Wave B Evidence (Hard-Close)¶

This evidence pack records the verified closeout state for Wave B: Local Runtime Completeness.

Wave B is hard-closed.

Scope Covered In This Evidence Update¶

• tree-sitter JNI startup crash in local core
• local packaging completeness for bundled service binaries
• startup hardening for model-path handling in native service wrappers
• direct local runtime smoke for core, speech-engine, and code-engine

Commands And Results¶

1. Local bundle packaging¶

cd maestro
./gradlew client:installServer -x downloadModels

• Result: BUILD SUCCESSFUL
• What it proves: maestro/client/static/local is populated via the current packaging path.

2. Tree-sitter JNI symbol namespace verification¶

nm -D maestro/client/static/local/core/lib/libjava-tree-sitter.so | rg "Java_ai_(arqon_maestro|serenade)_treesitter"

• Result: only Java_ai_arqon_maestro_treesitter_* exports are present.
• What it proves: local core now loads an Arqon-namespace JNI library, not the legacy Serenade JNI symbol set.

3. Local runtime smoke (direct run-pro)¶

(cd maestro/client/static/local/core/bin && timeout 8 ./run-pro)
(cd maestro/client/static/local/speech-engine && timeout 8 ./run-pro)
(cd maestro/client/static/local/code-engine && timeout 8 ./run-pro)

• Result:
• core: exits 124 (timeout) without JVM crash.
• speech-engine: exits 124 (timeout) after normal bring-up.
• code-engine: exits 124 (timeout) after loading all model groups and starting server.
• What it proves:
• all three local engines survive startup smoke on the local bundle.

4. Service health verification (17200/17202/17203)¶

bash -lc '
ROOT=/home/irbsurfer/Projects/arqon/ArqonMaestro/maestro/client/static/local
pushd "$ROOT/core/bin" >/dev/null
./run-pro >/tmp/arqon-core.log 2>&1 & CORE_PID=$!
popd >/dev/null
pushd "$ROOT/speech-engine" >/dev/null
./run-pro >/tmp/arqon-speech.log 2>&1 & SPEECH_PID=$!
popd >/dev/null
pushd "$ROOT/code-engine" >/dev/null
./run-pro >/tmp/arqon-code.log 2>&1 & CODE_PID=$!
popd >/dev/null
sleep 18
curl -sS http://127.0.0.1:17200/api/status
curl -sS http://127.0.0.1:17202/api/status
curl -sS http://127.0.0.1:17203/api/status
kill $CORE_PID $SPEECH_PID $CODE_PID
'

• Result:
• core={"status":"ok","c":"","g":""}
• speech={"status":"ok","c":"","g":""}
• code={"status":"ok","c":"","g":""}
• What it proves:
• local stack is healthy concurrently, not just one engine at a time.

5. Code-engine protobuf/sentencepiece crash fix¶

# source updates verified in:
# - maestro/code-engine/server/src/token_encryption.cc
# - maestro/code-engine/server/include/token_encryption.h

• Result:
• removed direct in-process SentencePieceProcessor::LoadFromSerializedProto(...) usage from TokenIdConverter.
• TokenIdConverter sentencepiece path now uses spm_encode CLI (ARQON_MAESTRO_SPM_ENCODE / SERENADE_SPM_ENCODE or default binary paths).
• What it proves:
• the previous code-engine segfault path is eliminated while preserving sentencepiece token-id behavior for contextual models.

6. Legacy null-model-path crash hardening¶

# source updates verified in:
# - maestro/speech-engine/server/server/speech_engine_server.cc
# - maestro/code-engine/server/server/code_engine_server.cc
# - maestro/speech-engine/bin/run-pro
# - maestro/code-engine/bin/run-pro

• Result:
• model env vars are now validated in C++ startup with explicit error messages.
• run-pro wrappers now default to bundled model directories when no path arg is passed.
• What it proves: missing-model configuration now fails clearly instead of dereferencing null env pointers.

Current Wave B Verdict¶

Wave B is hard-closed.

Close criteria met:

1. local packaging succeeds (client:installServer -x downloadModels)
2. bundled local core, speech-engine, and code-engine start cleanly
3. all three service health endpoints return status=ok together
4. previously blocking code-engine startup segfault is resolved

Residual follow-on (outside Wave B close criteria):

• keep gathering local UI and voice-command e2e evidence as part of ongoing runtime regression coverage