[bdbarnett]

Stacked on #4. A rack attached to the audio pump can now change its patch while it plays: plug or pull a cable, add or remove a module.

recompile() on an attached graph compiles the new graph while the pump plays on, then parks the pump at a block boundary, carries the state and trades patches with Graph.swap. Whatever pulls the graph (the pump, an output, a mixer) keeps the same object and hears the new patch from its next block. An instrument's connect(), disconnect(), add_module() and remove_module() now work with output="graph"; every change made in one turn lands together on the block after the turn. A module that still runs in Python is refused, and the old patch plays on.

It needs audiodsp's Graph.swap and carry-while-parked, PyDevices/audiodsp#350, which isn't released yet. Without it an attached recompile raises RuntimeError as before, and the new tests skip (MicroPython) or fail (CPython) on an audiodsp without it.

What was checked. tools/pump_swap.py plays the cable phrase on the pump, plugs a cable and opens an attenuator in one held batch mid-phrase, and compares every block the pump played with a rack built with the cable whose attenuator opens at the block the swap reports:

	CPython (service pump)	MicroPython unix (pthread pump)
swap	block 40, byte-equal	block 64, byte-equal
planted: state not carried	caught at 40	caught at 64
planted: swap a block late	caught at 40	caught at 70
planted: carry without parking	no fault (nothing renders between calls)	refused, old patch plays on whole

No block was dropped (the ring's overflow count stays 0). tests/test_graph_pump_swap.py (and its MicroPython twin) hold these. Scoped runs: test_graph_pump_swap, test_graph_live_patch, test_graph_pump and test_adapter pass on CPython (199); the three MicroPython modules pass.

What it costs. Eight swaps of a held chord per voice. "Held" is the time from asking the pump to park to letting it go, so how long it could not render:

voice	nodes	CPython held (worst)	compile	MicroPython held (worst)	compile
Mono101	26	0.045 (0.067) ms	0.52 ms	0.30 (0.41) ms	1.6 ms
DuoSynth	58	0.10 (0.13) ms	0.78 ms	0.45 (0.54) ms	2.5 ms
Poly6	99	0.20 (0.32) ms	1.9 ms	0.45 (0.51) ms	5.5 ms
PolyXa	150	0.13 (0.15) ms	2.6 ms	0.69 (1.01) ms	8.6 ms
Poly8	202	0.16 (0.18) ms	2.6 ms	0.63 (0.71) ms	8.1 ms
PolyA6	676	0.42 (0.54) ms	7.0 ms	1.49 (1.73) ms	54 ms

The compile runs while the pump plays, so only "held" eats into the ring's headroom.