ADR-224: Implement intent stages (PhraseVariator, VocabComputer, DVC entry-points)

[jodavis-claude]

ADR-224 — Intent Stages: PhraseVariator, VocabComputer, DVC Entry-Points

Summary

Implements Task 3 of the OOP pipeline refactor (ml/_spec_OopPipeline.md): the ml/pipeline/intent/ package, ml/pipeline/stages/conventions.py, and the first two DVC entry-points.

Files Added

ml/pipeline/intent/phrase_variator.py

• PhraseVariator class with injectable rng: random.Random
• generate(base_phrases, variations_per_phrase) returns list[TextSample]
• Ports _create_variation() and sanity_check() from the legacy 01_generate_phrases.py script; all random.* calls replaced with self.rng.*
• TextSample constructed with seed=0, content_hash=sha256(content), content=surface_form, label=command

ml/pipeline/intent/vocab_computer.py

• VocabComputer, VocabResult, PhonemeProvider, PhonemeNotFoundError
• compute(manifest, output_dir) extracts words from TextSample.label (splitting on _), not content
• Writes phoneme_list.txt and words_to_phonemes.json; no phoneme_trie.json
• VocabResult.ctc_blank_idx = len(phoneme_list)

ml/pipeline/stages/conventions.py

• All seven path-convention functions: manifest_path, split_manifest_path, sample_file_path, model_path, evaluation_predictions_path, evaluation_metrics_path, test_samples_path

ml/pipeline/stages/intent_01_generate_phrases.py

• CLI: --input-phrases, --output-dir, --variations-per-phrase, --subsample-rate
• Reads CSV, constructs PhraseVariator(random.Random(42)), applies subsample filter, writes Manifest[TextSample]

ml/pipeline/stages/intent_02_compute_vocab.py

• CLI: --input-manifest-dir, --output-dir, --phoneme-dict
• Includes _CmuPhonemeProvider for CMU Pronouncing Dictionary format
• Reads manifest, constructs VocabComputer, calls compute()

ml/test/pipeline/intent/test_phrase_variator.py

• Determinism tests (same seed → same output; different seeds → different output)
• Output structure tests (labels match commands, seed=0, content_hash integrity, variant count)
• Sanity-check tests (empty, too-short, no-letters, unrecognizable base phrase, too-long all rejected)
• Snapshot determinism check against original VariationGenerator

ml/test/pipeline/intent/test_vocab_computer.py

• Phoneme list coverage; words_to_phonemes mapping; ctc_blank_idx == len(phoneme_list)
• File output correctness (phoneme_list.txt, words_to_phonemes.json)
• Uses label, not content
• TestConventions: all seven convention functions verified

Build & Test

• .NET build: clean, 0 warnings, 0 errors
• .NET unit tests: 367 passed (312 AdaptiveRemote.App.Tests + 55 AdaptiveRemote.Speech.Tests)
• E2E failures are pre-existing infrastructure issues unrelated to this change:
  • Headless E2E: MessagePack.Annotations DLL not copied after git clean (fixed in this PR by adding direct package reference to AdaptiveRemote.App.csproj)
  • WPF E2E: AdaptiveRemote.runtimeconfig.json missing after git clean (pre-existing, not introduced here)

Design Notes

• TextSample.id is content-addressable (= content_hash) per the existing sample.py implementation — the spec's mention of uuid4() is superseded by the actual dataclass
• VocabComputer splits label words on _ (e.g. TV_ON → ["TV", "ON"])
• No digit-to-word substitution; no phoneme_trie.json
• Entry-points add ml/ to sys.path for direct invocation compatibility

[github-actions]

Test Results

401 tests  ±0   401 ✅ ±0   3m 2s ⏱️ +39s
  5 suites ±0     0 💤 ±0 
  5 files   ±0     0 ❌ ±0 

Results for commit 90db3ad. ± Comparison against base commit 1dffa03.

♻️ This comment has been updated with latest results.

[jodavis]

This review is suggesting a pretty large design change, so we'll start with this. The PhraseVariator needs to use VariationGenerator in order to keep variations consistent, otherwise we will end up with large differences when parameters change.