Voice asks which device you mean when two fit the name

[AboveColin]

When the action answer is sure but the entity answer is split between two exposed devices, the voice agent now asks "Do you mean A or B?" and does not fall back. The result has continue_conversation=True, so a satellite opens the microphone for the reply.

How it decides to ask (interpret._two_that_fit):

• Exactly two exposed devices each hold at least ASK_BACK_FLOOR (0.2) of the entity answer. A third device above the floor means that one question does not settle it, so the agent falls back as before.
• The two together reach the confidence floor the agent acts on.
• Their names differ, or the names are the same and the rooms differ. Then the question names the room: "Kitchen light (Kitchen) or Kitchen light (Office)". When nothing tells them apart, the agent falls back.

0.2 is not measured on a real instance. The comment says so.

How the reply is handled:

• The agent keeps the command, the first response and the snapshot, keyed by conversation id, until CONVERSATION_TIMEOUT (5 minutes, Home Assistant's chat session lifetime).
• The reply gets one Choice question over the two devices and "neither". Its state is {"command": ..., "reply": ...}. It goes through the same budget check, reservation and error handling as a command.
• A confident pick runs the first command with the entity answer settled. interpret runs again on the first sentence, so a brightness that the sentence named still comes from the text.
• A reply that picks neither, or picks without confidence, is handled as a new command.
• The question goes into the chat log as assistant content, so an LLM fallback agent sees it in the history.

The handler is split into _ask (the budget check and the call), _trace, _act (everything after interpret, unchanged), _ask_which and _resolve. The new line which_device is in strings.json and all 13 translations.

Tests: 12 new tests in test_conversation.py. They cover the question, the pick, neither, an unsure reply, another conversation, expiry, a third device, same name and same room, same name and two rooms, a Dutch pipeline, and a reply past the budget.

Changes after live testing on a development instance:

• A shared name is asked about when the model put most of its answer on "none". Measured: none 0.55, one Lamp 0.44, the other 0.01. The agent now reads the top device in that case.
• The reply's trace shows which device the reply picked, and no longer fails with a missing action key.
• The reply gets a second question, new_request: does it ask for something of its own? Pure picks scored 0.08 to 0.26, new commands 0.91 to 0.97. At 0.5 or more the reply is handled as a new command, so "never mind, turn off the lamp in the bedroom" turns that lamp off.
• The room of the satellite that heard the command settles a shared name, as it does for Home Assistant's own agent. The lookup is the same: the satellite entity's area, then its device's area. Live: "turn on the lamp" with two lights called Lamp turned on the one in the device's room, 4 of 4. A device in a third room, or in no room, still got the question.

Local gate: ruff, ruff format, mypy --strict, pytest (433 passed, 12 skipped), config_flow 100%, mkdocs --strict, hassfest. EXIT 0.

Part 3 of the 1.16 work. Stacked on #21.