[Stephen Belanger (Qard)]

Shutdown previously drained ingestion while accepted HTTP requests and background login could continue running. Coordinate both domains under the first shutdown call’s deadline, close both admission paths immediately, and share the final shutdown outcome across concurrent callers.

• Track at most 1,024 response tasks with atomic admission and completion registration. Runtime loss and timeout resolve their tokens, retire tracking, and cancel pending work, including tasks not yet polled. Dropping an observer still leaves accepted work running.
• Reserve lifecycle control capacity independently of response admission. A write failure still waits for accepted HTTP work; a deadline cancels both domains. Individual HTTP operation errors remain on their tokens without a shutdown failure history. ShutdownReport.delivered continues to count trace writes.
• Track background login and publish completion notifications outside internal locks so observer wakeups and unobserved result destruction can reenter scheduling safely. Reuse the bounded shared outcome implementation for span finalization and shutdown.

Stacked on #127. Public resource methods still use the existing async adapters; converting them to eager completion methods follows in the next layer.

Validation: 310 unit/integration tests and 9 doctests passed (existing ignored tests unchanged); strict all-target/all-feature Clippy, rustdoc, formatting, and diff checks passed. Regression coverage includes simultaneous HTTP/ingestion timeout, background login, concurrent shutdown, response saturation, runtime loss before first poll, admission races, reentrant publication, and explicit runtime progress obligations.