Make DialCacheConfig modes discriminated and explicit #19

Priority

P2 — Medium — completed by the observability and Redis-ownership simplifications.

Original problem

Earlier versions exposed overlapping constructor modes: Redis accepted either a direct client or lazy factory, while core metrics mixed disabled/custom modes with Prometheus-specific configuration.

Resolution

The ambiguous modes no longer exist:

• feat: decouple Prometheus metrics adapter #56 reduced core observability to one optional caller-supplied DialCacheMetricsAdapter. Omission disables instrumentation; vendor configuration lives in dialcache/prometheus and dialcache/datadog.
• feat: require caller-owned Redis clients #67 removed RedisConfig.createClient and RedisClientFactory. RedisConfig.client is now required and caller-owned.
• Plain JavaScript/untyped callers receive explicit runtime migration errors for the removed Redis factory and key-prefix modes, where silently changing resource ownership or cache identity would be unsafe.
• TypeScript and packed ESM/CommonJS declaration consumers reject the removed metrics options and assert the direct-client requirement, backend-neutral metrics shape, and absence of vendor factories from the root package.

Current evidence:

• Final core config shape
• Final Redis config and runtime guards
• Packed declaration contract
• Migration documentation

No further discriminated-union layer is needed: the remaining constructor properties are orthogonal optional capabilities, and Redis itself has one unambiguous required client mode.