Summary

Related to #245. Draft qualification infrastructure, not a completed placement decision or a production offloading change.

• Exercise the existing private RoundIngestion and native StreamRelay in equivalent inline and bounded-worker placements without widening the core API or adding another parser, folder or emitter.
• Reuse named cassette response inputs and native HTTP/SSE framing; keep complete payload/wire gates separate from timing. Cover bounded entries/logical bytes, concurrent feed/drain, cancellation, disconnects, primary errors and observed joins.
• Add explicit opt-in matrix/resource/queue qualification with shared-origin replay schedules, independent replica windows and unknown quantiles preserved. HTTP measurement deadlines cancel and await their batch, including warmups.

Production ingestion remains inline. Please advise whether this crate-internal opt-in arrangement and review/integration split are the preferred boundary. The earlier #245 scope question remains open. This is a substantial test-only candidate, not a request to infer readiness from its size or test count.

Test Plan

Current local candidate tree: 3ca01a413ca062bddd72c00064935cb740d56c8f, based on 0b98f4baa8963e4e51d2ae816813266bc651fcf0.

cargo +1.98.1 test --locked -p agentic-server-core --lib -- --test-threads=2
cargo +1.98.1 clippy --locked -p agentic-server-core --lib --tests -- -D warnings
cargo +1.98.1 fmt --all -- --check

• Final core library: 1,172 passed, 0 failed; 18 tests remained explicitly ignored. Strict Clippy, workspace formatting and all applicable candidate-file pre-commit hooks passed.
• The HTTP teardown regression covers inline/worker owners. A controlled old-behavior mutation failed the joined-client assertion; the fix was restored and verified. This is mutation evidence, not a pristine upstream baseline comparison.
• Earlier source 82b56831... had completed full-output gates and separately retained inline/fresh-worker records. Those remain historical evidence on that source pin, not this new tree's performance results. The interrupted prefix is excluded; the available arms are cross-session, not one continuous causal experiment.

The new tree's opt-in release/full timing was not rerun. No real-model/GPU performance, complete gateway transport/tool-loop/storage benchmark, exact OS thread peak or placement recommendation is claimed. Input/output queue bounds are currently coupled, not a complete independent asymmetric sweep. These remaining measurement/decision boundaries must be agreed before #245 can be considered complete.