midi: drop the dead time argument, and inject_midi with it

[dpwe]

The time argument was already dead

In amy_midi.c, time is declared AMY_UNSET_VALUE(time) (lines 321, 395) and never reassigned — so every real MIDI path (USB, gadget, UART, mac, sysex) has always passed it unset. amy_received_control_change() didn't even read its copy. Nothing outside src/ called these at all. The one caller that ever supplied a real value was pyamy's inject_midi(), i.e. the test suite.

That reflects the truth of the thing: a live MIDI message has no time of its own — it plays when it arrives. So the argument is gone from amy_event_midi_message_received() and the five amy_received_* handlers.

inject_midi goes with it

It was the weaker of the two Python entry points:

• It hardcoded len=3, so it could not send a 2-byte message at all — test.py already used inject_midi_bytes([0xC0, 5]) for program change.
• It handed a pre-formed message straight past the byte-stream parser.

inject_midi_bytes() runs the real convert_midi_bytes_to_messages(), so the tests now exercise running status and real-time interleaving the way actual MIDI input does. Net gain in coverage.

The ~16 test call sites that passed an absolute millisecond time move to a new amy_inject_midi_at(), the MIDI twin of the existing amy_send_at(): render up to the moment, then inject.

What deliberately did NOT change

midi_message_handler_to_queue() keeps its time parameter. It lives in midi_mappings.c, and patches.c:1081 drives it with a real e->time for the wave=AMY_MIDI osc — that's how a tick-scheduled event emits MIDI on the beat. amy_event_midi_message_received() now passes it an explicit unset instead of a laundered one.

Drive-by fixes

• pyamy.c had three bare return Py_None; (missing INCREF — each call over-decremented None's refcount). Now Py_RETURN_NONE.
• The Daisy port's midi_polyphony() demo was staggering notes 1000ms apart with the argument that never did anything; it now says so instead of pretending.

Testing

make test: 122 tests pass, and all 116 waveform comparisons are bit-exact (err=-100.0 dB) — no reference audio needed regenerating, despite the tests now routing through the byte-stream parser and the render-forward clock.

make check-c-api: no generated-API drift. Clean make with no new warnings.

🤖 Generated with Claude Code

[github-actions]

🎛️ AMY HW CI (AMYboard bench)

Flashed this PR's AMY (LoadTestChord: 6-voice Juno patch=1, one held note every 2 s) onto the physical AMYboard and measured the smoothed render load as the chord grows — back-to-back with the same sketch built at the PR's merge base, so Δ is this PR's own cost.

✅ PASS — the bench ran the test to completion.

notes held	main @ 5309014	this PR	Δ
1	1050	1045	-5
2	1200	1198	-2
3	1800	1799	-1
4	1962	1952	-10
5	2609	2610	+1
6	2766	2764	-2

Full chord settled render μs: 2765 (was 2767, Δ -0.1%) (peak 2771, 39 samples)

⬇️ Artifacts: serial log · load trace · report

_{Self-hosted bench (amyboardci). FAIL means only that the test could not run — the load values are informational, with no threshold and no audio compare. See tools/arduino_loadsweep/.}

[bwhitman]

⛓️ tulipcc integration PR opened

This merge was pinned into tulipcc for full-system CI: shorepine/tulipcc#1273

Test it there and merge that PR to move tulipcc onto this AMY.