cli: lean tool catalog exists but is off by default, and MCP server schemas join the prefix unbounded

[macanderson]

Problem

Two compounding prefix-size defects:

1. crates/stella-cli/src/discovery.rs:130-138: lean_from_env gates the 13-tool lean core behind STELLA_LEAN_TOOLS; unset = full ~46-tool catalog (~6k+ tokens of descriptions alone before JSON envelope) at position 0 of every request. The activation machinery (tool_search, ActivatedTools) is built and tested — it is simply not the default.
2. crates/stella-mcp/src/toolset.rs:501-518: every enabled MCP server's tools are forwarded with description and input_schema uncapped, from a remote server. One chatty MCP server with 50–100 tools can exceed the whole recall+memory+rules budget combined; nothing bounds it.

Fix direction

• Measure lean-mode's activation rate on the bench corpus before flipping any default (each activation mutates the schema block = one prefix invalidation; the trade is ~70% smaller prefix vs ~4-5 invalidations/session).
• Cap MCP schema contribution per server (bytes), truncating descriptions with a note, and surface a boot warning when a server blows the cap.

Verify

Witness for the cap: a fake MCP client with a 100KB schema set yields a bounded serialized block plus warning. For lean: bench arm comparison (cache ratio + task success), not just token counts.

Constraints

Schema-block changes invalidate the tools cache tier once per change — batch with other prefix-affecting PRs where possible. Related: #1848 (schema sort), #1837/#1839 (breakpoints/TTL).

[macanderson]

Triage 2026-08-06: P1 — both halves compound at position 0 of every request, and the unbounded MCP half is the sharper defect: one chatty server can swamp the entire recall+memory+rules budget with nothing even warning. Plan: as specced, split in two — the per-server byte cap with truncation note and boot warning (witnessed by the 100KB fake-client schema set) is safe to land alone and is the next step; the lean-catalog default flip waits on the bench activation-rate measurement, since each activation is a prefix invalidation and the trade must be measured, not guessed. Batch schema-block changes with #1848's sort fix to pay the one-time cache invalidation once.

[macanderson]

Triage note — half of this shipped; the other half has not.

Shipped (PR #1898, 273ecf4c): the unbounded MCP schema half. crates/stella-mcp/src/toolset.rs:86 defines MAX_SERVER_SCHEMA_BYTES = 32 * 1024, budget_segment (:546-562) truncates one server's contribution, and over_budget_servers() (:585) reports what was cut rather than dropping it silently.

Not shipped: the lean catalog is still off by default. crates/stella-cli/src/discovery.rs:130-138 returns None (= full catalog) when STELLA_LEAN_TOOLS is unset, and :193 still wires lean: lean_from_env().

So this issue should stay open, scoped to the default only. A fresh agent picking it up needs to touch discovery.rs and nothing in stella-mcp.