feat(db): add ai_chat_log so the ask endpoint is actually rate limited

`ask` has never been deployed, and the CI step in this branch will publish it.
Its limiter counts rows in ai_chat_log to cap requests at 8/min and 120/day per
IP, but no migration ever created that table, and the limiter fails open when
the count cannot be read.

Shipping `ask` without this would put an unauthenticated endpoint with
Access-Control-Allow-Origin: * in front of paid model APIs with no limit at all.

Applied to the live database as well.

supabase/migrations/20260730_ai_chat_log.sql

@@ -0,0 +1,28 @@
+-- Rate-limit ledger for the public `ask` edge function (news.pex.mom assistant).
+--
+-- `ask` is unauthenticated and sends `Access-Control-Allow-Origin: *`, so this
+-- table is its only throttle: 8 requests per IP per minute, 120 per day. The
+-- limiter counts rows here and fails open when the count cannot be read, which
+-- means shipping `ask` without this table would leave a public endpoint calling
+-- paid model APIs with no limit at all.
+
+create table if not exists public.ai_chat_log (
+  id         bigserial primary key,
+  ip         text        not null,
+  created_at timestamptz not null default now()
+);
+
+-- The limiter always filters on both columns together (ip = ? and created_at >= ?).
+create index if not exists ai_chat_log_ip_created_at_idx
+  on public.ai_chat_log (ip, created_at desc);
+
+-- Only the service role touches this table; the function reaches it through
+-- PostgREST with the service key, which bypasses RLS. Enabling RLS without any
+-- policy therefore keeps the behaviour intact while denying anon and
+-- authenticated clients, who have no reason to read a table of IP addresses.
+alter table public.ai_chat_log enable row level security;
+
+revoke all on public.ai_chat_log from anon, authenticated;
+
+comment on table public.ai_chat_log is
+  'Rate-limit ledger for the public ask endpoint. Holds IP addresses; prune rows older than the 24h window the limiter uses.';