fix: resolve review findings (audio-thread safety, robustness)

[yves-vogl]

Summary

• Fixes Unconditional per-block heap allocation in OvertureEngine::process() via IIR::Coefficients::make{High,Low}Pass #12 — OvertureEngine::process() unconditionally called IIR::Coefficients<float>::makeHighPass/makeLowPass for the Tight HPF and both Tone LPF stages every block, each new-ing a fresh ref-counted Coefficients object (up to 6 allocations/6 deallocations per processBlock() call on the audio thread). Replaced with juce::dsp::IIR::ArrayCoefficients (stack-computed) written directly into the already-allocated Coefficients storage via a new src/dsp/RealtimeCoefficients.h helper — the same non-allocating pattern already used by sibling plugins Lancet and Twist Your Guts.
• Fixes processBlock has no defensive clamp against block sizes exceeding the size declared to prepareToPlay #13 — processBlock()/OvertureEngine::process() had no clamp against a host handing more samples than the spec.maximumBlockSize declared to prepareToPlay()/prepare(); the oversampler's and DryWetMixer's internal buffers are both fixed to that size and only guarded by a jassert (compiled out in Release builds). OvertureEngine::process() now defensively splits any oversized block into chunks of at most the prepared maximum (each run through a new private processChunk()) before touching either internal — the same real, Release-safe chunking guard already used by sibling plugins Twist Your Guts and Miserere.

Test plan

• tests/AllocationTests.cpp (new): globally instruments operator new/delete (AllocationGuard.h/.cpp, ported from sibling plugin Seraph's proven pattern) and fails if OvertureAudioProcessor::processBlock() or OvertureEngine::process() ever allocates while Tight/Tone are being automated. Confirmed red (96 allocations observed) against pre-fix main, green after the fix.
• tests/EngineTests.cpp (new case): feeds a 64x-oversized block (8192 samples against a 128-sample prepare()) through OvertureEngine::process() and checks it produces output numerically identical to a correctly host-chunked reference run. Confirmed red (16175/16386 assertions failed) against pre-fix main, green after the fix.
• Full local suite green: cmake --build build --target Tests Overture_Standalone && ctest --test-dir build --output-on-failure — 54/54 passed (51 pre-existing + 3 new).
• Verified both new regression tests actually fail on pre-fix main via an isolated git worktree build (not just asserted from reading the diff).

🤖 Generated with Claude Code