[BUG] Bound MCP tool results before adding them to model history

[zoomote]

Problem

MCP servers can return arbitrarily large text payloads. Zoo Code currently allows those payloads to enter model-facing API history unchanged, so a single Git, browser, or other MCP result can exceed the context window and make condensation fail. This is the original failure reported in #648.

Scope

Bound MCP text results before they are added to model-facing history.

• Apply a provider-agnostic model-facing size/token ceiling to MCP string and text-block results.
• Preserve tool_use_id, error state, and valid tool call/result ordering.
• Leave image and other non-text content behavior unchanged unless required to preserve a valid result shape.
• Add a clear truncation marker containing the original and retained size.
• Keep this change independent of future full-output storage or MCP resource-link support.

Acceptance criteria

• A multi-megabyte MCP text result is bounded before the next model request.
• Mixed MCP content retains valid non-text blocks and bounded text blocks.
• Truncation preserves tool identifiers and provider-valid tool pairing.
• Error results remain errors after truncation.
• Focused tests cover string, structured text-block, mixed-content, and error results.

Out of scope

• Native read_file budgeting
• Condensation-input sanitization
• Sliding-window recovery behavior
• Full-output spill files or resource references
• Aggregate budgeting across parallel tool calls

Estimate

1 story point. Independently mergeable and non-stacked.

Parent tracking issue: #648