. "$BST_LIB"
# BST-TIMEOUT: 2400  # TWO python3 starts, and the probe's corpus is the only
#                    # thing here with real work in it.
#                    #
#                    # The two: a bare `import snowballstemmer', whose STDERR is
#                    # its own assertion, and the probe -- ONE start for 72
#                    # measurements, among them 41417 stemmed words checked
#                    # against Snowball's published output. Everything else
#                    # reads the filesystem or dpkg with shell builtins, or
#                    # greps one probe transcript. This package ships NO
#                    # executable, so there is no program to run as a real
#                    # process and no exit status to assert: nothing needs a
#                    # third start.
#                    #
#                    # THE PROBE'S WORK IS ~55000 STEM CALLS, measured at 1.35s
#                    # on the x86-64 control. The bound is not sized from that
#                    # figure but from the interpreter start, which is tens of
#                    # seconds on this arm before the first line of the program
#                    # runs, and from not knowing this arm's slowdown factor on
#                    # a pure-Python inner loop -- which is what a stemmer is,
#                    # and the one thing here that is NOT dominated by the
#                    # start. 2400 leaves the corpus a factor of ~1500 over the
#                    # control before it can time out. Nothing in this file is
#                    # quadratic in anything: the corpus is one pass per
#                    # language plus two more over the first 5000 English words.
#                    #
#                    # NOT MEASURED ON TARGET YET: this port's first on-target
#                    # run is the batch verify. The measured elapsed goes in the
#                    # test plan rather than in this comment, which would cost an
#                    # image build to keep the shipped suite byte-identical.
#
# python3-snowballstemmer 2.2.0 (poky/meta
# recipes-devtools/python/python3-snowballstemmer_2.2.0.bb, `inherit pypi
# setuptools3', PyPI sdist) on the wasm32-hwjs glibc image.
#
# ============================================================================
# WHAT THE PACKAGE IS.
# ============================================================================
#
# The Snowball stemming algorithms compiled to Python: 29 stemmers for 28
# languages, each one reducing an inflected word to its stem so that a search
# for `connection' also finds `connected'. `english' is Porter2, the algorithm
# Sphinx builds its search index with; `porter' is Dr Porter's ORIGINAL 1980
# algorithm, kept for reproducing old results and NOT a synonym for it. The
# generated stemmer modules are the whole package -- there is no compiler here,
# only its output.
#
# ============================================================================
# THE PURE-PYTHON PATH IS THE ONE UNDER TEST, AND IT IS CHOSEN AT IMPORT TIME.
# ============================================================================
#
# `__init__.py' opens with `try: import Stemmer' -- the PyStemmer C extension --
# and sets `cext_available' from whether that worked. When it works,
# `algorithms()' and `stemmer()' BOTH DELEGATE to the C library, and every
# corpus case below would be measuring PyStemmer rather than the Python this
# package ships.
#
# On this image python3 is one static binary with an empty lib-dynload and no
# dynamic loader (BRINTOS-1476), so PyStemmer could not load even if it were
# installed. But "could not" is a premise, and a premise is not a result: the
# suite asserts `cext-available' is `no' and that the import raises ImportError
# specifically. Without that pair, a green corpus would be evidence about an
# unknown implementation.
#
# The `ImportError' half is the sharper of the two. `__init__.py' catches ONLY
# ImportError, so an import that failed any other way -- which is exactly the
# shape BRINTOS-2046 describes, a CPython extension refused at `dlopen' on this
# image -- would propagate straight out of `import snowballstemmer' and take the
# whole package down. That the failure is the ordinary "no such module" one is
# a fact about this image worth holding.
#
# ============================================================================
# THE CORPUS IS SNOWBALL'S OWN PUBLISHED TEST DATA, PINNED TO THE RELEASE.
# ============================================================================
#
# `python3-snowballstemmer-corpus.txt' carries `word<TAB>expected-stem' pairs cut
# from snowball-data, the project's own published vocabulary/output pairs, at
# commit 0703f1d (2021-09-07) -- THE LAST COMMIT BEFORE snowball v2.2.0 WAS
# TAGGED (2021-11-10), which is the release this package is generated from.
#
# That pin is the load-bearing part, and it was established by measurement
# rather than assumed. On the x86-64 control, snowball-data at HEAD disagrees
# with snowballstemmer 2.2.0 on 57 English words, 838 French and 1265 German:
# the data tracks the algorithms, and the algorithms moved after 2.2.0. A
# fixture cut from HEAD would have produced a suite that was red on this image
# for reasons having nothing to do with this image -- the worst kind of red,
# because it looks exactly like a port defect.
#
# At 0703f1d, every one of the 28 corpora this package has an algorithm for
# agrees with it EXACTLY -- 0 mismatches over 1.6 million words, checked on the
# control before the fixture was cut. So a red here is a statement about this
# image and nothing else. That is what justifies carrying 759 KB of fixture.
#
# English is carried WHOLE, 29417 pairs, because Porter2 is the algorithm the
# port is really about. The other ten are STRIDED samples of 1200 -- strided and
# not head, so a sample spans the whole alphabet instead of the words beginning
# with `a', which is the difference between testing a stemmer and testing its
# first few hundred prefixes. The ten are chosen for the scripts and rules they
# exercise rather than for coverage of the roster: `porter' (the original, to
# hold it apart from english), french and german and spanish (Latin-1 accents
# and umlauts), russian and greek and armenian (three non-Latin scripts),
# turkish (the dotted/dotless `i', the classic case-folding trap), hindi
# (Devanagari), irish (its own initial mutations).
#
# WHAT A DIGEST WOULD NOT HAVE TOLD US. The probe reports a SHA-256 over every
# stem it produced, which is the assertion that catches a single wrong word
# anywhere in the corpus -- but it also reports `corpus_first_bad', the first
# disagreeing word with what was wanted and what was got, and a per-language
# count. A digest alone says the corpus moved without saying which word moved
# it, and the difference between those two is whether the next reader has to
# reproduce a 41417-word run to learn anything.
#
# ============================================================================
# THE ROSTER IS ASSERTED AS A COUNT, AN ORDER AND A DIGEST.
# ============================================================================
#
# `algorithms()' is `list(_languages.keys())' over a dict literal in
# `__init__.py'. The failure this port is exposed to is a packaging step that
# dropped stemmer modules: `stemmer("english")' would go on working while a
# dozen languages had quietly gone, and any hand-written sample of names would
# pass straight through that. So the roster is reduced to a digest over its
# sorted names, with the count beside it to say which way a mismatch moved, and
# each of the 32 shipped `.py' files asserted present by name as well.
#
# Three further cases say the roster is REAL rather than 29 names resolving to
# one class: every name constructs, the 29 objects have 29 distinct classes, and
# each class comes from its OWN module (`snowballstemmer.<name>_stemmer').
#
# ============================================================================
# TWO UPSTREAM QUIRKS ARE PINNED HERE RATHER THAN REPORTED ANYWHERE.
# ============================================================================
#
# `__all__' IS WRONG. It is `('language', 'stemmer')', and the module defines
# `algorithms' and `stemmer'. There is no `language' -- that is PyStemmer's name
# for the same thing -- so `from snowballstemmer import *' raises AttributeError.
# Asserted as AttributeError, so a future upstream fix shows up as a red that
# says what changed rather than as a silent behaviour change.
#
# STEMMING IS NOT IDEMPOTENT, and the suite pins the NUMBER rather than claiming
# the property. Re-stemming the stems of the first 5000 English words changes
# 165 of them: Porter2 is not defined to be a fixed point, and a caller who
# assumed it was would be wrong. A pinned count is a regression signal; a
# pinned property would have been a false claim.
#
# Neither is reported upstream. We do not file against third-party public
# repositories -- see docs/conventions/no-public-upstream-reports.md -- and for
# a behaviour we merely depend on, a pinned assertion is the durable record.
#
# ============================================================================
# THE ISO CODES ARE NOT ACCEPTED HERE, AND THAT IS A PORTABILITY TRAP.
# ============================================================================
#
# `modules.txt' lists `english,en,eng' and the like, and PyStemmer accepts all
# three. `stemmer()' on the pure-Python path lowercases its argument and looks
# it up in `_languages', WHICH IS KEYED BY FULL LANGUAGE NAMES ONLY -- so
# `stemmer("en")' raises KeyError. A caller who wrote that because the C
# extension took it gets a KeyError and nothing else. Both halves are measured:
# `English' and `EnGlIsH' resolve (the lowercasing works), `en', `eng' and `'
# raise, and the KeyError's message is asserted too, because it is the only
# thing that names WHICH algorithm was refused.
#
# ============================================================================
# THE STEMMER OBJECT IS REUSED ACROSS WORDS, WHICH IS A STATE HAZARD.
# ============================================================================
#
# A stemmer keeps its cursor state on the instance (`current', `cursor', `bra',
# `ket'), and `stemWords()' drives ONE instance over the whole list. The failure
# shape is a word leaving state behind that changes the next word's answer --
# the same defect python3-html2text's `handle()' has. It is checked three ways:
# the same word stemmed twice in a row, a long word stemmed between two short
# ones, and `stemWords(list)' asserted equal to a list comprehension over
# `stemWord()' across 5000 words, which is the same instance driven in a
# different order.
#
# ============================================================================
# THE PACKAGE'S SHAPE, AND ITS ONE DEPENDENCY.
# ============================================================================
#
# `Depends: python3-core' and NOTHING ELSE -- asserted as the whole line, not as
# a substring, because the point is the absence. The recipe declares no
# RDEPENDS at all and the scanner found nothing to add: the library imports
# nothing outside itself. Nothing in this round or outside it gated this port.
#
# The wheel says `Root-Is-Purelib: true' and the .deb ships 32 `.py', their 32
# `.pyc' and a five-file dist-info -- 69 files, no shared object, no script in
# /usr/bin. All of that is asserted, because "pure Python" is the premise the
# whole port rests on and BRINTOS-2046 does not reach a package that has no
# compiled extension to be refused.
#
# ============================================================================

P=python3-snowballstemmer

# The canary: this line proves the suite was entered and its first case
# reached, which is what tells a truncated run apart from an absent one.
bst_ok $P/suite-canary

require python3 $P/python3-present

SITE=/usr/lib/python3.12/site-packages
PKGDIR=$SITE/snowballstemmer
DI=$SITE/snowballstemmer-2.2.0.dist-info
PKGLIST=/var/lib/dpkg/info/python3-snowballstemmer.list
PROBE=$BST_DATA/python3-snowballstemmer-probe.py
CORPUS=$BST_DATA/python3-snowballstemmer-corpus.txt
tmp=${TMPDIR:-/tmp}
W=$tmp/bst-snowballstemmer.$$
mkdir -p "$W" || { bst_notok $P/workdir "cannot create $W"; exit 0; }

# ---------------------------------------------------------------------------
# Where it landed.
# ---------------------------------------------------------------------------
check $P/site-packages-dir test -d "$SITE"
check $P/package-dir       test -d "$PKGDIR"
check $P/package-init      test -s "$PKGDIR/__init__.py"
# setup.py maps package_dir snowballstemmer -> src/snowballstemmer, so a
# flat module beside the package would mean the mapping was ignored.
check $P/no-flat-module    test ! -f "$SITE/snowballstemmer.py"

# The 32 shipped modules by name: the 29 stemmers, plus the two the generated
# code is built on. Named individually rather than counted, so a red says WHICH
# language went missing -- which is the whole question when a roster shrinks.
for _m in arabic armenian basque catalan danish dutch english finnish french \
          german greek hindi hungarian indonesian irish italian lithuanian \
          nepali norwegian porter portuguese romanian russian serbian spanish \
          swedish tamil turkish yiddish; do
  check $P/stemmer-$_m test -s "$PKGDIR/${_m}_stemmer.py"
done
check $P/module-among       test -s "$PKGDIR/among.py"
check $P/module-basestemmer test -s "$PKGDIR/basestemmer.py"

# The .pyc are built at packaging time, not at first import: this rootfs is
# read-only to the guest, so a missing .pyc would be recompiled on every start
# of every program that imports the package -- a cost, not a failure, and one
# that would otherwise be invisible.
check $P/pycache-init    test -s "$PKGDIR/__pycache__/__init__.cpython-312.pyc"
check $P/pycache-english test -s "$PKGDIR/__pycache__/english_stemmer.cpython-312.pyc"
check $P/pycache-base    test -s "$PKGDIR/__pycache__/basestemmer.cpython-312.pyc"

# ---------------------------------------------------------------------------
# What the .deb actually contains.
#
# The file list is copied under $TMPDIR first: the `while ... done < file' below
# must run in THIS shell to keep the variable it sets, and a stdin redirect the
# top-level shell performs on the virtiofs root makes dup2(2) return 513, at
# which point ash prints `Unknown error 513' and EXITS -- ending the suite with
# every case so far counted as a pass (brintos/linux#73).
# ---------------------------------------------------------------------------
cat "$PKGLIST" > "$W/pkgf" 2>/dev/null || : > "$W/pkgf"
check $P/pkg-list-present test -s "$W/pkgf"

# Pure Python is the premise of this port: no shared object anywhere in the
# package, so there is nothing for BRINTOS-2046 (an extension refused at
# dlopen on this image) to reach.
check_eq $P/pkg-no-so "" "$(grep -E '\.so($|\.)' "$W/pkgf" 2>/dev/null)"

# And no executable: the package is a library only, which is why this suite
# runs no shipped program and asserts no exit status.
check_eq $P/pkg-no-bin "" "$(grep -E '^/usr/s?bin/' "$W/pkgf" 2>/dev/null)"

# Nothing but .py, .pyc and dist-info shipped. Catches a stray data file or a
# packaging step that swept in something from the source tree.
_bad=""
while IFS= read -r _f; do
  [ -f "$_f" ] || continue
  case "$_f" in
    *.py|*.pyc)      continue ;;
    */__pycache__/*) continue ;;
    */*.dist-info/*) continue ;;
  esac
  _bad="$_bad $_f"
done < "$W/pkgf"
check_eq $P/pkg-only-python "" "$_bad"

# 69 files: 32 .py, their 32 .pyc, and the five-file dist-info.
check_eq $P/pkg-file-count "69" \
  "$(grep -c '\.py$\|\.pyc$\|dist-info/' "$W/pkgf" 2>/dev/null)"

# ---------------------------------------------------------------------------
# dist-info.
# ---------------------------------------------------------------------------
check $P/dist-info-dir      test -d "$DI"
check $P/dist-info-metadata test -f "$DI/METADATA"
check_out $P/dist-info-version "Version: 2.2.0" cat "$DI/METADATA"
check_out $P/dist-info-top-level "snowballstemmer" cat "$DI/top_level.txt"

# `Root-Is-Purelib: true' is the wheel saying in its own words what every case
# above infers from the file list.
check_out $P/dist-info-purelib "Root-Is-Purelib: true" cat "$DI/WHEEL"
# py2 AND py3 tags: the package is version-agnostic, which is why nothing here
# depends on the interpreter being 3.12 beyond the .pyc paths.
check_out $P/dist-info-tag-py3 "Tag: py3-none-any" cat "$DI/WHEEL"

# The licence ships INSIDE the dist-info here (as COPYING), which is not what
# every pypi package does -- python3-webencodings ships none at all. Asserted
# because a BSD-3-Clause package that lost its copyright file is a licensing
# fact, not a cosmetic one.
check $P/dist-info-copying test -s "$DI/COPYING"
check_out $P/dist-info-copyright "Dr Martin Porter" cat "$DI/COPYING"

# METADATA's own summary carries the two counts, generated by setup.py from
# modules.txt: 29 stemmers, 28 languages. They are the package's own statement
# of its roster size, and the probe's `algorithms_count' is asserted against
# the same 29 independently.
check_out $P/dist-info-summary "29 stemmers for 28 languages" cat "$DI/METADATA"

# ---------------------------------------------------------------------------
# dpkg's view.
# ---------------------------------------------------------------------------
sed -n '/^Package: python3-snowballstemmer$/,/^$/p' /var/lib/dpkg/status > "$W/st" 2>/dev/null
check $P/pkg-known test -s "$W/st"
check_out $P/pkg-installed "Status: install ok installed" cat "$W/st"
check_out $P/pkg-version "Version: 2.2.0-r0" cat "$W/st"
check_out $P/pkg-arch "Architecture: wasm32" cat "$W/st"

# THE WHOLE Depends LINE, not a substring. The point of this case is the
# ABSENCE: the recipe declares no RDEPENDS, the scanner found nothing to add,
# and `python3-core' is what the build class supplies. A substring match would
# pass just as happily with three more names on the line, and those three names
# are exactly what would mean this port had acquired a gate it did not have.
check_eq $P/pkg-depends-exact "Depends: python3-core" \
  "$(grep '^Depends:' "$W/st" 2>/dev/null)"

# ---------------------------------------------------------------------------
# It imports, and it imports SILENTLY.
#
# The second case is not redundant: a DeprecationWarning or a failed lazy
# import that upstream swallowed would leave the first case green while every
# program using the package printed noise to stderr on every run.
# ---------------------------------------------------------------------------
check $P/import python3 -c 'import snowballstemmer'
check_eq $P/import-quiet "" "$(python3 -c 'import snowballstemmer' 2>&1)"

# ---------------------------------------------------------------------------
# The probe.
# ---------------------------------------------------------------------------
check $P/corpus-present test -s "$CORPUS"

if [ ! -f "$PROBE" ]; then
  bst_notok $P/probe-present "$PROBE did not ship"
  rm -rf "$W"
  exit 0
fi
bst_ok $P/probe-present

# 2100s, inside the file's own 2400s bound: the probe must be killed and
# reported by bst_probe -- which names the last line it printed, and so the
# language it stopped in -- rather than by the runner killing the whole suite,
# which would report nothing.
bst_probe $P/probe-run "$W/probe" 2100 python3 "$PROBE" "$BST_DATA"

OUT="$W/probe"
pv() { grep "^$1=" "$OUT" 2>/dev/null | head -n 1 | cut -d= -f2-; }

if [ -s "$OUT" ]; then
  bst_ok $P/probe-output
else
  bst_notok $P/probe-output "the probe emitted nothing -- python3 did not start, or died before its first line; every CASE below is unscorable"
fi

# `probe_done' is the LAST line the probe writes, so it is the only thing that
# distinguishes a complete transcript from one cut off mid-corpus -- which would
# otherwise read as a run whose later cases simply came back empty.
check_eq $P/probe-complete "yes" "$(pv probe_done)"
check_eq $P/emitted-count  "72"  "$(pv emitted_count)"

# --- identity -------------------------------------------------------------
check_eq $P/module-answered "$PKGDIR/__init__.py" "$(pv module_file)"
check_eq $P/is-package      "yes"    "$(pv is_package)"
check_eq $P/import-ok       "ok"     "$(pv import)"
check_eq $P/dist-version    "2.2.0"  "$(pv dist_version)"
# The package carries no `__version__'; the dist-info is its only version
# statement. Pinned so a release that adds one is noticed.
check_eq $P/no-dunder-version "no"   "$(pv has_dunder_version)"

# --- the pure-Python path -------------------------------------------------
# Together these two say the corpus below measured THIS package's Python and
# not a C extension that had shadowed it.
check_eq $P/cext-unavailable "no"          "$(pv cext_available)"
check_eq $P/cext-import-error "ImportError" "$(pv stemmer_ext_import)"

# --- __all__ names a function that does not exist -------------------------
check_eq $P/all-tuple        "language,stemmer" "$(pv all_tuple)"
check_eq $P/no-language-attr "no"               "$(pv has_language_attr)"
check_eq $P/has-algorithms   "yes"              "$(pv has_algorithms_attr)"
check_eq $P/star-import-raises "AttributeError" "$(pv star_import)"

# --- the roster -----------------------------------------------------------
check_eq $P/algorithms-count  "29"  "$(pv algorithms_count)"
check_eq $P/algorithms-sorted "yes" "$(pv algorithms_sorted)"
check_eq $P/algorithms-digest \
  "da9c8974aefe4bb54c2163e815dea315a6bfc16ad5a3274f039b6cfa30b03356" \
  "$(pv algorithms_digest)"
# A fresh list per call: a caller that mutates the result must not corrupt the
# roster for everyone else in the process.
check_eq $P/algorithms-fresh-list "yes" "$(pv algorithms_fresh_list)"

check_eq $P/constructed-count   "29"   "$(pv constructed_count)"
check_eq $P/construct-failures  "none" "$(pv construct_failures)"
# 29 names, 29 distinct classes, 29 distinct modules, each class from its own
# `<name>_stemmer' module. Without these, 29 names all resolving to
# EnglishStemmer would pass every count above.
check_eq $P/distinct-classes    "29"   "$(pv distinct_classes)"
check_eq $P/distinct-modules    "29"   "$(pv distinct_modules)"
check_eq $P/class-module-match  "none" "$(pv class_module_mismatch)"
check_eq $P/api-complete        "none" "$(pv api_missing)"

# --- name resolution ------------------------------------------------------
check_eq $P/uppercase-name  "EnglishStemmer" "$(pv uppercase_ok)"
check_eq $P/mixedcase-name  "EnglishStemmer" "$(pv mixedcase_ok)"
check_eq $P/reject-unknown  "KeyError" "$(pv reject_unknown)"
check_eq $P/unknown-message "Stemming algorithm 'klingon' not found" \
  "$(pv unknown_message)"
# The ISO codes are a PyStemmer feature and raise here. Asserted so the trap is
# a recorded fact rather than something a caller discovers in production.
check_eq $P/reject-iso2  "KeyError" "$(pv reject_iso2)"
check_eq $P/reject-iso3  "KeyError" "$(pv reject_iso3)"
check_eq $P/reject-empty "KeyError" "$(pv reject_empty)"

# --- the two calls, and instance reuse ------------------------------------
check_eq $P/stemword-type      "str"   "$(pv stemword_type)"
check_eq $P/stemwords-type     "list"  "$(pv stemwords_type)"
check_eq $P/stemword-running   "run"   "$(pv stemword_running)"
check_eq $P/repeat-stable      "yes"   "$(pv repeat_stable)"
# The same short word after a long one: cursor state left behind by
# `generalization' would show up here and nowhere else.
check_eq $P/after-long-word    "run"   "$(pv after_long_word)"
check_eq $P/empty-word         "[]"    "$(pv empty_word)"
check_eq $P/empty-list         "[]"    "$(pv empty_list)"
check_eq $P/stemwords-equals-map "yes" "$(pv stemwords_equals_map)"
check_eq $P/stemwords-map-n      "5000" "$(pv stemwords_map_n)"
# NOT a claim that stemming is idempotent -- it is not. The pinned count is a
# regression signal; the property would have been a false statement.
check_eq $P/idempotent-changed   "165"  "$(pv idempotent_changed)"

# --- the interpreter's encodings ------------------------------------------
# The corpus is Devanagari, Greek, Cyrillic and Armenian as well as Latin, and
# the probe reads it with an explicit encoding="utf-8" rather than the locale's.
#
# ONLY `default_encoding' IS ASSERTED, and deliberately. sys.getdefaultencoding()
# is a Python 3 invariant -- it is `utf-8' whatever the locale -- so a red there
# is a statement about the interpreter and cannot be a false alarm.
#
# `filesystem_encoding' and `preferred_encoding' are EMITTED AND NOT SCORED.
# They are the first two values to read when the corpus goes red, which is why
# the probe reports them; but they are properties of this image's locale rather
# than of this package, the probe does not depend on either (it passes
# encoding="utf-8" explicitly), and asserting them would risk a red that says
# nothing about the port. `corpus-read' below is the case that actually covers
# the reading.
check_eq $P/default-encoding "utf-8" "$(pv default_encoding)"

# --- the corpus -----------------------------------------------------------
check_eq $P/corpus-read  "ok" "$(pv corpus_read)"
check_eq $P/corpus-langs "11" "$(pv corpus_langs)"
check_eq $P/corpus-order \
  "english,porter,french,german,spanish,russian,greek,turkish,hindi,armenian,irish" \
  "$(pv corpus_order)"

# Per language: the count fixes that the fixture arrived whole, the bad count
# fixes that it agreed. Both, because a corpus that failed to parse would
# report zero mismatches over zero words and read as green.
for _l in english porter french german spanish russian greek turkish hindi \
          armenian irish; do
  check_eq $P/corpus-$_l-bad "0" "$(pv corpus_${_l}_bad)"
done
check_eq $P/corpus-english-n  "29417" "$(pv corpus_english_n)"
check_eq $P/corpus-porter-n   "1200"  "$(pv corpus_porter_n)"
check_eq $P/corpus-armenian-n "1200"  "$(pv corpus_armenian_n)"
check_eq $P/corpus-total-n    "41417" "$(pv corpus_total_n)"
check_eq $P/corpus-total-bad  "0"     "$(pv corpus_total_bad)"
# Names the first disagreeing word rather than leaving a digest to say only
# that something moved.
check_eq $P/corpus-first-bad  "none"  "$(pv corpus_first_bad)"
# The digest over every stem produced: catches a single wrong word anywhere in
# the 41417, including one the per-language counts would have to be read
# together to find.
check_eq $P/corpus-digest \
  "8d71cb354958134a9c63770b787ed3b2a8864430f89d7ad8d13e645e2705cc44" \
  "$(pv corpus_digest)"

rm -rf "$W"