perf(db): index LOWER(email) for the case-insensitive account lookup (#876)

[alex-dembele]

Closes #876

Migration 0068_users_email_lower_index: a non-unique index on LOWER(email), the predicate GetByEmail uses since #688. It's not unique on purpose: accounts written before #687 can be case-variant duplicates, and a unique index would fail the migration (and the boot with it) on such a database. docs/MIGRATIONS.md gains the read-only query that lists those duplicates.

Verification

Throwaway Postgres 18, server booted from this branch (AutoMigrate, then the SQL layer):

schema_migrations → 68
idx_users_email_lower → CREATE INDEX idx_users_email_lower ON public.users USING btree (lower(email))
down → DROP INDEX (0 indexes left)
seed 20 000 users + 'Karim@Alpha.test' / 'karim@alpha.test'
up with the duplicate present → DO (index back)   · up again → NOTICE already exists, skipping (idempotent)

EXPLAIN ANALYZE of the exact GetByEmail query, 20 002 rows:

without index: Seq Scan on users … Execution Time: 5.870 ms
with index:    Bitmap Index Scan on idx_users_email_lower  Index Cond: (lower(email) = 'karim@alpha.test'::text)
               Execution Time: 0.107 ms

Duplicate report: karim@alpha.test|2|{Karim@Alpha.test,karim@alpha.test}.

Go gate go build ./... && go vet ./... && go test ./... -race: BUILD_OK · VET_OK · TEST_OK, 80 ok, 0 FAIL. It ran with #882 applied temporarily, because master doesn't build (#881); that commit isn't part of this PR.

Honest remainders

• Merge fix(auth): master does not build — restore the #866 SAML handler (#881) #882 first.
• The index is built with a plain CREATE INDEX, which locks writes to users while it builds. That's milliseconds at today's sizes; CONCURRENTLY would need its own non-transactional migration.
• A unique constraint is not proposed: it waits on an operator running the duplicate report on real data.