fix(voice): stop the recorder writing over audio still in the resampler

[longcw]

Problem: The recorder writes each block up to _input_settled, which is the arrival time of the frame. A resampler emits only whole output frames, so the tail of the audio is still inside it, and the writer puts silence in that place. take() discards those samples later, because their place is behind the cursor.

Fix: _Track reports placed_through, the point the open run reaches, or None when the resampler holds nothing back. The encode thread keeps each write behind that point.

Context for reviewing and coding agents

How to see it

Three cases in tests/test_recorder_io.py are new and fail on main. test_the_writer_waits_for_samples_the_resampler_still_holds shows the loss directly: it takes a block up to the arrival time and asserts no zero samples inside it.

Blast radius

placed_through has one consumer, the cursor loop in _encode_thread. The end_run() added there is the third call in the file and sits beside the one from #7047 — that one ends a run at playback end, this one ends it when a source stops delivering mid-run. Nothing outside recorder_io.py reads either.

Why the shortfall is computed in _Track and not in the resampler

_Track already holds _run_start, _t0, _sample_rate and _run_samples, so the point the run reaches is derivable from state it keeps. No resampler API change is needed, and a track without a resampler returns None and never holds the writer.

What this does not change

A 48 kHz input has no resampler and takes the None path throughout. The two channels still share one cursor, so this only moves where that cursor may stop, never how the channels align.