The guard reported zero on 32 French values: short noun phrases were invisible (ZED-69) - #58

Closes ZED-69. Rebased onto master and retargeted after #54 was squash-merged.

ZED-50's [baseline] is empty and its docstring read that as "tests/ carries
exactly one French line". True of what the two signals detect, false of the
tree: ten test modules hold short French noun phrases — the description and
owner values that get copied verbatim into the golden trees — that both
signals call clean. Boutique en ligne has no accent and one function word;
Equipe Plateforme has neither. ZED-27 and ZED-60 found every one of them by
reading diffs.

The third signal

Lowering MIN_WORDS_PER_LINE to 1 is not on the table — ZED-50 tried it and got
75 English false positives. So nouns changes both halves at once: a different
list, read in a narrower place.

• CONTENT_WORDS is French nouns and adjectives with no English spelling,
  where FUNCTION_WORDS is grammar. REJECTED_CONTENT_WORDS records the
  candidates left out and why — demonstration, service and catalogue
  because they are spelled the same in both languages, and boutique because
  it is the fixture service name in some 60 paths, hostnames and slugs.
• _prose_values reads only the value side of a line (quoted spans, plus
  unquoted key: value so a French default in a plugin template counts) and
  drops anything identifier-shaped — no uppercase, no whitespace. That filter is
  what makes a threshold of one content word safe, and what keeps
  "name": "boutique", plateforme@example.net, helm/charts/boutique/ and
  fins-de-ligne.txt out of a signal that does not own renaming them.

The measurement

Run over the blobs at the ZED-60 merge base (master @ e2a81e5, unchanged on
master @ 068c1e0): 32 lines across ten test modules, and nothing anywhere
else — zero in src/forge/**/*.py, zero in every plugin template, zero in the
other nine test modules. Not one false positive.

tests/french_values_pre_zed60.txt holds those 32 lines verbatim, and
test_every_pre_zed60_french_value_fires asserts each one is exactly
{NOUNS} — which pins that the signal is new information rather than a third
reading of something accent or words already caught.

The issue says 22 lines across eleven modules; the fixture header reconciles
that against the machine's 32 across ten, rather than quietly picking one.

What it still cannot see

ZED-69 happened because an empty [baseline] was read as "no French here", so
that is now written down rather than left to be rediscovered: the module
docstring has a "What none of the three detects" section,
test_the_value_blind_spots_are_known pins it on real lines, and
french_baseline.txt points at it before the data. The known gaps are French
values that are only identifiers ("Boutique_Web" — ZED-59's scope), phrases
built only from words spelled the same in English, and a noun broken by a string
concatenation.

This PR is correct and early: it needs #57, and only #57

test_no_french_outside_the_baseline fails here on the 32 real French lines.
The D6 matrix run on ZED-74 confirmed it 3/3, and the conclusion was right —
this lands after the translations. Adding those 32 lines to [baseline] would
invert the mechanism the PR exists to build, so they are not there and
[baseline] is still empty.

Measured, not assumed: splicing #57 alone (ZED-60, the eleven inline-spec
modules) into this branch gives unknown (path, signal): [] across 267 scanned
files, with every allowlist entry still matching something.

ZED-74 sequenced this behind #55, #57, #60 and #46. Only #57 is load-bearing
for the gate — #55 (ZED-27) touches tests/golden/**, tests/specs/*.yml and
tests/fingerprints/*.json, none of which any SCANNED_GLOBS entry matches
(tests/**/*.py is the only tests/ glob), and #60 and #46 touch files the
signal already reports clean. That is a note for the sequencer, not a request to
jump the queue: #57 has not merged either, so the order costs nothing.

🤖 Generated with Claude Code