P1: every repo, route, scheduler path and test now speaks one async DB
interface (all/get/run/exec/transaction). better-sqlite3 stays underneath
for dev/tests via SqliteDb (statement cache, manual BEGIN/SAVEPOINT
nesting); openDb stays sync so fixtures remain one-liners. Zero behavior
change: 346/346 tests, typecheck clean.
P2: migrations-pg.ts — numbered Postgres migration set (001-init: bigint
paise, integer flags, TEXT ISO dates, TEXT JSON; reminder CHECK born in
final widened form) + schema_migrations runner.
P3: db-pg.ts — pg Pool behind the same interface: ?->$n rewrite outside
string literals, transactions pinned to one pooled client via
AsyncLocalStorage with savepoint nesting, int8/numeric parsed to JS
numbers so paise math is identical on both engines.
Also: reminder upsert INSERT OR IGNORE -> ON CONFLICT DO NOTHING
(portable, same changes-count semantics).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ORDER BY effective_from DESC alone picks an engine/plan-dependent winner
when two reminder_schedule rows for the same rule_kind share an
effective_from (a same-day cadence correction under rule 3's
new-row-never-edit makes this likely). Append id DESC — UUIDv7 ids are
time-ordered, so the latest insert wins deterministically on SQLite and
Postgres alike.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
reminder_schedule (id, rule_kind, effective_from/_to, day_offsets CSV,
subject, body) added to SCHEMA (CREATE TABLE IF NOT EXISTS, exec'd on every
open = the additive migration path). resolveSchedule(db, ruleKind, today)
picks the row active on the business date (from<=today, to NULL or >today,
latest from wins) with code-constant fallbacks (quote_followup 3/7/14 +
spec s7 subject/body; invoice_overdue 7/15/30). seedIfEmpty seeds one
open-ended dated row per kind, audited in the same transaction.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Thomas Joise