Add Redis write-path and end-to-end cache outcome telemetry

[lan17]

Priority

P2 — Medium — important hardening or API work after the safety foundations.

Problem

Current metrics time gets, fallbacks, and serialization, and observe serialized size. Redis writes have no client latency timer; watermark-suppressed writes have no dedicated outcome; and there is no full-wrapper duration/final-outcome signal.

Evidence:

• Current metrics adapter surface
• Write path without operational timing/outcome telemetry
• Suppressed write result is folded into control flow

Acceptance criteria

• Measure Redis write latency and stable outcomes: written, watermark-suppressed, skipped, and failed.
• Add end-to-end duration/outcome for local hit, remote hit, fallback, uncached pass-through, and error.
• Keep labels bounded and exclude IDs, arguments, and raw Redis keys.
• Extend custom metrics adapters compatibly, using optional methods or an approved versioned migration.
• Test every outcome plus metrics-adapter failure.
• Include instrumentation overhead in the performance budget.