Batch/throttle graph extraction instead of triggering an LLM call on every Stop hook

[LIOsDev]

mem::graph-extract (and the LLM-backed provider.compress() call it makes) is only wired to event::session::stopped, which the Claude Code plugin fires via the Stop hook at the end of every assistant turn — not just when a logical session truly ends. With GRAPH_EXTRACTION_ENABLED=true, this means an Anthropic API call fires every few minutes during active use, even though the graph itself is merged incrementally and cheap.

Could graph extraction support a batched/throttled mode — e.g. a debounce/interval (only extract if N minutes have passed since the last extraction for a session) or an opt-in nightly cron-style batch via the existing /agentmemory/graph/build endpoint — so users don't pay one LLM call per turn? Right now the only way to avoid the per-turn cost is to disable graph extraction entirely, which seems like an unnecessarily blunt tradeoff between cost and graph coverage.

Noticed GRAPH_EXTRACTION_BATCH_SIZE already exists in config.ts but isn't read anywhere in the codebase — might be a leftover hook for exactly this kind of batching.

[LIOsDev]

Filed #985 proposing a BatchAnthropicProvider (Anthropic Message Batches API) for background LLM routines. If it lands, this issue's per-turn graph-extraction cost could be solved by routing mem::graph-extract's provider.compress() call through batch mode instead of (or alongside) throttling — every Stop-hook-triggered extraction still fires, but cheaply, queued into the next scheduled batch flush rather than firing one paid call per turn.