refactor(benchmarks): subcommand-scoped CLI flags via subparsers

[lesnik512]

What

The benchmarks CLI registered command as one positional with choices and hung every flag (--messages, --repeats, --write-baseline, --markdown) on the single top-level parser. argparse accepted any flag with either command, but each flag is read by only one — so python -m benchmarks run --markdown parsed and silently did nothing.

Split into real run/check subparsers, each owning its own flags. Misplaced flags now error at parse time (error: unrecognized arguments: --markdown, exit 2).

How

• _build_parser() builds the two subparsers (run: --messages/--repeats/--write-baseline; check: --markdown); dispatch stays if args.command == "check".
• main(argv=None) takes an optional argv so the grammar is unit-testable without spinning up Postgres.
• No change to the happy paths the just recipes use (run …, check, check --markdown).

Testing

• 5 new grammar tests: run --markdown and check --write-baseline raise SystemExit; check --markdown / run --messages … --write-baseline parse to the right namespace; a subcommand is required. 20/20 in tests/test_benchmarks.py pass; ruff + ty clean.
• Verified live: python -m benchmarks run --markdown exits 2; --help lists the run/check subcommands.

Design: planning/changes/2026-07-17.04-benchmarks-cli-subparsers.md

🤖 Generated with Claude Code

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	1047	1.000	6.78	917	242	5000	5000	10000
consumer/w1/b100	1069	1.000	6.82	919	243	5000	5000	10000
consumer/w2/b10	1224	1.000	6.79	954	242	5000	5000	10000
consumer/w2/b100	1476	1.000	6.72	955	243	5000	5000	10000
consumer/w4/b10	1103	1.000	6.77	960	242	5000	5000	10000
consumer/w4/b100	1683	1.000	6.86	1035	244	5000	5000	10000
consumer/w1/b100/tfbs100	6994	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2227	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls/select_calls, exact; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).