[SaladDay]

What users see

A new hosted Session no longer waits for an idle allocation to suspend when its node is full. Core keeps one active slot free per node: an allocation idle for at least 2 minutes on a node with no free active slot but a free retained slot may be suspended without waiting demand. The suspended Session's next Turn resumes it (about 1.9 s). The idle timeout (idle_seconds, default 300 s) is unchanged; if it is shorter than 2 minutes the ordinary timeout already applies.

Changes

• SetCompute also allows running to quiescing when the allocation has been idle for 2 minutes (freeSlotIdle), its node passes the same deployment and node checks as demand suspension (nodes mode, not resetting, same installation, online, Active >= MaxActive), the node has a free retained slot (a suspended allocation keeps its retained slot, so without one the freed active slot admits no Session), and no reclamation is in flight on that node.
• The shared checks moved into fullNode; SetCompute loads the suspension demand once for both rules, and the demand path is otherwise unchanged. Busy and wake-requested allocations are still refused.
• docs/sandbox-provider.md and its Chinese mirror describe the rule in the Suspension paragraph and the Timeouts and limits table.

Measured

Before this change a new Session on a full node waited for a suspend of 5.5–7.8 s (median 6.2 s over 15 runs). Resuming the suspended Session takes about 1.9 s.

Tests

TestPressureSuspensionRequiresUsefulCapacity gains a full node idle past 2 minutes (allowed) a node with a free slot at the same idle (ErrNotIdle) and a full node whose retained slots are full (ErrNotIdle); the existing 1-minute no-demand case still returns ErrNotIdle. Ran go test ./internal/persistence/postgres/deploymentpg -run TestPressureSuspension and go vet ./... for services/core.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}