optional platform.memory backend — durability + cross-team continuation

[haasonsaas]

shared-memory-mcp currently solves coordination tax for agentic teams in-process — claim-based work distribution, 10:1 compression, dependency tracking, stale detection. Great until the process dies or a second team needs the same memories.

`evalops/platform`'s memory service (`internal/memory/memories/`) is the durable, multi-scope memory store that every other consumer in the org reads/writes. It supports workspace / user / team / project / agent scopes with typed retention and proto-typed records.

Ask

Add an optional backend:

```
SHARED_MEMORY_BACKEND=platform
SHARED_MEMORY_PLATFORM_BASE=https://platform.internal/...
SHARED_MEMORY_PLATFORM_TOKEN=
```

When set, claim / status / work-state operations persist to `memory.v1.MemoryService` (`Store`, `Recall`, `Update`, `Delete`) instead of the in-memory map. Falls back to the current behavior when unset.

Payoff

• Durability (process crash = no loss)
• Cross-team continuation (team A hands off to team B with context intact)
• Ranking inherits the recency + confidence + embedding blend landed in `evalops/platform#601` / `#603`
• Usage shows up in `evalops/console#15` (attribution panel) once that lands

Related

• `evalops/platform/internal/memory/` — the target backend
• `memory: opt-in sync of chat threads, meeting summaries, journal to platform kestrel#15` — same ask, different consumer; shared SDK would serve both

Scope

~1 week. Add the backend interface, a platform-client adapter, and tests that verify both paths satisfy the same contract.