Document the remaining WorkInterval experiment against current host scheduling

docs/rfc/193-work-interval.md

@@ -0,0 +1,57 @@
+# RFC 193: a scoped host scheduling experiment
+
+Related issue: [#193](https://github.com/incoai/splash/issues/193).
+Rechecked against main `1482f6bbc2abf12375332725302019594150b6cb`.
+
+## Current runtime and remaining proposal
+
+Main already reduces per-submit host work and batches buffer bindings.
+[`MetalBackend`](../../runtime/metal/MetalBackend.mm) has asynchronous command
+completion, host steps and a host-step worker that requests
+`QOS_CLASS_USER_INTERACTIVE`. Its worker QoS request is distinct from an
+engine-thread work interval around graph construction/submission. No
+WorkInterval integration or `SPLASH_ENGINE_WORK_INTERVAL` option is implemented.
+
+[Maintainer feedback](https://github.com/incoai/splash/issues/193#issuecomment-5975595014)
+keeps the work-interval proposal open while asking for measurement after the
+existing host improvements. Main has subsequently changed submission, host
+steps, streaming and decode scheduling further. Earlier third-party timings
+and the old RFC's host-gap assumptions are not Splash measurements of this
+revision. A scheduling hint does not establish a P-core placement guarantee.
+
+The unused flag/configuration header and tests that only asserted its strings
+are removed from this documentation proposal. They did not call a scheduling
+API, wrap an engine scope or demonstrate a performance improvement.
+
+## Experiment design
+
+First measure graph preparation, submission, completion and the interval
+between commands with matching current builds. Keep GPU command spans and
+host scopes separate; subtracting GPU time from wall time does not isolate
+pure scheduler cost. Existing
+[`decode-profile`](../../dev/benchmarks/decode_profile.mm) supports native
+operator/cycle diagnostics, and the [engine](../../runtime/engine/Engine.cpp)
+and [scheduler](../../runtime/engine/Scheduler.cpp) define the serving loop.
+
+If a material host scheduling delay remains, an opt-in implementation can
+compare a bounded hint around engine preparation/submission against the same
+code without the hint. API choice, supported OS behavior, ownership and
+cleanup remain design work. Include success, exception, cancellation and
+shutdown paths; do not retain a hint over blocking waits. Keep the native
+protocol, model geometry, memory plan and kernel code unchanged for the
+experiment.
+
+Use matched warmups and alternating controls, report B1 through B4 and full
+serving latency/throughput, and record power/thermal conditions. Include
+competing UI/audio work and both hint setup/teardown cost and host/GPU gaps.
+A host-phase improvement alone is insufficient to change a shared default.
+Output hashes should be reported; identical arithmetic kernels do not prove
+identical mixed-serving batches or output trajectories.
+
+## Review and validation scope
+
+This is a documentation-only RFC. There is no implementation, active flag,
+new timing result or speedup claim in this PR. The current production path
+and defaults are unchanged. The proposal remains experimental until an
+implementation and measurements justify it; it is independent of main's
+adaptive MoE verify-row policy in #487.