perf(observability): add prefix-cache telemetry metrics

[wsp1911]

Supersedes #907.

Credit to @harryfan1985 for the original work in #907. This replacement PR keeps and adapts the parts of that work that remain valid on top of current main, while leaving context compression aligned with 5e5738f, which already implemented prompt-cache-aware context compression on main via the full-prefix compression path.

What changed:

• add cache write token tracking for prefix-cache observability
• refine cache hit ratio accounting to use only input tokens from requests where cache telemetry was explicitly reported
• document the prefix-cache stability contract for contextual tool manifests

What is intentionally not included:

• dc992d6, which is already on main
• the context compression implementation from 3a5033c / 3acb3fb, because 5e5738f already implemented prompt-cache-aware compression on the current main branch and this PR's compression path takes a different approach

Commits:

• 3a312eeb perf(observability): add cache write token metrics
• 70378754 fix(token-usage): refine cache hit ratio denominator
• 2165bfa7 docs(agent-tools): document prefix-cache stability contract