fix(sql_base,base): make the Mapping contract truthful (contains, len, missing-key policy)

[thorwhalen]

Two commits fixing sqldol's Mapping-contract violations, both scoped to leave
default behavior unchanged:

1. fix(base): make SqlBaseKvReader.__len__ return a real count — __len__
   returned a SELECT's rowcount, which is -1 on drivers that don't pre-buffer
   (SQLite among them), so len(store)/list(store) raised ValueError on those
   backends. Now issues SELECT count(*). Also narrows __getitem__'s bare
   except: to except Exception: (no longer swallows KeyboardInterrupt/
   SystemExit) and adds an opt-in keyword-only missing_key_policy ('empty'
   default = today's behavior, 'raise' makes in/.get() truthful per
   collections.abc.Mapping).
2. fix(sql_base): make \k in store` agree with `store[k]`—SQLAlchemyPersisterhad nocontains, so it inherited the brute-force scan from dol.base.Collection, which compares against iter's yield — but this persister's iteryields ORM row objects, never keys, sok in persisterwasFalsefor every key, including keyspersister[k]resolves happily. Adds a getitem-basedcontains`.

Refs #2 (fixes the __contains__ half only; making __iter__ yield keys instead
of ORM rows is deliberately out of scope — not backward compatible, a known
consumer is written against the row-yielding behavior — and is pinned by a new
regression test). Does not close #2.

Dependents check (fleet_dependents.json lists raglab-app as the sole
dependent of sqldol): raglab_app imports only SqlDictReader/SqlDictStore
(both SqlBaseKvReader/SqlBaseKvStore subclasses, commit 1's class) — never
SQLAlchemyPersister/SQLAlchemyStore/SQLAlchemyTupleStore (commit 2's
classes), so the __contains__ change cannot reach it. Commit 1 keeps
missing_key_policy defaulted to 'empty' (byte-for-byte prior behavior,
pinned by test_missing_key_policy_defaults_to_the_legacy_empty_behavior), and
the __len__/except Exception changes are strictly corrective. raglab_app's
own test suite is intentionally disabled in its CI ([tool.wads.ci.testing] enabled = false — "private Streamlit app with no test suite; its modules use
script-style imports and run code at import time") and its full install needs a
pg_config/postgres toolchain not present on this box, so I could not execute it
directly; the source-level compatibility read above is the substitute. Recorded
in DECISIONS.md.

Branch sat pushed with green CI and no PR for two weeks (thorwhalen/fleet_stuff
cleanup). Verified: master has not moved since the branch was cut (no rebase
needed). sqldol has no pyproject.toml/wads CI config (setup.py/setup.cfg
only), so gated with a plain venv + pytest instead of wads ci-local: 20/20
tests pass (sqldol/tests/test_base_mapping_contract.py,
sqldol/tests/test_sqlalchemy_store_contract.py). Hosted CI on the branch was
already green from the original 2026-09-08 push.

Refs #2

🤖 Generated with Claude Code

[thorwhalen]

Post-merge review notes (design, not fixed here):

1. With the default missing_key_policy='empty', k in store is still True for every key, and that is the default anyone gets. __contains__ does not have to depend on the policy. An explicit __contains__ that runs SELECT EXISTS(... WHERE key = :k) would make in truthful under both policies and would not touch the store[missing] is None callers the default protects. I would add that and keep missing_key_policy for __getitem__/get only.
2. SqlBaseKvStore._mk_column_filter builds its WHERE clause by f-string interpolation into text(). A key containing a quote breaks the query, and it is an injection vector. Use self.table.c[col] == value, the way __getitem__ already does.
3. The same rowcount-on-SELECT trap this PR fixed in __len__ was still in __setitem__. On SQLite, writing to an existing key inserted a duplicate row. Fixed in Fix: SqlBaseKvStore.__setitem__ duplicated rows instead of updating (rowcount trap left by #4) #5.