[SaladDay]

What users see

Faster cold start of a hosted Session on a full node. No behavior or API change.

Changes

• A placement that failed because nodes are full (placement.ErrNodeUnavailable) now hints every existing node lane, so demand suspension and checkpoint expiry start immediately instead of on the lane's next 5 s tick.
• A freed node slot (compute saved as "suspended", ReleaseAllocation, ReleaseLostCompute) wakes the manager's placement loop; direct lanes free no node slot and send nothing, instead of waiting for its next 5 s tick.
• Completing Environment initialization wakes the Worker scheduler instead of waiting for the next 250 ms scan.
• Tickers stay as the recovery path; hints only accelerate.

Measured

Test cluster, 5 Sessions, node max_active 3 (Core logs + node journal):

• Placement failing for capacity to the node lane starting suspension or release: 1.4-4.0 s wasted.
• Suspension or release finishing to the next placement attempt: 2.2-4.9 s wasted.
• Initialization completing to the pending input being picked up: up to one 250 ms scan.

Tests

• TestCapacityFailedPlacementHintsNodeLanes: a capacity-failed placement hints the node lane.
• TestFreedSlotWakesPlacement: a freed slot wakes placement and coalesces.
• go vet ./... (services/core) and go test ./internal/execution -run 'Placement|Wake|Suspend|FreedSlot|Capacity' pass.

---

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