Title-case Mac in case repair unless a mask records the name (#1014)

docs/design/decisions/R4-case-repair-reads-the-unjoined-mark.md

@@ -10,7 +10,7 @@
 - Case repair's conjunction test reads the parse's `conjunction` tag rather than re-deriving the word's class (`juan e-f smith` → 'Juan E-F Smith'); the particle test still keys on lexicon membership, by decision, so `replace()` and a parse part there — 2026-08-29 #458.
 - A token carrying `UNCLASSIFIED_TAG` (stamped by `replace()` only) falls back to the vocabulary for the per-word conjunction-or-initial question and never for the per-part particle question: `replace(family='velasquez y garcia')` → 'Velasquez y Garcia' — 2026-08-29 (#458 review).
 - `initials()` has no such fallback, since it is handed no lexicon and guessing the default one erases fields parsed under a custom one; `replace(family='de la vega').initials()` → 'j. d. l. v.' is accepted until #464 gives that view a crossing — 2026-08-29 (#458 review, then the review of PR #463).
-- A single-case word the parse put in the suffix role from the acronym vocabulary repairs to capitals unless a mask records its casing, asked before the Mac/Mc convention (`john smith mba` → 'MBA', `mcse` → 'MCSE'); the gate is the role, so a spliced suffix repairs alike and a name word sharing the spelling does not — 2026-09-13 #459.
+- A single-case word the parse put in the suffix role from the acronym vocabulary repairs to capitals unless a mask records its casing, asked before the Mc convention (`john smith mba` → 'MBA', `mcse` → 'MCSE'; the Mac/Mc convention until 2026-11-02 #1014); the gate is the role, so a spliced suffix repairs alike and a name word sharing the spelling does not — 2026-09-13 #459.
 - The guard on the generation exception is the pair, the suffix role and the `vocab:suffix` tag: a connective read as the generation repairs as the generation (`Carod i` forced → 'Carod I'), while a connective the suffix field merely holds keeps its lowercase (`Smith, John, and` forced → 'John Smith and', a spliced `suffix='i'` → 'i'), and a lone middle connective stays lowercase (`Carod i Rovira`) — 2026-09-20 #461/#397.
 - The exceptions map holds case masks and nothing else, a value being its key's own letters and digits recased, laid over the word as the writer punctuated it wherever the word stands (`phd`, `ph.d.`, `PH.D.` → 'PhD', 'Ph.D.', 'Ph.D.'; `phd smith` → 'PhD Smith'); `md` → 'MD' and `phd` → 'PhD' are deliberate deviations from 1.4.0 through 2.3.0 — 2026-09-23 #459, DECIDED.
 - A value that does not spell its key's letters and digits is a `ValueError` at every `Lexicon` entry point, compared as alphanumerics, whose message offers a recased key that constructs; a mask the word's letters miscount against is declined and the next clause repairs — 2026-09-23 #459, THE CHECK.
@@ -29,6 +29,7 @@
 - A mask outranks the particle lowercase, in the shipped map (`SEÁN Ó MURCHÚ` → 'Seán Ó Murchú', with `ní` and `ua`) and in a caller's (`van` → 'Van' gives 'Ludwig Van Beethoven'); `mac` and `binti` carry none and stay lowercase — 2026-10-04 #604.
 - A listed acronym credential joined by a hyphen or slash repairs whole in capitals in the suffix role (`john smith, otr/l` → 'OTR/L', `john smith, fnp-c` → 'FNP-C', which was 'Fnp-C' before), and a recorded casing for the whole token wins in any role — 2026-10-21 #808.
 - A word's full stops divide it into runs written as words of their own (`j.r.r. tolkien` → 'J.R.R. Tolkien', `st.john smith` → 'St.John Smith', forced `J.Smith` → 'J.Smith'), and a word opening with a caseless modifier letter takes its capital on the letter after it (`AHMAD IBN-ʿALĪ` → 'Ahmad ibn-ʿAlī', `ʻIOLANI` → 'ʻIolani'), a digit not being passed over (`3rd`); the cost, an unlisted abbreviation whose ending follows its full stop reading as two words (the Sardinian `DUT.RA` → 'Dut.Ra', where 1.4.0 through 2.5.0 gave 'Dut.ra'; and an abbreviated second word reads as its spaced spelling does, `dr.philol.` → 'Dr.Philol.', where they gave 'Dr.philol.'), is accepted, a listed casing being the remedy, which the Italian abbreviated salutations carry (`SIG.RA` → 'Sig.ra', `dott.ssa` → 'Dott.ssa') — 2026-10-22 #838; 2026-10-22 #840.
+- Case repair title-cases a word opening Mac and gives it an inner capital only from a mask, the Mc convention standing (`JOHN MACHADO` → 'John Machado', `JAMES MACKENZIE` → 'James Mackenzie', `RONALD MACDONALD` → 'Ronald MacDonald' by its mask); the shipped masks are the Mac surnames written MacX by most of their bearers, at least three writings, in en.toml as `family` readings carrying `cased`, and again in gd.toml and ga.toml by origin, a chunk after a full stop taking its own mask as a hyphen part does — 2026-11-02 #1014 (Derek), DECIDED.
 
 Not standing: the rewording of rules/R4's vocabulary promise and its `Juan McDonald` row, carried out — 2026-08-29, DEBT DISCHARGED; the correction of mechanisms.md#RENDER-HONORS-THE-PARSE, carried out — 2026-09-23, mechanisms; the whitespace raises the branch review left for the orchestrator to file, filed and fixed as #541 — 2026-09-24 #459/#492/#478 (branch review); the measurements, which decide nothing — 2026-09-23 #459/#478.
 
@@ -113,3 +114,4 @@ Accepted costs, deferred to the rescoped #459 rather than relitigated here: the
 - 2026-10-22 #840 — THE REMEDY FOR #838's ITALIAN COST SHIPS: the abbreviated salutations and letter-heading honorifics whose ending follows their full stop are titles with `cased` masks (decisions/salutation-titles), so `SIG.RA MARIA ROSSI` repairs to 'Sig.ra Maria Rossi' again, as 1.4.0 through 2.5.0 gave it, and `dott.ssa` to 'Dott.ssa'. The mask is asked before the chunk split, the precedence the 2026-10-22 #838 entry relied on. The accepted cost stands for an unlisted ending (the Sardinian `DUT.RA` → 'Dut.Ra', CLDR's `Dut.ra Prof.ra`). A spaced `Sig. ra` is out of reach: it joins as a title run, `ra` being a listed title, and forced repair writes 'Sig. Ra'.
 - 2026-10-28 #357 — THE SHIM'S HALVES OF THIS RULE'S ENTRIES GO WITH THE SHIM (decisions/3-0-reevaluations' 2026-10-28 #357, REMOVAL entry): the v1 pickle load as a second producer of `UNCLASSIFIED_TAG`, leaving `replace()` the one; the first-parse raise for a v1 `Constants` carrying a mismatched mask (2026-09-23 #459, A DECIDED EXCEPTION), the core `Lexicon` raising on it as before; and the filter of v1 `Constants` entries with its folded cross-set checks (2026-10-02 #541, 2026-10-02 #582). A punctuated mask such as 1.4.0's `M.D.` stays legal in a `Lexicon`. SUPERSEDES: the 2026-09-23 #459, A DECIDED EXCEPTION, 2026-10-02 #541 and 2026-10-02 #582 entries, all about translating a v1 `Constants`, and the 2026-08-29 (#458 review) entry's second producer, the v1 pickle load.
 - 2026-10-28 #939 — A CONNECTIVE BETWEEN PARTICLES: rules/R4's all-particles Except sets aside a connective that is not also particle vocabulary, standing between the part's words, as rules/R2 does, so `FRITZ FREIHERR UND VON BAR` should repair to 'Fritz Freiherr und Von Bar' where it gives 'Fritz freiherr und von bar', a deviation #939 tracks; decided and measured in decisions/R2's 2026-10-28 #754 (Derek), A CONNECTIVE AMONG PARTICLES entry.
+- 2026-11-02 #1014 (Derek), DECIDED — A NAME OPENING MAC TAKES ITS INNER CAPITAL FROM A MASK, NOT FROM ITS SPELLING. The fall-through had capitalized after any `ma?c` prefix with two or more letters behind it since 0.2.3, so a single-case `MACHADO` repaired to 'MacHado', `MACIAS` to 'MacIas', `MACIEJEWSKI` to 'MacIejewski'; #56 asked in 2015 for an exceptions list beside the regex, and only the length check shipped. Measured over the mixed-case writings of every word the regex matched, each counted as its writer typed it, in the ORCID family-name box and the names of the UK Electoral Commission donors, GOV.UK people and ClinicalTrials.gov officials local corpora (2026-11-02): 4,616 writings of 970 words, the old convention matching the writer's spelling in 16.3% and plain title case in 83.7%; over the same four sources the Mc convention matched 96.8% of 6,312 Mc writings (94.3% of ORCID's 2,764), the rest mostly names Mc opens that are not Gaelic, `Mchunu` 24 the commonest, then `Mchedlishvili` and `Mchome` 5 each. The Gaelic Mac names split by family, so no shape decides them: `MacDonald` 143 against `Macdonald` 77, `Mackenzie` 59 against `MacKenzie` 44, `Mackay` 73 against `MacKay` 26, `MacLeod` 39 against `Macleod` 18. The mask records the spelling where its bearers lean one way: a word written MacX in more than half of at least three writings carries one, 42 words, which brings the measured match to 91.2% (in-sample, since the masks were chosen from the writings they are scored on; at least two writings gave 60 masks and 92.0%, at least five 25 and 90.2%). Recipe: for each source, read the field named (`family` for ORCID's `raw.jsonl`, `name` for the other three) and take every match of `(?<![^\W\d_])(Mac[^\W\d_]{2,})(?![^\W\d_])` whose letters after `Mac` are not all capitals; key it by its lowercase and count it MacX where its fourth letter is a capital; the comparators are the old rule (every word MacX), plain title case (none), and title case with a mask on each word that is MacX in more than half of at least three writings; the Mc figures are the same with `Mc` for `Mac` and the third letter. The masks stand on `family` readings, the first `cased` a name reading carries: an entry holding only such a reading compiles, so the build does not warn it inert. They are in en.toml, the anglicized spellings being English ones (the Gaelic writes `MacDhòmhnaill`, the Irish `Mac Cárthaigh`), and again in gd.toml, 37 of them, and ga.toml, the Irish `maccarthy`, `macmahon`, `macmanus` and `macdermott`, so a strict choice of a Gaelic pack, whose bearers are likeliest to be written so, keeps them: `parser_for(locales.GD, strict=True)` repairs `RONALD MACDONALD` to 'Ronald MacDonald' (decided, Derek, 2026-11-02). `mackrill`, an English name, is in en.toml alone. Every plain choice keeps all 42, no file holding another reading of them; a strict choice holding none of the three files keeps none, and a strict GA choice keeps only the Irish four (`RONALD MACDONALD` → 'Ronald Macdonald'). A chunk of a word that a full stop divides takes its own mask, as a hyphen part does, so the masks reach `J.MACDONALD SMITH` → 'J.MacDonald Smith', which the old convention gave and #838's parity between a chunk and a hyphen part kept; credential chunks take theirs too (`Phd.Md` forced → 'PhD.Md', `Msc.Ed.` forced → 'MSc.Ed.'): of the readings of ClinicalTrials.gov's officials that move, 125 forced and 1 default are of names holding no `mac` in any case, all credential chunks, against 611 forced and 8 default of names holding one (`snapshot.py --local ctgov-officials --base` the parent commit, 2026-11-02). A caller's lower-case mask on a one-letter or particle word reaches a chunk too (`{'y': 'y'}` gives `J.Y. SMITH` → 'J.y. Smith', where it gave 'J.Y. Smith'), as it reaches the spaced spelling ('J. y. Smith'); no shipped mask does. The cost is a Gaelic name written MacX that no mask records, `SHIRLEY MACLAINE` → 'Shirley Maclaine', and a family that writes Macdonald, which a forced repair recases to the mask. A caller's own spelling goes in `Lexicon.capitalization_exceptions`.

docs/design/decisions/R5-case-repair-gate.md

@@ -2,8 +2,8 @@
 
 ## Standing
 
-- Case repair acts only on a name written entirely in one case, and a name whose words outside the suffix are written in more than one case is kept as written, whether that casing is right not entering into it (`juan mcdonald` → 'Juan McDonald', `SHIRLEY MACLAINE` → 'Shirley MacLaine', `Shirley Maclaine` kept), unless repair was asked for anyway (`Shirley Maclaine` forced → 'Shirley MacLaine'); the refusal predates the git history and the override shipped in 0.4.0, and the suffix's place in the test is the 2026-09-23 #492, DECIDED line's — 2026-08-29 (#407 arc); 2026-08-29, WHY R5'S STATEMENT WAS REWORDED; 2026-09-23 #492, DECIDED.
-- The boundary example of rules/R5's main fence is the forced `Shirley Maclaine` line, the refusal being R5's effect rather than its absence, and no example witnesses an already-correct mixed-case name, kept and repaired being the same string there — 2026-08-29, why the gap survived; 2026-08-29, WHY R5'S STATEMENT WAS REWORDED.
+- Case repair acts only on a name written entirely in one case, and a name whose words outside the suffix are written in more than one case is kept as written, whether that casing is right not entering into it (`juan mcdonald` → 'Juan McDonald', `RONALD MACDONALD` → 'Ronald MacDonald', `Ronald Macdonald` kept), unless repair was asked for anyway (`Ronald Macdonald` forced → 'Ronald MacDonald'); the refusal predates the git history and the override shipped in 0.4.0, and the suffix's place in the test is the 2026-09-23 #492, DECIDED line's — 2026-08-29 (#407 arc); 2026-08-29, WHY R5'S STATEMENT WAS REWORDED; 2026-09-23 #492, DECIDED; the examples' name, 2026-11-02 #1014.
+- The boundary example of rules/R5's main fence is the forced `Ronald Macdonald` line (the forced `Shirley Maclaine` until 2026-11-02 #1014), the refusal being R5's effect rather than its absence, and no example witnesses an already-correct mixed-case name, kept and repaired being the same string there — 2026-08-29, why the gap survived; 2026-08-29, WHY R5'S STATEMENT WAS REWORDED.
 - The override has one route, `capitalized(force=True)` per call, the v1 `force_mixed_case_capitalization` attribute having gone with `Constants`; the rule names none — 2026-10-28 #357.
 - `capitalized_forced` is a test-side pseudo-field the doc runner resolves, not a parser field — 2026-08-29, `capitalized_forced` is a test-side pseudo-field.
 - Forcing repair does not make the output independent of how the input was cased, so rules/R5 does not state that it does; the property is pinned in tests/test_capitalization.py, where `JUAN Y GARCIA` and `juan y garcia` both give 'Juan y Garcia' since a one-case name stopped reading a bare capital connective as an initial — 2026-08-29, WHAT THE OVERRIDE DOES AND DOES NOT PROMISE; 2026-08-29, and therefore NOT stated in rules; 2026-09-13 #383/#479.
@@ -42,4 +42,4 @@ Open: none in this file.
 
 - 2026-09-24 #492 — DECIDED (Derek, 2026-09-24): where repair was not forced, a SUFFIX token written in more than one case is kept as written; one written in a single case is repaired like any other token, and force repairs both, unchanged. It is the gate's own one-case test asked of each suffix on its own (`_render._in_one_case`, one predicate for both), and it is the 2026-09-23 #492 reasoning read to its end: a suffix is left out of the gate because a cased one says nothing about how the writer cased the NAME, and for the same reason a mixed-case one is the writer's spelling of the credential, which repair has no evidence to overrule. THE REGRESSION THAT SURFACED IT, measured by the branch's code reviewer and re-measured 2026-09-24: once the gate stopped counting the suffix, a one-case name with a conventionally mixed-case listed acronym was admitted and the acronym clause wrote the acronym in capitals — `john smith EdD` → `John Smith EDD`, `JANE DOE, DSc` → `Jane Doe DSC`, `juan garcia PsyD` → `Juan Garcia PSYD` — where the parent 4d0680e6 and every release 1.4.0 through 2.3.0 returned those names untouched. They now give `John Smith EdD`, `Jane Doe DSc` and `Juan Garcia PsyD`. The accepted `B.Tech.` → `B.TECH.` on the default path (decisions/R4, 2026-09-23) was the same shape, and its default-path half is superseded (decisions/R4, 2026-09-24). THE COST, accepted: "as written" keeps a garbled spelling with the deliberate one, so `juan garcia Iii` gives `Juan Garcia Iii` on the default path (`Juan Garcia III` forced) and `john smith Mba` gives `John Smith Mba`, where the branch before this decision gave `III` and `MBA`; every release returned both names untouched. A rules/R5 boundary line pins `Iii`. Why not keep only a spelling the vocabulary does not know how to repair: that would re-derive, per word, a judgment about which mixed spellings are deliberate, which is the evidence R5 says repair cannot read. The masks added the same day (decisions/R4) are the other half: they serve the one-case names this rule cannot reach (`john smith psyd` → `PsyD`). No corpus name written before this decision moves by it (measured 2026-09-24 against the branch before it, over this tree's 1375 names: the only movers are the new rules.md lines). Pinned by `tests/v2/test_render.py::test_a_suffix_written_in_more_than_one_case_is_kept_as_written`, core and facade; a mutant dropping the per-token test fails it, `test_the_gate_leaves_the_suffixes_out` and two rules/R5 lines, and one dropping its `force` guard fails it and the forced `EdD` line.
 - 2026-10-28 #357 — THE OVERRIDE HAS ONE ROUTE (decisions/3-0-reevaluations' 2026-10-28 #357, REMOVAL entry). The 2026-08-29 entry recorded two, `capitalized(force=True)` and the v1 `force_mixed_case_capitalization` attribute the facade resolved when no per-call value was given; the attribute went with `Constants`, as decisions/3-0-reevaluations' 2026-10-28 #925 (Derek), WARNINGS entry foresaw. rules/R5's statement, written to cover both, needs no change to cover one. SUPERSEDES: the 2026-08-29 entry recording the override's TWO routes.
-
+- 2026-11-02 #1014 — THE EXAMPLES' NAME: rules/R5's fence named `Shirley Maclaine`, whose forced repair to 'MacLaine' was the Mac convention's, which decisions/R4's 2026-11-02 #1014 entry retires; with no mask for `maclaine` the forced line would return the name unchanged and witness nothing, so the fence names `Ronald Macdonald`, whose forced repair writes its mask, 'MacDonald'. The gate is unchanged.