feat(agent-memory): Phase 7 — discovery marker

[jamby77]

Stacked on #254 (Phase 6 — consolidate).

What

Phase 7 of @betterdb/agent-memory: a discovery marker so BetterDB Monitor auto-discovers memory stores on any Valkey it watches, reusing the shared registry protocol from agent-cache.

• MemoryDiscovery registers a marker on __betterdb:caches keyed by store name: type: 'agent_memory', prefix, version, protocol_version, capabilities: ['recall','consolidate','reinforce'], stats_key: {name}:__mem_stats, plus started_at/pid/hostname.
• Heartbeats on an unref'd interval, refreshing a TTL'd __betterdb:heartbeat:{name} key and the marker; sets __betterdb:protocol with NX.
• Detects a name collision with a different cache type and throws; overwrites a same-type marker.
• MemoryStore gains an opt-in discovery option (registers on construct) and close() to stop the heartbeat and delete the marker.

Design / reuse notes

• The agent-cache DiscoveryManager uses a method-style Valkey client and hard-codes its cache type in the collision check, while MemoryStore is built on a .call-only client. Rather than refactor the merged manager (and its FakeClient tests), this reuses the protocol as the single source of truth — agent-cache now additively re-exports its discovery constants (REGISTRY_KEY, PROTOCOL_KEY, HEARTBEAT_KEY_PREFIX, TTL/interval, PROTOCOL_VERSION, MarkerMetadata) — and implements the register/heartbeat loop against .call, so Monitor reads the memory marker identically to a cache marker.
• Registration is fire-and-forget (construction stays sync); a rejected registration (collision) can't leak as an unhandled rejection, and close() awaits readiness before teardown.
• Best-effort writes (an onWriteFailed hook is in place for the Phase 9 observability wiring).

Tests

• discovery.test.ts (8): marker shape/capabilities/stats_key · protocol NX + heartbeat TTL · collision throws (and writes nothing) · same-type overwrite · stop deletes heartbeat · tickHeartbeat re-writes · interval fires (fake timers) · best-effort on write failure.
• MemoryStore.discovery.test.ts (2): opt-in registers on construct + close() tears down · disabled → no registry I/O.

@betterdb/agent-cache unchanged behaviorally (253/253 still green after the additive export). agent-memory: 65/65 green · tsc clean · prettier clean.

---

Note

Low Risk
Discovery is opt-in and additive; registry I/O is best-effort and mirrors existing agent-cache patterns without changing recall/remember paths.

Overview
Adds opt-in Valkey discovery for @betterdb/agent-memory so BetterDB Monitor can find memory stores the same way it finds agent caches.

MemoryDiscovery registers an agent_memory marker on the shared __betterdb:caches registry (field {name}:mem to avoid clashing with an agent-cache of the same name), sets protocol/heartbeat keys from the agent-cache protocol, runs an unref’d heartbeat loop, throws on cross-type name collision, and treats registry writes as best-effort.

MemoryStore accepts discovery?: boolean | MemoryDiscoveryConfig, starts registration asynchronously on construct, and adds close() to await registration and tear down the heartbeat key.

@betterdb/agent-cache only additively re-exports discovery constants/types (REGISTRY_KEY, PROTOCOL_KEY, heartbeat TTL/interval, MarkerMetadata, etc.); existing cache behavior is unchanged.

^{Reviewed by Cursor Bugbot for commit 73d91d7. Bugbot is set up for automated code reviews on this repo. Configure here.}