[AdaAibaby]

First step toward #3763 (raising sandboxes-per-node). Observability only, off by default — nothing is reclaimed, paused, or shared here. It establishes the data floor that later density work (live balloon reclaim, same-template page sharing, pool-pressure auto-pause) needs to justify and tune itself.

Why

Guest RAM is faulted into the hugepage pool lazily (UFFD + hugetlbfs), so a node's real ceiling is the working set actually faulted in, not the RAM sandboxes nominally declare — and an idle sandbox keeps holding pool pages it no longer uses, because nothing reclaims them at runtime. Today there is no per-sandbox visibility into either number, so the density conversation is all estimates.

What

Two per-sandbox gauges, gated behind the sandbox-density-metrics flag (default false → no new series until an operator enables it):

• orchestrator.sandbox.mem.working_set — host bytes faulted into the pool for the sandbox since resume. Read from the UFFD handler's existing ServeStats().Bytes (no new fault-path instrumentation) via a new Sandbox.MemoryWorkingSetBytes() accessor. Upper bound on current pool residency.
• orchestrator.sandbox.mem.reclaimable_estimate — pool memory the node holds that the guest is not actively using: max(0, workingSet − guestInUse), where guestInUse is the envd-reported used RAM. Combined with free-page reporting: when the sandbox's balloon returns freed pages itself (BalloonModeReporting), the pool is self-reclaiming, so the estimate is reported as 0 rather than double-counting pages the device already discards.

The estimate feeds a later reclaim decision, so it is floored at zero and never over-reports. reclaimableBytes is a pure function with table tests: gap, floor-at-zero (never negative), no-guest-report upper bound, free-page-reporting → 0, zero footprint.

Safety / scope

• Off by default: zero behaviour change and zero added metric cardinality until enabled.
• No hot-path change: working-set comes from counters the UFFD handler already maintains; the metrics callback just reads them.
• Pure-function test + existing SandboxObserver regression suite pass; go build/go vet clean.

Refs #3763. This is PR 1 of a planned sequence; it intentionally ships only the measurement so the subsequent reclaim/sharing/tiering PRs can cite real per-sandbox data.