Keep node links from summarised turns grounded, so a long chat's correct link is not replaced with a request to confirm #3508

Problem

In a chat long enough to be summarised, a correct link to a node that an earlier turn looked up or created is replaced with "I'd like to help with that. Could you confirm what you'd like me to do?".

The fabricated-id check in packages/agent/src/local_agent/agent_loop.rs accepts a nodespace:// id in a reply only if a tool produced it in this session: this turn's tool results, tool messages in history, or the system's own records of earlier turns' lookups and writes (grounded_node_uris_from_history). When maybe_summarize_history drains old turns into one [Conversation summary] system message, those tool messages and records go with them. The summary is the model's text, so it is deliberately not read for grounding (CONVERSATION_SUMMARY_PREFIX). Every id that only those drained turns produced is then treated as invented.

The check errs in the safe direction, a real link refused rather than a dead one let through, and the cost is recorded in the comment on grounded_node_uris_from_history and in architecture/prompt-skill-system.md. It is still a wrong answer to the user, in exactly the long chats where earlier records are most likely to be referred back to.

Fix

When maybe_summarize_history drains messages, collect the nodespace:// ids grounded by the drained Tool messages and system-written System messages (the same rule grounded_node_uris_from_history applies, so the summary itself contributes nothing). Keep them in a separate, system-written message beside the summary, for example "Nodes referenced earlier in this conversation: …", which the grounding rule reads like any other system record. Ids the user typed or the model wrote stay ungrounded.

A bounded list (the most recent N) keeps a very long chat's prompt in check.

Acceptance criteria

• After summarisation, an id that a drained tool result or system record produced is still grounded.
• An id that appeared only in drained user or assistant text, or only in the summary, is still not grounded.
• The carried-forward list is written by the system, bounded, and survives a second summarisation.
• Unit tests cover both directions, and each fails without the change.
• The comment on grounded_node_uris_from_history and the paragraph in architecture/prompt-skill-system.md no longer describe this as an accepted cost.