afrikaans: a word-final ⟨c⟩ is [k], not the soft [s] (#757)

[christopherthompson81]

Closes #757.

The call

The rule is "soft [s] BEFORE a front vowel". Word-finally there is no following vowel, so the soft branch cannot apply — but it did:

out += "eiyêéè".includes(w[i + 1] ?? "") ? "s" : "k";   //  includes("") === true

franc → frans, arc → ars, bloc → blɔs. #756 preserved that verbatim, correctly — a data move must not change output — and left the decision to this issue.

The decision is that the accident inverted the rule rather than extending it: no following letter means the soft condition is unmet, not met. That is an argument from the rule's own statement, and it needs no corpus.

⚠ Not referee-confirmed, and it cannot be

af.wiktionary-af.tsv contains no word-final ⟨c⟩ at all… the letter name C → sɪə reaches output through the initialism path, not this rule. Both halves of that were wrong — see the addendum.

The corpus's one word-final ⟨c⟩ is the letter name C → sɪə, and it does traverse this branch. It simply misses both ways (s, now k, against sɪə), which is why the score is unmoved.

So af is unmoved at 1658/2220 (74.7%), exactly as predicted. The new goldens pin the rule being self-consistent and say so in the comment; they are not evidence of a measured improvement, and shouldn't be read as one.

Found along the way, deliberately not bundled: #758

The ⟨c⟩ code rule runs before the fixed table, so "ch": "ʃ" never fires — chemie → kɦiəmi, chirurg → kɦirœrχ. kɦ- is not a possible Afrikaans onset, and unlike #757 this one is referee-visible.

I measured the naive fix (let the digraph win): +1 word, but it breaks Christus → ʃrəstœs (should be [kr-]) and pushes chemie to [ʃ] where the referee wants [x]. Afrikaans ⟨ch⟩ is three sounds by source — [ʃ] French, [x] Greek/Dutch, [k] before ⟨r⟩ — so it needs a conditioning decision, not a reordering. Filed as #758 with those numbers so the next attempt starts from data.

npm run ci green — 228 files / 3141 tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr

---

Review addendum — my stated reason was false

I claimed the referee holds no word-final ⟨c⟩, and that its C → sɪə entry reaches output via the initialism path. Both wrong, and the review caught it:

• C is a word-final ⟨c⟩, and eval.ts feeds af through phonemizeWord directly, with no normalization pass — so that entry goes through this exact branch, s → k with this change.
• The initialism normalizer only fires on runs of ≥2 letters, so a lone C never reaches it.

My original grep was case-sensitive and lowercase-only. I then saw the C entry in later output and explained it away with an assumption I never checked — the same shape of error as the "corpus isn't in-repo" claim in #748.

The conclusion survives: s and k both miss against sɪə, so the folded backbone is unmoved at 1658/2220 and this change remains rule-consistency rather than a measured gain. But the reason has been corrected in the test comment, because as written it would have told the next reader that nothing exercises the branch.

The underlying gap — a bare single letter read as a sound instead of spelled as its name (C → k, B → p) — is filed as #761.

Test also moved into the g2p describe block, where it belongs.