context-graph: track and analyse memory served to and used by the model

[antejavor]

Problem

Context Graph serves memory to the harness's model through read tools — sessions-graph's recall today, resources-graph's resource(address) as designed in Map: Resources component. We have no way to see what was served, whether the model actually used it, and whether it saved anything (e.g. a re-fetch avoided). Without that, we can't tell useful memory from noise, tune retrieval, or show the cache earns its keep.

Context from the Resources component design

• Every resource call records a Touch marked served from memory with an outcome: hit, subsumed (answered from a broader cached listing), or miss (What a Resource cache hit means).
• A miss followed by a real fetch (WebFetch / gh / MCP) is first contact; a hit followed by a real re-fetch of the same address means the model judged memory too stale — both are observable as Touch sequences.
• A PreToolUse nudge tells the model an address is in memory; whether the model then switches to resource is observable too.
• Touches are private to their user (HAD_SESSION); any analysis must respect that scoping.

Questions to answer

• What to record per serve: which tool, what was returned (Resource / memory rows), outcome, size.
• How to detect use: did the served content influence the next actions/answer (cited, acted on, ignored, overridden by a re-fetch)?
• Which signals are analysis (reporting-only) vs. inputs to retrieval/refresh behaviour.
• Where it lives: per-component, or a shared served-memory record across recall and resource.
• Relationship to the eval harness (context-graph/eval) — offline benchmark vs. live usage analytics.