test: Index every table file, not a list of extensions

[henrikottesorensen]

The one root cause behind the 27

BrailleSpecTests.TableCache indexed tables/ through an extension allow-list:

.Where(f => Path.GetExtension(f) is ".ctb" or ".utb" or ".uti" or ".dis" or ".cti" or ".dic")

Upstream 3.39.0 ships 55 .tbl files, and .tbl is not on that list. So none of them reached
lou_indexTables — and a .tbl is precisely a metadata header (#+language:, #+grade:,
#+type:) plus include lines, i.e. exactly the file lou_findTable is meant to match on.

liblouis applies no filter of its own. indexTablePath() in metadata.c walks LOUIS_TABLEPATH
with listDir() and hands lou_indexTables every non-directory entry; lou_indexTables then
drops any file analyzeTable finds no metadata in. A filter in the harness could therefore only
ever be wrong in one direction — dropping tables upstream would have indexed. It was.

That single omission is the whole of the group the specs README described as "27 fail on table
resolution, and are probably one root cause rather than 27": 20 queries matched nothing, 7 settled
for a lower-scoring table from another language or code.

Measured

A C probe linked against the built osx-arm64 liblouis.dylib 3.39.0, indexing
LibLouis.NET.Tables/tables twice — once with the filter, once with every file — then running each
excluded spec's query:

query	with the filter	every file	spec asserts
language:bn grade:1	(no match)	bn.tbl	bn.tbl
language:hr type:computer dots:8	(no match)	hr-comp8.tbl	hr-comp8.tbl
locale:pa grade:1	(no match)	pa.tbl	pa.tbl
language:es type:literary grade:1	es-no.utb	es.tbl	es.tbl
language:ar grade:1	he-IL.utb	ar.tbl	ar.tbl
language:en region:en-GB grade:2	en-ueb-g2.ctb	en_GB.tbl	en_GB.tbl
language:en region:en-US grade:2	en-ueb-g2.ctb	en_US.tbl	en_US.tbl
language:en region:en-US type:computer dots:8	en-nabcc.utb	en_US-comp8-ext.tbl	en_US-comp8-ext.tbl
language:cmn-Hans region:cmn-CN system:cmn-traditional variant:no-tone	zhcn-g1.ctb	zh_CHN.tbl	zh_CHN.tbl

Every query from the 27 resolves to exactly the table its __assert-match names. Run over all 128
distinct queries in the suite, the only resolutions that change are those 27's — nothing the 100
specs already here depend on moves.

What this restores

25 of the 27, taking the suite from 100 specs to 125, over roughly 70 languages. The
recorded skipped tally moves with them: test option: mode 14 → 19, test option: typeform
287 → 339.

Green on net8.0 and net10.0 locally (208 tests each).

The two that stay out, and why

Both now resolve their tables correctly and fail on something else. Diagnosed and written into the
specs README rather than fixed here, since they are not this root cause:

• en-us-g2.yaml (6/20) — the harness carries flags: forward into a following bare tests:
  block; upstream does not. lou_checkyaml.c declares int testmode = MODE_DEFAULT; inside the
  loop over the flags/tests keys, so a tests: block with no flags: of its own runs forward.
  The six failures are cases the harness runs backward and upstream runs forward — back-translating
  five identical cells and expecting five different bullet characters cannot pass either way.
• es-g0-g1.yaml (8/992) — BrailleSpec.Unescape handles \xNNNN, \yNNNN, \uNNNN, \\
  and \", but not the \n, \r and \s this spec uses, so they reach liblouis as a literal
  backslash plus a letter. liblouis's own parseCharsInternal also takes
  \e \f \n \r \s \t \v \w and \ZNNNNNNNN. Worth completing carefully: \s appears across the
  specs, so it changes the input of cases that pass today — spaces.yaml is the obvious one to
  re-check.

Also recorded in the README: the 7 specs the old list did not name (ar-ar-g1, ar-ar-g1_harness,
en-GB-g2, en-us-comp8-ext-back_harness, en-us-comp8-ext-for_harness, en-us-g2, zh-chn),
and the fact that CopyToOutputDirectory never deletes — so a stale bin/ keeps tables upstream
has since removed, and with the filter gone those get indexed too.

🤖 Generated with Claude Code