feat(cache): two-tier cache + per-node request coalescing

[java-jaydev]

Why

Market-data calls hit a rate-limited upstream. Bursts of identical requests (the common case when several agents ask for the same quote) each cost an upstream call, and slow-changing data was re-fetched as aggressively as live quotes.

What

A two-tier read-through cache in front of the Toss Open API, with per-node single-flight coalescing.

• L1 — Caffeine near-cache (in-process). get(key, loader) is atomic per key, so concurrent identical requests on a node collapse to one upstream load. L1 honors the caller's per-type TTL via Caffeine variable expiry, so a single node caches correctly with no Redis deployed.
• L2 — Redis shared cache (opt-in, TOSS_CACHE_L2=true). Adds cross-instance sharing. Any Redis error degrades to a cache miss and never breaks a tool call.
• Per-type TTLs: quotes/orderbook 2s, trades 3s, intraday candles 10s, daily candles 1h, stock info 6h.
• Cache keys normalize comma-separated symbols (trim + sort), so 005930,000660 and 000660,005930 share one entry.

Tools still return the raw upstream JSON string — the cache is transparent to the MCP contract.

Scope

L1 single-flight is per-node. Cross-node request coalescing is not implemented; the shared L2 narrows but does not eliminate the concurrent-miss window across instances. HTTP (Streamable) transport is deferred to Phase 2.5.

Testing

18 tests. RedisL2CacheIT uses Testcontainers (disabledWithoutDocker = true) and is skipped locally without a Docker daemon — this PR run is its first execution on a Docker-enabled runner.

Suite	Covers
MarketDataCacheTest	L1 hit/miss, single-flight
MarketDataCacheTtlTest	per-type TTL via variable expiry (time-controlled Ticker)
MarketDataServiceTest	per-type TTL selection, symbol normalization
RedisL2CacheDegradeTest	Redis failure degrades to miss
RedisL2CacheIT	real Redis round-trip + TTL expiry (Testcontainers)
CacheWiringTest	exactly one L2Cache bean across all three property modes