[jiri-god]

An unrelated graph rebuild currently resets every latency-compensation delay line. Even adding an unconnected node can briefly drop a delayed signal to zero. Graph scratch storage is also resized/allocated while publishing a new rendering sequence, and changing device block lengths can resize graph output storage in the callback.

This preserves delay histories for matching source/destination node IDs, ports and delay length (not temporary scratch-buffer indices). History transfer and publication are protected by the existing render lock. New scratch audio/MIDI storage is allocated before acquiring publication locks, and old storage is destroyed afterward. Graph output capacity is prepared ahead of time and reused for smaller variable blocks; device input scratch and MIDI buffers are reserved during preparation.

The MIDI reserve is 64 KiB, not an allocation-free guarantee for arbitrary SysEx traffic or plugin output. Graph publication still uses locks, and history matching is O(new operations × old operations); this does not promise glitch-free arbitrary live graph editing.

Validation on Apple Silicon macOS, RelWithDebInfo:

• GraphContinuityTests mixes two constant sources with a 257-sample latency difference. On unmodified d0226686, adding an unconnected node drops all 128 samples in the next block from 0.75 to 0.5.
• The regression also covers varying smaller block lengths and verifies that a changed delay length starts a new history rather than inheriting the old ring.
• With this patch: all 59 CTest suites pass (3.63 s).

Based directly on main and independent of #1211/#1212. Linux/Windows validation is pending CI.

Combined validation: the four independent branches (#1211–#1214) merge without conflicts and all 62 CTest suites pass together on Apple Silicon macOS (4.06 s). Upstream Build/REUSE runs currently report action_required pending maintainer approval for this fork; no cross-platform CI success is claimed.