Struct MockCombinedRuntime 

pub struct MockCombinedRuntime {
    pub available: Option<u64>,
    pub total: Option<u64>,
    pub pending: Option<u32>,
    pub capacity: u32,
}

A test-only runtime handle carrying BOTH a VRAM signal (available/total, read by the VRAM guard) AND a live-load signal (pending/capacity, read through BackendStreams) — the exact shape of the production DeviceRuntimeHandle, which answers both would_fit and pending_work_count off one handle. Used by the C2 headline-gate test to build a 2-device branched graph whose arms are both VRAM-fit but differ in live load.

Fields

available: Option<u64>

VRAM available bytes (the guard’s available_bytes).

total: Option<u64>

VRAM total bytes (the guard’s total_bytes).

pending: Option<u32>

Live in-flight count (pending_work_count).

capacity: u32

Advertised slot capacity (slot_capacity).

Trait Implementations

impl BackendRuntime for MockCombinedRuntime

fn available_bytes(&self) -> Option<u64>

Bytes currently available for new allocations on this backend’s device. None when the backend genuinely cannot measure (no OS query exposed, no vendor API supports it).

fn total_bytes(&self) -> Option<u64>

Total memory on this backend’s device. Static after first call; implementations cache unconditionally. None for backends with unbounded notional capacity (e.g. Reference returns Some(u64::MAX), or None if the backend prefers to advertise “unknowable”).

fn as_backend_streams(&self) -> Option<&dyn BackendStreams>

Tier-2 upcast seam: a runtime handle that also exposes a deferred-execution queue model returns Some(self) so a selector holding only a &dyn BackendRuntime (the type the route picker’s BackendRuntimeLookup hands out) can reach the BackendStreams live-load surface without naming the concrete handle type.

fn would_fit(&self, size: u64) -> FitStatus

Predictive fit-check: would an allocation of size bytes likely succeed given current state? Returns FitStatus::Tight when projected post-allocation usage crosses the pressure threshold (default 0.85 of total).

impl BackendStreams for MockCombinedRuntime

fn pending_work_count(&self) -> Option<u32>

Number of slots (streams / queues / command buffers) currently busy with submitted-but-not-yet-finished work, for this handle’s device. None when the backend dispatches synchronously and has no queue concept (selectors then have no load signal — fall back to static cost, never treat None as “idle”).

fn slot_capacity(&self) -> u32

Maximum concurrent in-flight work this backend advertises. Used by the runtime route picker for load tiering (count / slot_capacity) and dispatch lookahead sizing.

fn flush(&self) -> Result<()>

Block until all submitted work on this backend’s slots has completed. Used at realize boundaries and by training-loop barriers (gradient accumulation, optimizer step).