Refactor Cognito auth into an async, persistence-aware token lifecycle

まず evnex/api.py の認証フロー、特に 120-157、167-207、218-273、275-285、302-310 行を読み、次に tests/test_auth.py の 162-197 行と README の認証例を確認してください。完了とは、非同期認証、永続化されたトークン更新、401 に対する回数制限付きリカバリ、認証の一元化、新しい認証テストを含む、列挙された受け入れ条件がカバーされていることを意味します。

Context

MFA support in #85 / v0.6.0 adds the missing Cognito challenge flow and fixes several immediate reliability problems. The current API now works reasonably for a CLI or a single long-lived Evnex object.

It is still the wrong abstraction for a Home Assistant integration or another long-running application that must persist credentials safely across restarts.

Problems

1. Refresh-token resumption still requires and retains the password

Evnex.__init__() requires both username and password, even when a refresh token is supplied:

https://github.com/hardbyte/python-evnex/blob/main/evnex/api.py#L120-L157

The README says "the refresh token alone is enough", but callers still need to provide a real or dummy password:

https://github.com/hardbyte/python-evnex/blob/main/README.md#L104-L113

For Home Assistant, the password should ideally be used only during interactive bootstrap/reauthentication and then discarded. Normal startup should require only the durable refresh-token session.

2. Refreshed token state cannot be reliably persisted

pycognito mutates access_token, id_token, and refresh_token internally during authentication and refresh. python-evnex exposes properties for reading them, but provides no notification or async callback when the token set changes:

https://github.com/hardbyte/python-evnex/blob/main/evnex/api.py#L167-L207

https://github.com/hardbyte/python-evnex/blob/main/evnex/api.py#L275-L285

A caller can manually inspect the properties after a request, but it cannot reliably or atomically persist every replacement token set. This becomes a correctness issue if Cognito refresh-token rotation is enabled, and is awkward even when only access/ID tokens change.

Token persistence should be owned by the application, with the library explicitly publishing each newly issued token set before other requests proceed.

3. The public authentication API is synchronous

authenticate() and respond_to_mfa_challenge() perform network operations synchronously:

https://github.com/hardbyte/python-evnex/blob/main/evnex/api.py#L218-L273

Internal API calls correctly wrap Cognito operations in asyncio.to_thread(), but a Home Assistant config flow calling the documented public methods must know to add its own executor/thread boundary. This is easy to get wrong and inconsistent with the otherwise async client.

All public methods that can perform network I/O should be async.

4. Public consumers are coupled to pycognito

MFA is surfaced through pycognito exception classes and challenge.get_tokens():

https://github.com/hardbyte/python-evnex/blob/main/README.md#L62-L102

This leaks the implementation dependency into integrations and makes it harder to change the Cognito implementation later. It also models MFA as a special mode + code call rather than an explicit Cognito challenge containing the challenge name, opaque session, and parameters.

Even if Evnex currently uses only SMS and TOTP, python-evnex should expose its own typed challenge/result API.

5. Authentication retries are coupled to generic request retries

A 401 calls _ensure_valid_token(). If check_token() considers the JWT unexpired, the request fails without forcing one refresh. Conversely, if renewal keeps reporting success, TokenRefreshedError can consume the full five-attempt generic retry budget:

https://github.com/hardbyte/python-evnex/blob/main/evnex/api.py#L302-L310

https://github.com/hardbyte/python-evnex/blob/main/tests/test_auth.py#L162-L197

Authentication recovery should have independent semantics:

1. Send with the current valid token.
2. On 401, force one single-flight refresh.
3. Retry the request once.
4. Surface a reauthentication-required error if it still fails.

It should not share the generic transient-network retry counter, particularly for command endpoints that may not be safe to send repeatedly.

6. Authentication is applied separately to every API method

@refresh_token_if_expired is repeated across resource methods. This is easy to omit when adding a new endpoint and mixes transport/auth concerns into each operation.

Authentication and one-time 401 recovery should happen in one central request path or an injected httpx.Auth-like component.

Proposed direction

Introduce a separate async authentication component, while retaining a backwards-compatible convenience constructor initially.

For example:

@dataclass(frozen=True)
class TokenSet:
    access_token: str
    id_token: str | None
    refresh_token: str
    expires_at: datetime | None


@dataclass(frozen=True)
class AuthChallenge:
    name: str
    session: str
    parameters: Mapping[str, str]


class EvnexAuth:
    async def get_access_token(self) -> str: ...
    async def force_refresh(self) -> TokenSet: ...
    async def start_authentication(self, username: str, password: str) -> TokenSet | AuthChallenge: ...
    async def respond_to_challenge(self, challenge: AuthChallenge, response: str) -> TokenSet | AuthChallenge: ...

The application should be able to supply an async token-update callback or token store:

async def save_tokens(tokens: TokenSet) -> None: ...

Evnex would then accept an auth object and an httpx.AsyncClient, without needing to retain a password itself.

Acceptance criteria

• A client can resume using a refresh token without supplying or retaining a password.
• Every successful authentication/refresh returns or publishes the complete current token set.
• Token updates can be persisted atomically by the caller.
• All public network operations, including initial auth and MFA responses, are async.
• Public exceptions and challenge objects do not expose pycognito types.
• Concurrent API calls perform at most one refresh.
• A 401 triggers at most one forced refresh and one request retry.
• Auth retry behaviour is separate from generic transient HTTP retries.
• Authentication is enforced through one central request path rather than per-method decorators.
• Tests cover refresh-token-only startup, token-update persistence, concurrent refresh, MFA challenge round-tripping, and bounded 401 recovery.

This does not need to block the current MFA release. It is a follow-up refactor to make the new support robust for Home Assistant rather than merely functional in a single-process example.