 insertions, 10 deletions

diff --git a/test/test_calendar_of_data.ml b/test/test_calendar_of_data.ml
index d4f8a8b..9261ebc 100644
--- a/test/test_calendar_of_data.ml
+++ b/test/test_calendar_of_data.ml
@@ -69,14 +69,16 @@ let sha256_of_file path =
    don't transcribe it from output" pinning discipline. *)
 let test_sha256 () =
   Alcotest.(check string) "data/of/calendar-2002.sexp SHA-256"
-    "1b4f2117133dea1c70989708d94e6f292f3ce23fce6221391191d6fcf6042d19" (sha256_of_file path)
+    "91479d8b5d92d8e35c2ae34f239d5623a2a59cca2f02d51a8cbbad52edd46b3a" (sha256_of_file path)
 
 (* Counts independently re-derived from the shipped file
-   (`grep -c '(slug'`; `grep -oP '(?<=\(rank )[A-Za-z_]+' | sort | uniq -c`).
-   206 = 209 raw table rows minus the 3 dates (1 Jan, 6 Jan, 25 Dec) already
-   computed by Temporal_of.named -- see test_excludes_temporal_of_dates
-   below, which closes the loop on that exclusion rather than merely
-   asserting the count moved. *)
+   (`grep -c '((slug'`; `grep -oP '(?<=\(rank )[A-Za-z_]+' | sort | uniq -c`).
+   208 = 209 raw FIXED-date table rows minus the 3 dates (1 Jan, 6 Jan, 25
+   Dec) already computed by Temporal_of.named (see
+   test_excludes_temporal_of_dates below, which closes the loop on that
+   exclusion) PLUS the 2 movable entries fix round 1 added (Sacred Heart
+   of Jesus, Immaculate Heart of Mary -- see test_movable_entries): 209 -
+   3 + 2 = 208. *)
 let test_load_and_counts () =
   let l = load () in
   Alcotest.(check int) "208 entries" 208 (List.length l.L.entries);
@@ -92,7 +94,19 @@ let test_load_and_counts () =
   Alcotest.(check int) "208 Feast (OF admits no Commemoration_only status)" 208
     (count (fun e -> e.L.cel.Cel.status = Cel.Feast));
   Alcotest.(check int) "every entry tagged Precedence_of.universal_layer" 208
-    (count (fun e -> e.L.cel.Cel.layer = PO.universal_layer))
+    (count (fun e -> e.L.cel.Cel.layer = PO.universal_layer));
+  (* Fix round 2: classify_subject's "maria"/"b.m.v" SUBSTRING heuristic
+     over-matched 12 saints who merely carry "Maria" in their own name
+     (Maximilian Mary Kolbe, John Mary Vianney, Mary Magdalene, ...) as
+     Bvm -- replaced with an explicit BVM_DATES table of the 13 genuinely
+     Marian entries (12 fixed + the movable Immaculate Heart of Mary).
+     Pinned here so a regression to the old substring rule -- or a new one
+     with the same failure shape -- reddens immediately rather than only
+     showing up if/when a future Subject.Bvm-keyed rule is added. *)
+  Alcotest.(check int) "13 Bvm (12 fixed + Immaculate Heart of Mary)" 13
+    (count (fun e -> e.L.cel.Cel.subject = Sub.Bvm));
+  Alcotest.(check int) "9 Lord" 9 (count (fun e -> e.L.cel.Cel.subject = Sub.Lord));
+  Alcotest.(check int) "186 Saint" 186 (count (fun e -> e.L.cel.Cel.subject = Sub.Saint))
 
 let test_slugs_unique () =
   let l = load () in
@@ -109,9 +123,13 @@ let test_slugs_revalidate () =
       | Error msg -> Alcotest.failf "slug %s failed re-validation: %s" (S.to_string e.L.cel.Cel.slug) msg)
     l.L.entries
 
-(* Every date is a Fixed month/day; DS.fixed already rejects an out-of-range
-   day against the leap-year maximum, so this also proves no entry claims a
-   nonexistent day (e.g. 30 February). *)
+(* Every FIXED-date entry's month/day pair already went through DS.fixed,
+   which rejects an out-of-range day against the leap-year maximum, so
+   resolving every entry (fixed AND the 2 movable Easter_offset ones, via
+   DS.resolve's own Easter_offset branch) in a real leap year also proves
+   no fixed entry claims a nonexistent day (e.g. 30 February) and that both
+   movable entries resolve at all. Not "every date is a Fixed month/day"
+   (stale since fix round 1 -- 2 of 208 are Easter_offset). *)
 let test_dates_resolve_in_a_leap_year () =
   let l = load () in
   List.iter