Bound every MCP response: short lists, paging and partial item reads #302

Only memory_recall (maxChars), item_history (limit) and relation_list (depth, limit) let the caller size a response. The rest are fixed or unbounded:

• task_list has no limit (internally 300 root tasks) and returns every full description. On 2026-09-17 it returned 266,000 characters for this project, too large for an agent's tool output.
• item_search returns fixed 30 memories, 20 pages and 20 tasks without saying whether more exist.
• item_get always returns the whole item; pages can hold 100,000 characters.

Principles

• Lists carry no details. A list returns what identifies, sorts and filters an entry; item_get returns the rest.
• Lists page with one shared contract: limit with a small default, cursor, and total plus hasMore in the response. Explained once in the system guide instead of on every tool.
• Single items read in parts: offset and take in characters with a high default, the server reporting where to continue.
• maxChars stays where sections share one budget, which is recall only (#215).

Constraints

• The tool list is 20 characters under its budget in McpDiscoveryTests (18,996 of 19,016). New parameters need shorter descriptions elsewhere or a deliberate, recorded raise of the budget.
• Visibility: totals, cursors and outlines never reveal items or text the caller may not read.

Related: #219 (complete memory inventory, uses the list contract below), #215 (recall notices, done), #291/#295 (bag children page with the same contract).