test(cli): add subprocess contract tests for train/encode/decode

[dinesh-git17]

Summary

Closes Task 4-2 by adding 12 subprocess-level contract tests that pin the bpetite CLI to its stdout/stderr boundary, its exact JSON summary shape, and every documented failure mode.

Why

Task 4-1 shipped the CLI behind a single deterministic contract (FR-33/FR-34 channel discipline, exact train summary keys, decode raw-text semantics, typed exits), but nothing enforced it at the subprocess boundary. In-process tests collapse the stdout/stderr split, so a regression like a stray print() on stdout or a missing non-zero exit on an error path would slip through. These tests drive the real installed bpetite entry point through subprocess.run, so every assertion witnesses the same surface a user or CI smoke test would see.

Changes

• tests/test_cli.py — 12 new tests covering:
  • train happy path (JSON summary shape pinned to exact required key set; corpus_bytes tied to the fixture's on-disk size; JSON absent from stderr on all five key substrings plus the full json.dumps string, so a double-write regression is caught)
  • train progress lifecycle ("Training started: planned=" and "Training complete: merges=" — substrings that only the _train_with_progress._on_event console.print lines produce, not the panel titles, so deleting either branch fails the test)
  • train every-100-merges branch (new progress_corpus_path session fixture generates a deterministic ~15 KB synthetic corpus via fixed-seed random.Random(0xBADC0FFEE); with --vocab-size 480 the run plans 224 merges and empirically fires 2 merge events, asserted via "Training merges:" on stderr and absent from stdout)
  • train failure modes: nonexistent input, invalid UTF-8 input (reusing invalid_utf8.bin), save without --force to existing path, save with --force overwrites cleanly
  • encode happy path (compact JSON array via json.dumps(ids, separators=(",", ":")) equality, no spaces, stderr empty under non-interactive subprocess)
  • decode happy path (raw text bytes, stderr empty; roundtrip of "Hello, world!" through encode → decode)
  • decode failure modes: unknown token ID ([999999]), invalid UTF-8 byte sequence ([128], lone continuation byte), missing model file
• tests/conftest.py — added tiny_corpus_path session fixture (CLI subprocess tests need the actual file path, not the decoded string already exposed by tiny_corpus)

Validation

• uv run pytest — 192 passed (12 new, no regressions)
• uv run ruff check . — clean
• uv run ruff format --check . — clean
• uv run mypy --strict — clean

Manual checks:

• Three Codex review rounds; rounds 1 and 2 surfaced real assertion-uniqueness and branch-reachability gaps (panel title vs lifecycle line, 4-merge run never firing the every-100 branch, JSON summary never checked against stderr), all fixed in this PR. Round 3 came back clean.
• Subprocess invocation resolves the bpetite entry point via Path(sys.executable).parent / "bpetite" rather than uv run bpetite, so each test avoids nested uv resolution inside pytest and runs fast on both macOS and Linux CI targets.

Risks / Follow-ups

• The every-100-merges test adds ~0.6 s to the test suite (full pytest: 1.65 s → 2.38 s) because training 224 merges dominates subprocess startup. Acceptable for a single coverage-motivated case.
• Ruff S311 (non-crypto RNG) is suppressed inline on the synthetic corpus generator with a scoped # noqa: S311; the RNG is deterministic fixture generation, not a security surface.
• Task 4-3 (TinyShakespeare download helper) is the next Phase 4 task.

[github-actions]

bpetite workflows

Workflow	Status	Comment if failure and where
tests	success	ok
lint	success	ok
syntax	success	ok
format	success	ok
types	success	ok
build	success	ok
cli-smoke	success	ok
determinism	success	ok
policy-guard	success	ok
ci-meta	pending	waiting

PR #33: tracked workflows are still running.