Add Sandbox diagnostics and system observability plan

[sam2tom]

Summary

• Add allocation-aware Sandbox diagnostics to the Core Web Dashboard: observed coverage, unexpected sampling gaps, parked allocations, memory pressure, and top measured targets.
• Preserve missing Runtime, CPU, memory, lifecycle, and token chart values as gaps rather than plotting zero. Conflicting same-second observations of one allocation are counted as an ambiguous sample gap.
• Document the existing Runtime/Session telemetry inventory and a bounded next-phase collection design for request, terminal Turn, actual model, tool, Sandbox, and collector-health signals. This PR adds no new backend metrics API.

Validation

• CI=true make check-claude-sdk check-web check-mcode-harness passed: Web 783 tests, client 409 tests, and 147 fixture browser cases, including wide/narrow Dashboard rendering.
• Dashboard focused tests and TypeScript typecheck passed.
• Independent blind review found two issues (same-second allocation ties and zero memory denominator); both were fixed and covered by tests.
• make check reached Agents API store tests with a disposable local PostgreSQL 16 database. One unrelated E2B helper timeout test failed under the parallel suite and passed when rerun alone.
• check-agents-executor cannot compile its Linux-specific Rust filesystem code on this macOS host (OFlags::PATH and st_dev platform types). The full aggregate gate is therefore not claimed green.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

[sam2tom]

Superseded by #98. The Sandbox diagnostics commit is already included in #98, and future Dashboard observability work will continue there.