[brenorb]

Regular tests that execute a program directly can leave requested verbose output buffered when the test returns, catches an execution failure, or panics. Those paths bypass the Signer's log flush.

Flush the remaining buffers when the regular test context is dropped, before existing node cleanup. Output already flushed by the Signer or explicitly by the caller stays cleared and isn't printed again. SDK execution and fee-estimation buffering are unchanged: direct SDK callers still own per-execution flushing, and a later execution can replace an earlier buffer for the same input. This change doesn't promise a complete execution history.

Fixes #170.

Validation:

• On upstream a1bf060, node-free probes reproduce missing traces for direct successful and failed execution, including after a regular context is dropped. Explicit flushing prints the buffered DBG/jet output.
• Three new subprocess-based tests fail before the fix while the quiet/manual-flush/Signer controls pass. All six integration tests pass afterward, covering Debug and Trace success, caught failure, unwinding, default quiet behavior, explicit flush, and actual Signer fee estimation without duplicate output.
• cargo test --locked --offline -p smplx-sdk -p smplx-test: 104 tests pass. One ignored helper is intentionally invoked in fresh subprocesses so each scenario owns its GlobalConfig and its stderr can be checked.
• All cases use synthetic transactions and an unused local Esplora endpoint; no node or broadcast is required. Extra quiet-mode failure/panic probes emit no execution trace.
• cargo clippy --locked --offline -p smplx-sdk -p smplx-test --all-targets -- -D warnings and cargo fmt --all --check pass.
• Independent agent review found no blocking issue. Existing SDK documentation work in docs(sdk): document programs, verbosity and utilities in rustdoc #194 remains separate.

Written by Codex.

[Arvolear]

LGTM!