Add bounded pedal multi-delay audition and profiling package

[Denys]

Manager summary

Adds a deterministic host-audition and timing-evidence package around the existing five-mode PedalDelayEngine without changing the pedal UI, public parameter contract, engine implementation, or target firmware.

This makes every compact mode reproducibly renderable and measurable for engineering review:

• DIGI
• TAPE
• MOD
• REV
• FREEZE

The result is deliberately named HOST_AUDITION_PACKAGE_READY, not HOST_LISTENED_ACCEPTED or TARGET_READY.

What changed

• adds a standalone C++17 runner under DaisyHost/tools/pedal_multidelay_audition/;
• adds 15 deterministic cells, three per mode: reference / character / stress;
• emits 15 stereo float WAVs, 15 per-cell JSON records, an aggregate audition manifest, host timing CSV, and environment JSON;
• records finite/non-silent status, peak, RMS, tail RMS, max adjacent step, stereo-difference RMS, and deterministic FNV-1a checksum;
• implements bounded host profiling with:
  • 1,000 warm-up blocks;
  • 50,000 measured blocks per mode;
  • arithmetic mean;
  • p99 and p99.9 via offline std::nth_element;
  • observed maximum, explicitly not WCET;
  • p99.9 margin/utilization;
  • observed execution deadline misses/rate/lateness;
  • timer-bracket overhead and environment metadata;
• adds the corrected Real-Time DSP Profiling System v1.1 candidate;
• adds a source-grounded work-package contract separating host render evidence, host timing, human listening, and physical target feasibility;
• adds a scoped GitHub Actions gate that runs existing focused PedalDelay tests on Windows and the bounded renderer/profiler on Ubuntu.

Public and parallel-work contract

Unchanged:

• app id pedal_multidelay;
• mode and slot identity/order;
• canonical parameter ids and descriptor schema;
• freeze operations;
• bypass/trails and tap/hold behavior;
• PedalDelayEngine.h/.cpp;
• PedalDelayCore.h/.cpp;
• JUCE editor/processor and board-profile/UI files.

The final diff contains exactly seven newly added files. No pre-existing source, shared tracker, JUCE/UI file, engine file, core wrapper, or target-firmware file is modified.

Exact final evidence

PR head:

f65de8be96b268576e121e65f5423b4852b9c44c

Base inspected:

041bfcbb7975316606ece8f1ccbacbab9c0e1482

Scoped multi-delay workflow

GitHub Actions DaisyHost Pedal Multi-Delay Audition run 31696153621: SUCCESS.

Actually executed on the final PR head/merge ref:

• Windows Server 2022 / MSVC 19.44 Release-focused build;
• existing focused PedalDelay CTest suite: 37/37 PASS;
• Ubuntu 24.04 / GCC 13.3 Release bounded runner build;
• bounded audition/profiling CTest: PASS;
• audition cells: 15/15 PASS;
• generated evidence files: 33;
• artifact ID: 9179550895;
• artifact size: 10,391,426 bytes;
• artifact SHA-256: 90f4a152b95475c305fe783a9749724c5fb3a9fcc024190219e705cadd66d865.

Host-only timing evidence

Configuration:

48 kHz
48 frames per channel
1,000 warm-up blocks
50,000 measured blocks per mode
50 s measured audio per mode
std::chrono::steady_clock
GCC 13.3.0 Release

Mode	Mean µs	p99 µs	p99.9 µs	Observed max µs	p99.9 load	Observed execution deadline misses
DIGI	2.7948	2.982	8.256	25.014	0.8256%	0
TAPE	6.7337	9.204	14.239	45.173	1.4239%	0
MOD	2.4914	2.583	7.941	21.335	0.7941%	0
REV	3.3256	3.614	9.085	53.195	0.9085%	0
FREEZE	7.4889	10.443	14.039	32.848	1.4039%	0

These measurements are regression/characterization evidence for the CI host only. They are not converted into Cortex-M7 margin.

Repository checks

• Fix Style run 31696153590: SUCCESS using the repository-pinned clang-format 10 path;
• Build All run 31696153589: SUCCESS for repository libraries, embedded examples, and distributable generation; deploy step skipped for the PR event as designed.

Build All is not the full DaisyHost/JUCE milestone gate and is not presented as such.

Known real-time safety hold

The runner calls PedalDelayEngine directly. The existing live PedalDelayCore::Process() wrapper still contains a scratch-vector growth fallback for an oversized callback. That shared wrapper file is intentionally untouched here.

Current evidence boundary:

ENGINE_PROCESS_ALLOCATION_FREE: VERIFIED_SOURCE
CORE_WRAPPER_ALLOCATION_GUARANTEE: HOLD
RT_SAFETY_EXCEPTION_REQUIRED: YES

The smallest coordinated follow-up is to preallocate before audio start, fail closed on an oversized block, add a negative no-growth test, and rerun the focused/full DaisyHost gates.

Diagnostic precursor runs

Earlier source-branch runs intentionally exposed and corrected:

• a freeze character cell whose final tail was below the declared evidence threshold;
• an unsuitable Linux route for the complete DaisyHost/JUCE configure;
• missing selective submodule initialization in the Windows focused-test job.

Those diagnostic failures did not modify the current runtime engine. The exact final-head workflow and repository style/build checks listed above are green.

Explicitly not claimed / not run

• human listening acceptance;
• full serialized DaisyHost 337/337 plus Hub/Render/VST3/Standalone milestone gate;
• manual standalone or DAW/VST3 interaction;
• ARM/Daisy build, flash, or callback timing;
• target SDRAM/cache behavior;
• physical audio, controls, electrical, EMI/EMC, mechanical, or production readiness;
• parity with the current portable product DSP/firmware implementation.

Review focus

1. bounded timing/statistics implementation and terminology;
2. scenario/event semantics and evidence schema;
3. host-vs-target claim boundary;
4. whether the isolated tool/workflow should remain standalone or later be integrated by the main DaisyHost integrator;
5. the separately recorded PedalDelayCore allocation-safety follow-up.