One-shot owner-scoped capture under a heap limit and an absolute deadline.

A snapshot owner reads a source once, at start, and must not grow its own heap doing so. The work therefore runs in a separate monitored process with :max_heap_size set to kill, and the calling owner waits on three things at once: the reply, the worker's death, and the death of the process the capture belongs to.

The outcome is reported in one vocabulary — :owner_down, :heap_exceeded, :worker_failed, :deadline_exceeded — and each caller maps it onto the codes its own source contract publishes. A capture's own return value is handed back wrapped in {:ok, …}, so a capture that answers {:error, …} is never confused with a failure of the capture machinery itself.

Whatever ends the wait, the worker does not outlive it: a deadline or a dead owner kills the worker and reaps its :DOWN before returning, so no capture keeps reading a source nobody is waiting for.