fix: reject directory args in keywords CLI

[cardmagic]

Fast follow to #163. Closes the one review item that did not land before the merge, and stops the Hound noise.

1. Directory arguments leaked a raw errno

command_fit globbed arguments without a file check, and command_extract only checked File.exist?. A directory reached File#read and surfaced the errno through the generic StandardError rescue:

$ keywords fit -m d.json corpus
Error: Is a directory @ io_fillbuf - fd:6 /path/corpus     # exit 1

$ keywords extract corpus
Error: Is a directory @ io_fread - /path/corpus            # exit 1

Every other input error in the tool exits 2 with a clear message, so this path was the odd one out.

It also broke the workflow the README leads with. A shell expands keywords fit corpus/* before the CLI sees it, so a subdirectory arrives as its own argument and the whole run fails:

$ keywords fit corpus/*        # corpus/ holds a.txt, b.txt, subdir/
Error: Is a directory @ io_fillbuf - fd:6 /path/corpus/subdir    # exit 1

Behavior now

Command	Before	After
keywords fit corpus/* (subdir present)	errno, exit 1	fits the files, exit 0
keywords fit corpus	errno, exit 1	Error: No files to fit, exit 2
keywords extract corpus	errno, exit 1	Error: "..." is a directory, not a file, exit 2
keywords fit corpus/a.txt corpus/NOPE.txt	No files matched, exit 2	unchanged
keywords fit 'corpus/*.md'	No files matched, exit 2	unchanged

fit skips directories and keeps the files around them. A path that matches nothing still fails, so typos are still caught. An argument set that yields no file at all fails with No files to fit.

This also removes a split introduced by the per-argument guard in #163: a quoted 'corpus/*' and an unquoted corpus/* behaved differently, because the CLI expands the first itself. Both now do the same thing.

2. Hound reported 96 offenses that CI does not see

Hound flags Metrics/LineLength. RuboCop renamed that cop to Layout/LineLength in 0.78 (December 2019). Hound also flags at stock defaults (LineLength 80, MethodLength 10, ClassLength 100, AbcSize 15) while .rubocop.yml sets 140, 25, 250, and 30 and excludes this file. It flags Style/Documentation and Style/FrozenStringLiteralComment too, which .rubocop.yml disables outright.

So Hound never read .rubocop.yml. This adds .hound.yml to point it there.

Note that Hound may still not honor it. Its RuboCop is too old to parse plugins: (needs 1.72+) or NewCops (needs 0.90+), so it can fall back to defaults again. If the noise continues, the next step is to disable the linter with ruby: {enabled: false} or remove the Hound app, since the CI lint job already runs RuboCop 1.89 against .rubocop.yml on every PR and is the real gate.

Verification

Tests written first, confirmed failing with exit 1, then fixed.

• 709 runs, 0 failures, 1 skip
• bundle exec rubocop: 56 files, no offenses
• rbs-inline + rbs validate: pass
• Built the gem, installed it into an isolated GEM_HOME, and drove the real keywords binary through the table above plus 13 regression checks from the feat: add keywords CLI tool for text vectorization (#122) #163 review rounds: bigram labels, min_word_length models, snake_case tokens, missing-model guards on transform and info, empty stdin, negative -n, --min-df/--max-df/--ngram validation, and the 600-file run under ulimit -n 256. All hold.

Four tests added: glob with a subdirectory, bare directory to fit, directory to extract, and a nonexistent path among valid ones as a regression guard.

[houndci-bot]

Some files could not be reviewed due to errors:

.rubocop.yml: Layout/LineLength has the wrong namespace - should be Metrics

.rubocop.yml: Layout/LineLength has the wrong namespace - should be Metrics
Warning: unrecognized cop plugins found in .rubocop.yml
Warning: unrecognized cop Naming/MethodParameterName found in .rubocop.yml
Warning: unrecognized cop Lint/UselessConstantScoping found in .rubocop.yml
Warning: unrecognized cop Minitest/MultipleAssertions found in .rubocop.yml
Warning: unrecognized cop Style/OneClassPerFile found in .rubocop.yml
Error: Unknown Ruby version 3.1 found in `TargetRubyVersion` parameter (in .rubocop.yml).
Supported versions: 2.1, 2.2, 2.3, 2.4, 2.5

[greptile-apps]

Greptile Summary

This PR normalizes directory handling in the keywords CLI and configures Hound to use the repository’s RuboCop settings.

• Filters directories from fit inputs while preserving unmatched-path errors and rejecting an empty resulting file set.
• Rejects directory arguments to extract with a usage error.
• Adds regression coverage for directory and missing-path behavior.
• Adds .hound.yml pointing Hound at .rubocop.yml.

Confidence Score: 5/5

The PR appears safe to merge with no actionable regressions identified.

The changed CLI paths correctly exclude directories from fitting, reject directory extraction with the intended usage error, preserve unmatched-path failures, and include focused regression coverage.

Important Files Changed

Filename	Overview
lib/classifier/keywords/cli.rb	Adds regular-file filtering for fit inputs and explicit directory rejection for extract without introducing an actionable defect.
test/keywords/cli_test.rb	Covers mixed file/directory globs, directory-only fit input, missing fit paths, and directory extract input.
.hound.yml	Enables Hound’s Ruby linter and directs it to the existing RuboCop configuration.

_{Reviews (1): Last reviewed commit: "fix: reject directory args in keywords C..." | Re-trigger Greptile}