Follow-up to #2336 (umbrella) / PR #2463 (partial sweep). Covers Area 8 deferred from that PR.

Context

The global `runPreparedReply` mock in `src/auto-reply/reply/get-reply.test-mocks.ts:39-40` replaces the real function with `vi.fn(async () => undefined)`, preventing any integration test from exercising the `resolveReplyDirectives` → `runPreparedReply` boundary. This is one of the five shields that hid #2334 from CI until a user-facing crash surfaced.

Scope

• Delete the global `runPreparedReply` shield mock in `get-reply.test-mocks.ts:39-40`.
• Add a regression test that drives `runPreparedReply` with `resolvedThinkLevel: undefined` (or whatever the replacement contract is post-gut(auto-reply): finish thinking-level and model-selection sweep — follow-up to #2335 #2336) through the real `resolveReplyDirectives` chain — no fabricated fixtures.
• Consider a lint rule that flags `as never` / `as any` in test arrangement code (every fabricated shape is a pending production crash).

AC

• Mock removed, existing tests still pass (may require fixture updates where they implicitly relied on the shield).
• New regression test fails on the resolveReplyDirectives → runPreparedReply integration boundary when the chain is broken.
• `pnpm check` + `pnpm test` pass.

Context

• Parent: gut(auto-reply): finish thinking-level and model-selection sweep — follow-up to #2335 #2336
• Sibling shipped: PR gut(auto-reply): partial thinking-level + model-selection sweep — Areas 1+2+4 of #2336 #2463
• Related incident: gut(auto-reply): remove dead thinking-level and model-selection plumbing — TypeError on first-turn messages #2334 (the crash this shield hid)