packaging: make the installed wheel importable and the CLI runnable

[Gavin-Borges]

Fixes packaging blockers that prevented an installed SESTRAV wheel from being imported or run. All changes are packaging-only; no model, data, or scientific code is modified, and every frozen model artifact is byte-identical.

Problems fixed

• import sestrav failed: the wheel shipped a top-level src package, never sestrav (pyproject include was ["src*"], and no sestrav/ package existed).
• functions.* was missing from the wheel: the CLI predict command imports functions.stage1..stage4 at runtime, but functions/ is a top-level directory that include = ["src*"] excluded, so an installed sestrav predict raised ModuleNotFoundError: functions.
• the release smoke test masked the break: it used import sestrav ... || from src import features, and the fallback passed on the broken wheel, keeping the job green.

Changes

• pyproject.toml: expand the wheel package set to include = ["sestrav*", "src*", "functions*"] so the CLI stages and the new facade package ship. The src/functions layout is unchanged (they are mutually coupled), so no internal import statements change.
• add sestrav/init.py: a thin facade that re-exports the public surface (features, main, __version__) from src, so import sestrav works for an installed wheel.
• .github/workflows/release.yml: drop the || fallback and make the smoke test real - it now asserts import sestrav, import functions.stage1_..., and sestrav --help.

Safety

• No file under models/ is touched. rf_31feature_integrated.joblib and every other joblib/pth artifact remain byte-identical (sha256 verified before/after).
• A pickle-header audit confirmed no frozen artifact references any src.*, functions.*, or __main__.* class; all classes resolve to sklearn/xgboost/numpy or torch state_dicts, so the packaging change cannot affect model loading.

Verification

• Built the wheel; confirmed sestrav/, src/, and functions/ (all 5 stage files) ship; top_level.txt now lists functions/sestrav/src; entry point intact.
• Isolated installed-layout import test (wheel on PYTHONPATH from a neutral cwd): import sestrav, import functions.stage1_peptide_generation, and src.cli.main all resolve from the installed location (not the repo source).
• Frozen RF sha256 unchanged (e2acd33...); integrity trio green (selftest 18/18; harness 20 PASS / 1 WARN / 0 FAIL / 2 SKIP).

Note on scope

This makes a future PyPI publish safe (the de-masked smoke test will genuinely validate the wheel); it does not itself publish. The release.yml:121 pip install "sestrav==X" --index-url pypi.org step still activates only once the package exists on PyPI.