34. French verse is counted a line at a time, and the count is an estimate¶

Context¶

ADR 0032 shipped the French lexicon and stopped at 80 of 121, leaving 41 rows blocked and dividing them: 18 want stress, which French does not have, and 23 want a syllable count. Its D5 declined to reach those 23 and said why — Lexique's orthosyll, phon and syll columns were present and deliberately unused, because the 23 need a line-level count and not a per-word one, and that was "a separate decision, gated on a separate measurement". This record is that decision, and the measurement did not come out well.

French cannot be counted word by word. denckring.procedures.syllable_count summed pack.syllable_count(word) over a line, which is right for English and German and wrong here. A French final mute e is a syllable before a consonant, elides before a vowel or a mute h, and never counts at the end of a line. Lexique gives femme, une, belle and porte all nbsyll=1 in citation form; in verse each is frequently 2. French verse is entirely syllable-counting, so alexandrine — the one row named after a French metre — would have been wrong about most real French even with a perfect dictionary.

The go/no-go was the alexandrine share, and it was chosen before the work started. The chapter spec made D3 conditional on one figure: the share of Racine's and Hugo's alexandrines that the line counter scores at exactly twelve. That figure is pass/fail per line rather than a coverage percentage, which is why it was picked; the bar was 94%, and it was not reached. See D5.

Decision¶

D1. A pack answers for a whole line, and the default is today's per-word sum. LanguagePack.line_syllables(line) -> (total, estimated) joins the protocol. BasePack implements it as the loop syllable_count.line_syllables used to run inline, so English and German are unchanged by construction — neither pack overrides it — and tests/test_line_syllables.py pins both halves of that: the default equals the per-word sum, and English's line counts over a fixture are what they were. FrenchDataPack overrides it with the elision rules.

The seam lives on the pack rather than in the procedure because "how long is this line" is a question about the language, and the procedure is the wrong place to know that a mute e exists. It is the same move ADR 0030 made for is_vowel_phoneme after assonance_constraint answered a phonological question with CMUdict's convention and was wrong in German. The 23 syllabic rows then run in French without one of them being edited, which was the test of whether the seam sits in the right place.

D2. French declares syllables, syllables.dictionary, syllables.heuristic and phonemes, from a table of 125,653 rows. syllables.txt.gz carries (nbsyll, SAMPA phon, orthosyll) per spelling. French is the first pack in this project that can honestly declare syllables: the capability means a segmentation of the written word, orthosyll is exactly that (car-ros-se), and ADR 0030 removed the same capability from denckring-en-data, which had declared it and never implemented it. German declines it still, because its transcriptions carry no division of the spelling.

D3. stress is still refused, and 103 is no longer a projection. ADR 0032 D5 and spec D2 both stated 103 as the honest ceiling. It is now a measurement: French runs 103 of 121, the remaining 18 rows are blocked on stress and on nothing else, verified mechanically against meta.requires rather than asserted. They are accentual metres — the sonnet family, iambic pentameter, dactylic hexameter, sapphic, alcaic, ottava rima, rhyme royal, spenserian, heroic couplet, trochaic tetrameter, double dactyl, elegiac couplet, curtal sonnet, ballade, blank verse, proteus verse — and they are not French forms. Declaring stress to reach a larger number would be the false-capability defect ADR 0030 fixed twice in one day.

D4. The French line count is exact=False, always. ADR 0012's flag already exists to say so and every syllabic report carries estimated_words. Diérèse and synérèse are not decidable from spelling — extraordinaire is four syllables or five depending on the poet — and D5 is the evidence rather than the assertion. plint, the mature tool in this space, calls its own handling "a liberal estimate" for the same reason.

D5. A diérèse rule ships, and it is a STOP, not an ACCEPT. Scored over lines interior to a speech block, because the first and last line of a speech may be half an alexandrine shared between speakers — Hugo's unfiltered run has a 266-line spike at exactly 6, the caesura — and excluding any line holding a word Lexique does not carry, which cannot test an elision rule:

	baseline	shipped	ceiling
Racine, Mithridate (n=1213)	89.3%	90.3%	96.6%
Hugo, Hernani (n=1283)	79.3%	80.2%	84.8%

The ceiling column is a perfect per-site oracle: what the counter would score if every diérèse decision in the text were made correctly by an editor. A mechanical rule captured 1.0 of the 7.3 points that oracle would take on Racine. The accept bar was 94%. Three variants were measured and none reached it, so the chapter stops here with what it has rather than spending more of the budget or lowering the bar.

The shipped rule is narrow: every j after any consonant is one extra syllable, gated on the word ending in the -ion(s) suffix or starting with dia-. It generalises further than its 90.3% suggests — it fires on 18 distinct words across both poets and 23 of its 27 firings are correct — but it is not a solved rule and the count stays estimated (D4).

Record the rules that failed, because they are the plausible ones. The chapter plan's own proposed starting rule — a glide after a consonant cluster ending in a liquid — scored 84.5% / 74.7%, roughly five points below the baseline of doing nothing. Two blanket probes were catastrophic (49.5% / 36.6% and 46.2% / 33.8%), because the plurality of every "glide after a consonant" site in Lexique is a verb or adjective ending (-iez, -ions, -ier, -ien) that is reliably synérèse. Anyone proposing the cluster rule again should read this paragraph first; it sounds right and it is worse than no rule at all.

D6. The aspirated-h list is vendored data, extracted from frwiktionary and expanded through Lexique. Lexique cannot distinguish aspirated from mute h — haricot is aRiko and hôtel is otEl, the letter simply dropped — and the distinction decides whether the schwa in front of it elides. h_aspire.txt.gz holds 3,370 entries: 3,285 from a scan of {{h aspiré}} in the French section of each page, plus 85 from expanding those lemmas through Lexique's lemme column. The expansion looks negligible and is not: Wiktionary already marks most inflected pages directly, so expansion adds only what it missed, and what it missed includes hais — the high-frequency verb forms are exactly the ones that appear in verse. Only 324 of the 3,370 entries are known to Lexique as lemmas at all, and those expand to 740 forms. ({{asp|fr}}, which an earlier draft guessed at as an alternative template, occurs zero times in the dump.)

D7. One table row per spelling, chosen by freqlivres. Lexique carries several rows per homograph, and the pack is asked about a spelling rather than about a reading, so the most frequent reading wins. The cost is stated in Consequences rather than hidden here.

Consequences¶

French runs 103 of 121, up from 80, and denckring eval --all goes from 439 to 487 passing cases, the 48 new ones French across the 23 unblocked rows. denckring status is unchanged at 155 · 130 · 121 · 121 · 25: this record adds no catalogue row and implements no procedure. It moves one language's coverage and nothing else.

parent the verb is counted as parent the noun. D7's consequence, and it is not a rare shape: French homographs that differ in pronunciation across parts of speech (président, fils, est, parent) are counted, transcribed and rhymed as whichever reading freqlivres ranks first. A caller asking for the syllable count or the rhyme key of the loser gets a confident wrong answer rather than a refusal, which is this project's least favourite failure shape. Fixing it needs a part-of-speech capability no pack provides — the same blocker homosyntaxism, verbless_prose and chimera already sit behind — so it is recorded, not solved.

extraordinaire is a known mute-e miss and is inside the measured 90.3%, not against it. Lexique gives it nbsyll 5 and orthosyll ex-traor-di-nai-re, also 5, because traor covers two phonetic syllables in one orthographic segment. The mute-e test compares the segment count against nbsyll and so cannot see the final e here. The failure is silent: the word contributes 5 where verse frequently wants 6, and nothing in the report says which word was wrong. Words shaped like this are the residue the ceiling column of D5 measures.

Schwa is rendered /ø/. Lexique writes le as l2, and SAMPA 2 spells both /ø/ and the schwa; the table maps it uniformly, so phonemes("le") is ['l', 'ø'] where a French phonology reference would say /lə/. This is faithful to a source that cannot distinguish the two, and it is harmless for rhyme and assonance, where they are near-merged — but a caller reading phonemes() as a phonological transcription is being told something slightly false, and finding that out by inspection would be worse than reading it here.

syllables.heuristic is now satisfiable through line_syllables as well as through syllable_count, and the guarantee behind it is weaker than it was. Before this record the capability meant one method. French's override reads its own tables directly rather than calling syllable_count, so a pack can now satisfy the capability by two independent paths and a row that declares it has no way to say which one it will get. tests/test_requires_honesty.py tracks both names and a new test asserts the two hand-maintained capability tables agree — adding the method to one table alone silently disarmed the sufficiency check for eleven rows once already. No pack abuses the looser contract today; nothing prevents one.

No installed pack lacks phonemes any more, so that refusal has no in-suite example. tests/test_prosody_robustness.py covered "a pack that genuinely cannot answer" with French, because French had no pronunciation data; German had held the same role until chapter 4 gave it a dictionary. That test now uses sonnet and stress, which is a real refusal but a different one. The phonemes refusal path is exercised only by the french-without-data CI job on a core-only install, which is a job that can be deleted by someone who does not know it is load-bearing. Its comment says so.

Feminine/masculine rhyme alternation is not modelled, and French classical prosody requires it. rhyme_key returns the rime suffisante baseline — phonemes from the last vowel to the end of the word, since French has no lexical stress to anchor a stronger definition. A French classical poem that never alternates feminine and masculine rhymes is not a well-formed poem, and rhyme_scheme will pass it. This is a real gap and a different one: the project models the alternation in no language, so it is not something French is missing relative to English or German.

The caesura is unchecked. alexandrine counts twelve syllables and says nothing about the hemistich. plint does not check it either, for the reason it gives: the hemistich has to fall at a sensible grammatical position, which is not decidable from the data here.

A word outside Lexique degrades rather than refuses, and the two behaviours now sit side by side. syllables() and phonemes() raise MissingCapability for a word the table does not carry; line_syllables falls back to a vowel-run estimate and increments estimated_words instead, because one unknown word must not blank the count for a whole line of real verse. Both are defensible and they are not the same rule, so a caller comparing a per-word answer against a per-line one can see a word counted in one and refused in the other.

Alternatives considered¶

Leaving the sum in syllable_count.line_syllables and special-casing French there. It would have put a fact about French mute e in a procedure module that has no other business knowing one, and the next language with a line-level rule would have added a second branch beside it. The seam on the pack costs a protocol method and one default implementation.

A diérèse parameter, letting the caller declare the poet's intent. A reasonable future request and not designed here. It would turn D4's estimate into an exact count for a caller willing to annotate, which is the honest shape for a decision that is editorial — but it is a parameter on 23 rows and its own decision.

Vendoring a G2P engine (phonemizer/espeak) rather than a table. A runtime dependency, TTS-oriented, and it answers pronunciation rather than the orthographic syllabation D2 rests on. It would not have improved D5, which is bounded by editorial intent and not by phonetic accuracy.

Deriving the elision rules from plint. GPLv3 against this project's Apache-2.0. Its documented rules are facts about French and were learned from; its implementation was not read, and none of it is here.

Stopping the chapter at 80. The genuine option, and the one the go/no-go existed to force. What was measured is a diérèse bar that D5 misses, on top of a mute-e baseline of 89.3% on Racine that the same harness reports. The chapter ships because that baseline is the bulk of what the 23 rows depend on and because exact=False is a truthful label for it — not because 90.3% cleared anything. Had the baseline itself been the poor number, this record would have said the chapter stopped, and no comparison against a naive per-word sum was measured to make the baseline look better than it is.