Instrument KV cache tiers and enforce matched performance promotion gates

[i386]

Context

PR #1632 adds an exact node-local durable L3 cache, but every optimization added to that work needs a matched before/after gate. This issue owns the measurement contract for the expanded KV-tier work; it complements the broader cache telemetry in #1097.

The current L3 implementation can produce roughly 9,500 geometry-aligned segment references for a realistic 19K-token prefix. Storage layout, copying, I/O scheduling, admission, and tier routing therefore need measurements that expose where time and bytes are actually spent.

Coordinated in Buzz channel #skippy-kv.

Work

• Add phase timings for lookup, export, encode, queue, write, read, verify, decode, import, and commit.
• Count logical bytes, unique physical bytes, bytes copied, bytes allocated, bytes read/written, file operations, and peak transient memory.
• Record L1/L2/L3 hit, partial-hit, miss, refusal, admission, promotion, demotion, eviction, and cold-fallback reasons.
• Capture queue depth, read-ahead depth, write-behind depth, compaction debt, and time blocked by cache maintenance.
• Report restored tokens, suffix-prefill tokens, TTFT, TPOT/decode-event p50/p95/p99, throughput/goodput, CPU time, and device utilization.
• Extend the release harness with identical cold/current/candidate A/B workloads across representative prefix lengths, growth patterns, repetition rates, cache pressure, and concurrency.
• Run the discrete CUDA/non-unified-memory cell as well as unified-memory Metal and every supported release backend/model-family cell required by feat(skippy): add configurable node-local L3 KV cache #1632.
• Hash inputs, binaries, configuration, order, and outputs so comparisons are attributable to an exact build and workload.

Promotion gate

• Exact deterministic output identity and corruption/fallback tests remain green.
• A candidate must beat the current feat(skippy): add configurable node-local L3 KV cache #1632 implementation on its stated primary metric under an identical workload; storage-size improvement alone is not sufficient.
• No candidate may regress the existing disk-off p99 decode-event latency gate beyond 5%.
• Publish raw artifacts and a concise comparison table, including negative results.

Relationship to other issues

This is the common evidence gate for the packed L3, streaming import/export, value-aware policy, L2 RAM, codec, and non-prefix experiments. It does not itself choose or implement those mechanisms.