fix(sprout-agent): cap tool-result text at 50 KiB with middle elision

[tlongwell-block]

Problem

A single MCP tool result can carry up to 8 MiB of text into agent history (MAX_TOOL_RESULT_BYTES, raised from 256 KiB in #602 to let view_image return real images). Tool results are re-sent on every request, so a couple of fat cats blow past the context budget, force a handoff, and the agent keeps only a lossy ~8k-token summary — it feels reset mid-task. Reported by Wes in #sprout-bugs; the investigation itself got handed off four times mid-research, which is all the repro anyone needs.

Fix

Split the budget at the existing chokepoint, tool_result_content() in mcp.rs — the single path every MCP tool result takes before entering history:

• Text capped at 50 KiB per result (new SPROUT_AGENT_MAX_TOOL_RESULT_TEXT_BYTES, validated 1 KiB..=8 MiB). 50 KiB matches the shell caps in sprout-dev-mcp, goose, and pi; codex's 10 KB is the aggressive end.
• Images keep the 8 MiB total budget dev-mcp: add view_image tool #602 needed — they're genuinely large and already accounted via estimated_bytes.
• Middle elision, not head-only: head and tail survive, with an inline marker reporting bytes elided. Build logs put the verdict at the end; head-only truncation is how an agent confidently retries a command whose failure reason was cut off. (Codex's strategy; pi/goose tail-keep makes the same point.)

Why here

• Not sprout-dev-mcp: its tools already self-limit (shell = 50 KiB + artifact spill). Fixing only the tools leaves every other MCP server able to flood history — goose has exactly this gap. Defense belongs in the harness.
• Not request-assembly or compaction-time: source-time truncation is deterministic and prompt-cache-friendly — what the model saw once is what it sees forever. Codex truncates at history-record time; results enter our history through this one path, so source-time alone covers it.

Outcome

Worst-case turn (64 tool calls) drops from "up to 512 MiB of history" to ≤3.2 MB of text. Context fills from work, not from one fat cat. Handoffs become rare instead of routine.

Verification

• cargo test -p sprout-agent — 110 passed, 0 failed (full suite, incl. 5 new tests: middle-elision, budget-untouched, image-exempt, UTF-8 boundary stress)
• cargo clippy -p sprout-agent --all-targets clean, cargo fmt --check clean
• Merges clean against origin/main