@@ -100,9 +100,9 @@ public void SkippedEntriesMatchTheRecordedTally()
         var expected = new SortedDictionary<string, int>(StringComparer.Ordinal)
         {
             ["test option: expected_typeform"] = 17,
-            ["test option: mode"] = 14,
+            ["test option: mode"] = 19,
             ["test option: outputPos"] = 5067,
-            ["test option: typeform"] = 287,
+            ["test option: typeform"] = 339,
             ["testmode: display"] = 5,
             ["testmode: hyphenate"] = 1198,
         };
@@ -198,9 +198,15 @@ private static string DescribeSkipped(IReadOnlyDictionary<string, int> skipped)
     // the hundred-odd specs do not each re-run the same lou_findTable queries.
     private static readonly Lazy<Dictionary<string, string>> TableCache = new(() =>
     {
-        LibLouis.Instance.IndexTables(
-            Directory.EnumerateFiles(TableDirectory)
-                .Where(f => Path.GetExtension(f) is ".ctb" or ".utb" or ".uti" or ".dis" or ".cti" or ".dic"));
+        // Every file in the directory, not a list of extensions. liblouis itself indexes whatever
+        // LOUIS_TABLEPATH contains without looking at the name (metadata.c, indexTablePath ->
+        // listDir), and lou_indexTables already skips a file it finds no metadata in, so a filter
+        // here can only ever be wrong in one direction: it drops tables upstream would have
+        // indexed. That is what it did - the extension list predated the .tbl files, so all 55 of
+        // them were invisible, and .tbl is exactly where upstream now puts the metadata that
+        // lou_findTable matches on. Queries that should have selected one either found nothing or
+        // settled for a worse match from another language.
+        LibLouis.Instance.IndexTables(Directory.EnumerateFiles(TableDirectory));
 
         var resolved = new Dictionary<string, string>(StringComparer.Ordinal);
 
 
@@ -3,7 +3,7 @@
 Copied verbatim from `upstream/liblouis-<version>/tests/braille-specs/`. Re-copy when the upstream
 version is bumped; the diff is the set of expectations that changed.
 
-100 specs are here, from liblouis 3.39.0, covering roughly 60 languages. Upstream ships 160.
+125 specs are here, from liblouis 3.39.0, covering roughly 70 languages. Upstream ships 160.
 
 The selection below was made against 3.33.0. Specs upstream has added since have not been tried yet:
 `en-g3-dictionary.yaml`, `en-g3.yaml`, `en-gb-comp8.yaml`, `et_harness.yaml`, `it.yaml`,
@@ -17,26 +17,35 @@ The selection below was made against 3.33.0. Specs upstream has added since have
 200,000+ cases each, around 800,000 in total. They would dominate every run for little extra signal.
 Add them behind a switch if wanted.
 
-### 39 specs that do not yet pass
+### 14 specs that do not yet pass
 
 Not excluded because they are wrong — excluded because nobody has worked out yet whether the
 disagreement is this harness or the wrapper, and a red suite that is expected to be red is worse
 than a smaller green one. They divide into three groups.
 
-**27 fail on table resolution, and are probably one root cause rather than 27.** 20 report no table
-matching a query, 7 resolve to a table other than the one the spec's `__assert-match` names. The
-queries themselves look well formed — `language:bn grade:1` — so the likely cause is which tables
-reach `lou_indexTables`, or which of them liblouis manages to analyse. Worth starting here: it is
-the largest group and the cheapest to be wrong about.
+**2 fail on a construct this harness reads differently from upstream.** Both resolve their tables
+correctly and disagree on a handful of cases each; in both the harness is the one that is wrong.
 
 ```
-bn.yaml cs_harness.yaml eo-g1_harness.yaml es-g0-g1.yaml ethio-g1_harness.yaml hi_harness.yaml 
-hr-8dots_harness.yaml hu-hu-g1_braille_input_backward.yaml hu-hu-g1_braille_input_forward.yaml 
-hu-hu-g1_dictionary_numbers.yaml hu-hu-g1_harness.yaml kmr.yaml litdigits6Dots_backward.yaml 
-lv_harness.yaml ml.yaml nl-g0_harness.yaml pa.yaml 
-pl-g1.yaml sl-g1.yaml tr.yaml
+en-us-g2.yaml  6/20
+es-g0-g1.yaml  8/992
 ```
 
+`en-us-g2.yaml` has a `flags: {testmode: bothDirections}` block followed by three bare `tests:`
+blocks. This harness carries the flags forward; 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 cases that fail here are ones the harness runs
+backward and upstream runs forward — back-translating five identical cells and expecting five
+different bullet characters cannot pass in any case.
+
+`es-g0-g1.yaml` uses `'\n'`, `'\r'` and `'\s'`. `BrailleSpec.Unescape` handles only `\xNNNN`,
+`\yNNNN`, `\uNNNN`, `\\` and `\"`, so those three reach liblouis as a literal backslash plus a
+letter and translate to two cells. liblouis's own parser (`parseCharsInternal` in
+compileTranslationTable.c, reached from `lou_checkyaml` via `_lou_extParseChars`) also takes
+`\e \f \n \r \s \t \v \w` and the 8-digit `\ZNNNNNNNN`. Worth fixing with care: `\s` in particular
+is used across the specs, so completing the list changes the input of cases that pass today —
+`spaces.yaml`, in the next group, is the obvious one to re-check afterwards.
+
 **10 disagree on translation output**, a fraction of their cases each. These are the interesting
 ones: either a per-case option this harness does not model, or a real difference between the wrapper
 and upstream.
@@ -59,8 +68,24 @@ zh-tw.yaml  2/20
 unindented line (`- [` then `disingenuous, ...]`), which libyaml accepts and YamlDotNet 18.1.0
 throws on.
 
+### What used to be here: the 27 table-resolution failures
+
+Gone as of the fix in `BrailleSpecTests.TableCache`, recorded because the shape of the mistake is
+worth not repeating. The harness indexed `tables/` through an extension allow-list that predated
+upstream's `.tbl` files, so all 55 of them were invisible to `lou_indexTables` — and `.tbl` is
+exactly where upstream keeps the `#+language:` / `#+grade:` / `#+type:` metadata `lou_findTable`
+matches on. 20 queries therefore matched nothing and 7 settled for a worse match from another
+language: `language:ar grade:1` resolved to `he-IL.utb`, `language:en region:en-US grade:2` to
+`en-ueb-g2.ctb`. liblouis applies no filter of its own — `indexTablePath` in metadata.c hands it
+every file on `LOUIS_TABLEPATH` and drops the ones with no metadata — so a filter here could only
+ever be wrong in one direction. 25 of the 27 passed as soon as it was removed.
+
 ## Restoring one
 
 Copy it back from the upstream tarball and run the suite. Note that liblouis caches compiled tables
 process-wide, so when bisecting a disagreement, use a cold process per variant — otherwise the
 result depends on which test compiled a given table first.
+
+`CopyToOutputDirectory` never deletes, so a `bin/` from an older liblouis keeps tables upstream has
+since removed, and those get indexed alongside the current set. If a query resolves to a table that
+is not in `LibLouis.NET.Tables/tables`, delete `bin/` and build again.