Add byte-aware and per-use-case local-cache capacity controls

[lan17]

Priority

P1 — High — resolve before broad production adoption or large-scale use.

Problem

The local layer is one global LRU per DialCache instance, and every entry has weight 1 regardless of key/value size. Large object graphs can make memory effectively unbounded, while one high-cardinality use case can evict unrelated hot data.

Evidence:

• Single global LRU and entry weighting
• Every stored value has size 1
• README acknowledges entry-count semantics

The review's heap probe estimated roughly 259–270 bytes of LRU metadata per entry before key and value memory.

Open decisions

Choose an estimated-weight callback or another explicit model; decide whether per-use-case limits, reservations, or separate pools are supported. Do not claim exact JavaScript heap measurement.

Acceptance criteria

• Define behavior for an item larger than its budget.
• Add entry-count, estimated-weight, eviction-reason, and rejected-item telemetry.
• Make localMaxSize: 0 behave and report as disabled rather than as a permanent miss layer.
• Test large values, noisy-neighbor eviction, TTL expiry, and global/per-use-case interactions.
• Benchmark accounting overhead.