[Stephen Belanger (Qard)]

Experiment events previously registered first and queued their trace later, only after the caller polled the async method. A shutdown between those steps could close delivery admission before the event reached the queue. Make Experiment.log admit its write immediately and resolve registration during bounded delivery preparation; its Completion returns the row ID after full batch delivery. Experiment.flush now captures a cutoff immediately and returns the ordinary flush report token.

• Reserve response capacity before admitting the dependent write. Response overload creates no write obligation; queue overload retires the response reservation and drops the registration future unpolled. Dropping an observer leaves accepted work running.
• Resolve an immutable experiment destination inside the existing preparation pipeline. Registration errors fail the logical batch and covering flush, then retire with that batch; fresh events can retry registration without a failure ledger. Keep remaining async experiment methods pinned to the same execution binding while their conversion follows.
• Reuse span event normalization for the complete event snapshot. Return typed admission/execution errors when adapting delivery outcomes to BraintrustError, preserving API preparation causes.
• Detach canceled commands and barriers before notifying/dropping them outside service locks. Deferred commands may own the last queue producer, whose destructor reenters the service. Rejected rows with unresolved destinations log a diagnostic rather than initiating registration to produce a debug dump.

Stacked on #130. Other experiment methods and client span updates remain subsequent layers.

Validation: 327 unit/integration tests and 9 doctests passed; existing ignored tests unchanged. Strict all-target/all-feature Clippy, rustdoc, formatting, and diff checks passed. New regressions cover plain-thread unobserved progress after owner drop, shutdown before first poll, terminal registration failure and fresh retry, both admission limits without extra writes/registration, deferred-registration timeout, reentrant command cleanup, and stopped runtimes holding the last queue producer.