Fidelity: callback-answer discipline, message-id/edit/stale lifecycle, Redis test storage (v0.3.0)

[0x216]

Why

teremok's promise is test green ⇒ works in production. Two bug classes recently slipped a fully-green MockBot suite and only surfaced on a live bot:

• Eternal spinner — a callback handler returned without callback.answer(). Real Telegram leaves a loading spinner on the button for ~30s; MockBot auto-acked every callback, so the test passed.
• "Screen out of date" / menu freshness — handlers that gate on callback.message.message_id == stored_menu_id (delete-and-resend menus) have a stale branch that a mock minting a fresh carrier id on every tap can't reach — and returning a new id from an edit (Telegram keeps the id) mismodels the stored id itself.

Both are pure fidelity gaps, the same kind v0.2.0's validation layer set out to close.

What's in this PR

1. Callback-answer discipline — kills the eternal-spinner class (opt-in: strict_answer=True)

• A second AnswerCallbackQuery for the same id raises the genuine TelegramBadRequest the live API returns (query is too old … or query ID is invalid), via the same check_response route as validation.
• A dispatched, handled callback the handler never answered fails the step with CallbackNotAnswered.
• DispatchResult.assert_answered() (+ .answered / .callback_query_id) asserts a single step without global strictness; MockedSession.answered_callbacks records ids regardless of the flag.
• Opt-in, default off so existing suites that dispatch un-answered callbacks stay green (see migration note). Catches the bug the ElfBot team had to hand-roll a _StrictAnswerSession for.

2. Message-id + edit + stale-message lifecycle — kills the menu-freshness class

• Edits now keep their message_id (an inline-message edit returns True); a fresh send still gets a new id.
• MockBot.sent_messages / last_message expose the on-screen message history, so a test taps a button on an earlier, now-stale message (MockCallbackQuery(message=old_msg, data=...)) and reaches the handler's "screen out of date" branch. Regression test walks: open menu → tap current → bot pushes a new screen → tap the old one → out-of-date branch fires.

3a. Redis-backed FSM storage for tests (optional redis extra)

• fake_redis_storage() → a real aiogram RedisStorage over in-process fakeredis (same JSON serialization + key builder as prod, no server). Passed as MockBot(router, storage=...).
• Surfaces FSM bugs MemoryStorage hides — e.g. stashing a non-JSON value in state data: green on memory, raises on Redis (test included).

3b. Concurrent-update driving — deferred to a follow-up issue

Reliable per-step request attribution under asyncio.gather needs a contextvar stamp threaded from feed_update into MockedSession.make_request; a naïve gather() would mis-slice requests[start:]. Design sketch is in docs/superpowers/specs/2026-07-18-…-design.md and the tracking issue, rather than ship a half-baked version.

Compatibility / migration

• No breaking changes. strict_answer and storage are new keyword-only options defaulting to prior behaviour; DispatchResult gained two fields with defaults. Existing public API and all pre-existing tests unchanged.
• Migration: opt in per-bot with mock_bot(router, strict_answer=True) (or assert per-step with result.assert_answered()). A future major version may flip the default.

Checks

• Full suite green (130 passed, +19 new across test_callback_answer.py, test_message_lifecycle.py, test_redis_storage.py).
• ruff check . clean, mypy src strict clean, coverage 96% (gate ≥90%).
• Docs updated: README, CHANGELOG (0.3.0), quirks table (#17–20), design spec.

Note: the separate coverage-table CI leg is independently stale vs aiogram 3.30.0 (new *EphemeralMessage* methods + subscription update) — pre-existing on main, intentionally not regenerated here to keep this diff focused on the feature.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UxTypHUUDsvB6coXwEKyaX