Give the library a layer model, and an architecture contract to keep it

[Varfalamei]

Why

The library had no declared structure, so every change added a module and the shape eroded. Three concrete symptoms, all verified:

1. The research layer was loaded inside Triton. recommender_als.py imported rrf_fuse from policy/fusion.py, which executes policy/__init__.py, which imports PolicyModel and rectools' model_from_config. A broken import in offline-only code would have taken down inference for every model in the bucket, including als_youtravel.
2. The co-visitation algorithm existed three times. covis_kernel.py had been extracted to unify the serving and research copies, but only the serving caller was ported, so the kernel was a third copy sitting at the package root with no layer.
3. Four modules were named model.py, and smartrec_lib.model (settings) vs smartrec_lib.models (a model) differed by one character while meaning opposite things.

What

CLAUDE.md at the repo root states the architecture contract: four layers with imports flowing strictly downward, a table saying where a new file belongs, naming rules, and the 13 invariants that break production silently when violated (the pickle is a bare dict, so a renamed attribute does not raise — it reverts to the __init__ default and the model quietly serves degraded results).

model.py          L0  contract: RecomItems, Strategy, *Settings   (frozen by pickles)
kernels/          L1  pure algorithms: cooccurrence, fusion, constraints
recommenders/     L2  the servable models                         (paths frozen by pickles)
serving/          L2  Triton python backend
research/         L3  rectools-native CoVisModel and PolicyModel
evaluation/       L3  offline protocols

• policy/fusion.py, policy/constraints.py → kernels/; research no longer loads in serving.
• covis_kernel.py → kernels/cooccurrence.py, and research/covis.py is ported onto it, so the algorithm exists once and the two callers differ only in the parameters they pass.
• models/ and policy/ → research/; the two remaining model.py are the two that are forced, one by the pickles in S3 and one by Triton's file-name convention.
• Recommenders import RecommenderModel from recommenders.base instead of from their own package, so __init__.py line order stops being a correctness requirement.

Also fixes the red CI on master

test_fit_popular and test_fit_als pinned exact result lists over tied items. The fixture holds 100 rows over 99 items, so 98 tie at one unique user; test_fit_popular expected three arbitrary members of that tie, and not even the genuinely most popular item, which filter_viewed drops. They now assert what the fixture determines. This is why the suite passed locally and failed in CI with no behaviour change.

Adds unit tests for the kernel, which no test reached before — including top_k truncating a tie, the regime where the two callers can genuinely diverge.

Safety

Nothing frozen moved. Verified against a freshly trained als_covis artifact: the pickle references smartrec_lib.model and smartrec_lib.recommenders.recommender_covis and nothing else, so artifacts already in S3 load unchanged; a dill round-trip still serves covis_session. 99 tests pass under PYTHONHASHSEED 0, 1, 42 and 12345; flake8 hard gate is clean.

Deploy note: this is a code change, so it needs a runtime repack (make pixi-pack-build-docker + upload) before the retrain that ships it — recommend() comes from runtime_env.tgz, not from the pickle.