A passing process is not a passing instrument

Baseline matrix

zig build test
zig build test -Doptimize=ReleaseSafe
zig build test -Dlive=false
zig build -Doptimize=ReleaseSafe

The #29 reconciliation baseline is 218 tests, plus a standalone null-backend C audio executable. This is a dated baseline, not an eternal promise that future additions cannot change the count.

The suites cover synth/goldens, app glue, config, hostile protocol dispatch, generic callbacks, real socket timing/reconnection/pressure, and codecs. Null-backend tests never open real speakers or request microphone permission.

Real reference demo

bash demo/run_demo.sh

This builds an unmodified reference ninjamsrv and a headless reference client core. Six instrument intervals include five play bars and one rest. The receiver decodes real audio with measured energy. Silence must fail the negative control, and equal-input runs must produce equal upload hashes.

Reference main is resolved once to an exact SHA. NINJAM_REF_SHA selects an explicit pin; NINJAM_REPO can supply a clean matching checkout. External checkouts are not reset or deleted.

Fresh run evidence is written under demo/evidence/run-<timestamp>-<pid>. KUJ_EVIDENCE_DIR can choose another fresh path. receipt.json records phases and cache outcomes even when no demo runs.

The shared provisioner validates source/build fingerprints and artifact hashes before using warm state. Transient network retries are bounded. Compilation and demo assertion failures are not hidden behind generic retries.

Read the provisioning policy.

Backpressure is measured, not narrated

zig build repro-backpressure -Doptimize=ReleaseSafe

The full session runs on a constrained stream socket. Its peer stops reading for four seconds, keeps sending pings, then drains and decodes fresh uploads on the same connection. Assertions cover clock progress, counted drops, frame alignment, and non-silent recovery.

The recorded #14 reproduction ran 7.2 seconds, started ten bars, counted six drops, and decoded three recovered uploads. This is a bounded fixture, not a prediction of how many bars your network will drop.

Before/after evidence.

Fuzzing and CI

src/proto_fuzz.zig drives full sessions with a fixed corpus and seeded random server streams. Arithmetic consumers matter as much as parsers. The former hostile channel-ID and zero-tempo crashes are fixed and covered.

Coverage-guided zig build test --fuzz has documented Zig 0.16 runner/linking limitations; the ordinary corpus and seeded runs still execute in every test run. Do not claim a coverage-guided campaign merely because that flag exists.

CI runs native Linux/macOS build and test matrices plus a Windows build-test-windows leg (Debug and ReleaseSafe build/test, render and fart smokes), render/energy/silence smokes, and the macOS reference demo. build-test and demo-e2e are required on main. Docs have their own validation/publishing workflow; a docs build does not replace runtime evidence.