fix: adopt ty 0.0.73 optional-import inference

[lesnik512]

Fixes #155.

What broke

The weekly dependency check's lint job failed on run 32002388707; every pytest job passed. ty 0.0.73 (2026-08-18) changed how it infers the optional-import fallback in faststream_outbox/subscriber/usecase.py:

try:
    import asyncpg as _asyncpg
except ImportError:  # pragma: no cover
    _asyncpg = None  # ty: ignore[invalid-assignment]

It no longer errors on the assignment; it widens _asyncpg to <module 'asyncpg'> | None. That produced 4 diagnostics:

• warning[unused-ignore-comment] — the invalid-assignment ignore no longer suppresses anything (:52).
• error[unresolved-attribute] x3 — the module-level annotations "_asyncpg.Connection" at :359, :418, :462 now resolve against the union, and the None arm has no Connection.

The in-function uses at :440/:443 were unaffected, because the if _asyncpg is None: return None guard at :431 narrows them.

Fix

Split the type name from the runtime name. _asyncpg stays the guarded, nullable runtime handle; asyncpg joins the existing if typing.TYPE_CHECKING: block and backs the annotations — the same treatment AsyncConnection / AsyncEngine already get in this file. No runtime behavior change, no public API change, no test change (tests/test_unit.py patches ..._asyncpg.connect, a name this PR leaves alone).

Pinning ty<0.0.73 was considered and rejected: just install runs uv lock --upgrade and the lint group deliberately floats ruff/ty, so #153 set the precedent of adopting the upgrade. The new inference is also the more correct one.

Planning lane: Tiny (linter tweak, no change file) — same as #153.

Drive-by

Justfile -> justfile, with the CLAUDE.md link retargeted. The link would otherwise 404 on a case-sensitive checkout. The Justfile mentions left in planning/ are historical records of past changes and are intentionally untouched.

Verification

Run locally on this branch:

• just lint-ci — eof-fixer, ruff format --check, ruff check --no-fix, ty check, planning/index.py --check all pass.
• uv run pytest tests/test_unit.py tests/test_fake.py tests/test_client_contract.py --no-cov — 400 passed, 18 skipped.

Not run locally: just test (the full suite incl. tests/test_integration.py) — no Docker or local Postgres on this machine. CI covers it across 3.11-3.14 + free-threaded.

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	839	1.000	6.79	916	242	5000	5000	10000
consumer/w1/b100	883	1.000	6.84	920	243	5000	5000	10000
consumer/w2/b10	990	1.000	6.80	965	242	5000	5000	10000
consumer/w2/b100	1253	1.000	6.78	997	263	5000	5000	10000
consumer/w4/b10	885	1.000	6.73	942	243	5000	5000	10000
consumer/w4/b100	1420	1.000	6.87	1037	243	5000	5000	10000
consumer/w1/b100/tfbs100	4828	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2177	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).