Runs the address reap's writes on SQLite

[johnnyt]

Fixes StatifierRouter.Addresses.reap/3 on SQLite. Refs sr-1rgb.

The bug

The private stamp/3 and delete/2 of StatifierRouter.Addresses named their rows with fragment("? = ANY(?)", a.id, ^ids), Postgres's own form, since 0.8.0's :primary_key option. On SQLite the statement fails with "no such function: ANY", so every reap of a SQLite host that finds a row to stamp or delete fails. The new tests reproduce it on the base commit, under both key types.

The fix

Both functions now use fragment("? IN (?)", a.id, splice(^batch)): an IN list with one bound parameter per id, which Postgres and SQLite both take. Each id is still bound as the table handed it over, never cast, so the rule StatifierRouter.Schema.Id documents holds as before: a text id made of digits stays a string.

The batch bound: a list of more than 500 ids is written in batches of 500, one statement each (the private in_batches/2, @ids_per_statement). 500 ids plus the stamp's time stays under SQLite's smallest limit on bound parameters (999, the default before SQLite 3.32) and far under Postgres's 65535, so a :limit above the default cannot build a statement either adapter refuses. The ANY form bound one array whatever the count; a single IN list without batches would have made a large :limit fail on Postgres too. The counts reap/3 answers are the sums over the batches, the same numbers as before, and no ids still sends no statement. At the default :limit of 1000, a reap that stamps or deletes more than 500 rows now sends two statements for that write where it sent one.

Postgres answers are unchanged: the existing reap tests in StatifierRouter.CreateModesTest and StatifierRouter.PrimaryKeyTest pass unchanged; neither file is in this diff.

Tests

• test/statifier_router/sqlite_reap_test.exs, StatifierRouter.SQLiteReapTest, on the SQLite test repo: a reap that stamps two delivered parcels' rows and deletes an orphan's and one stamped a day ago, then deletes the two it stamped an hour on, under the default integer key and under a :primary_key text key whose ids are digits with a leading zero; and a reap over 1201 rows, past two batches.
• test/support/status_store.ex, StatifierRouter.StatusStore: a storage adapter answering fetch_execution/2 from a map, so the SQLite repo needs no execution table.
• StatifierRouter.SQLiteReapTest and the existing StatifierRouter.SQLiteMigrationsTest both start StatifierRouter.SQLiteRepo under its module name, so both are in the :sqlite_repo ExUnit group, whose modules never run at once. The first CI run of this branch failed on exactly that race (already_started); without the group a local --repeat-until-failure 100 run of the two modules failed on its first run, and with it 100 runs passed.

Sabotage, one row per check, each restored byte-equal from a copy before the next:

Mutation in addresses.ex	Result
stamp/3 and delete/2 back on ? = ANY(?)	the three SQLite tests red on their reap assertion
each id cast to an integer before it is bound	the text-key test red: the reap stamped and deleted nothing
in_batches/2 answering the last batch's count alone	the batch test red: 201 stamped for 1201
delete/2 on NOT IN	six CreateModesTest reap tests red on Postgres

Records and changelog

• ADR-0002 gets a dated foot Note: its Note of 2026-09-27 accepting the :primary_key Amendment says stamp/3 and delete/2 bind the ids with = ANY. The new Note says what they bind with now and that the Amendment's rule, that the package never casts an id it binds, holds. It decides nothing and removes no line (git diff origin/main -- docs/adr/ is additions only).
• changelog.d/sr-1rgb.md under ### Fixed.

Provenance

Two engineering choices: an IN list with splice/1 over a per-id OR of ? = ? comparisons (one flat list rather than a nested expression SQLite's depth limit would bound), and a batch of 500 ids, for the limit above. No dependency changes: mix.exs and mix.lock are not in this diff.

The group line in test/statifier_router/sqlite_migrations_test.exs is a change to a test module this bead did not otherwise touch, forced by the new module sharing its repo.

Gate

Full mix quality on this head: Format, Compile, Isolated tests, Doc links, Dependencies, Credo, Docs, Tests (425 of 425 passed, 97.7% coverage) and Dialyzer all passed; Doctor, Gettext and Sobelow skipped as not installed.