feat(economy): compress Agent/Task prompt at PreToolUse time

[claudioemmanuel]

Summary

• New squeez compress-prompt subcommand reads stdin, runs markdown compression via the existing compress_md pipeline (Ultra mode), and head-truncates with a notice if still over budget; echos unchanged when under the token budget
• New agent_prompt_max_tokens config field (default 2000, 0 = disabled) with parser arm in Config::from_str and entry in the default config.ini written by squeez setup
• hooks/pretooluse.sh: new Agent/Task branch pipes tool_input.prompt through squeez compress-prompt and replaces it only when the output is strictly shorter than the input
• src/hosts/claude_code.rs: PreToolUse matcher extended from "Bash" to "Bash|Read|Grep|Glob|Agent|Task" so the new branch fires
• tests/test_compress_prompt.rs: integration test spawns the debug binary; asserts short prompt echoes verbatim, large markdown prompt is shorter than input, and key headings survive

Motivation

In observed production sessions a single plan.md (1,754 lines, ~7,000+ tokens) was injected verbatim into every Agent/Task spawn. With multiple agent invocations per session this compounds quickly. squeez already compresses Bash output and Read/Grep/Glob inputs; this PR closes the gap for the prompt field, which is the single largest uncompressed input vector for multi-agent workflows.

Test plan

• cargo test — all tests pass (162 unit + 3 new integration in test_compress_prompt.rs)
• cargo build --release — binary builds clean, zero-dep constraint preserved
• Manual: echo "# Plan\n\nshort" | squeez compress-prompt — echoes unchanged
• Manual: pipe a 1500-line markdown file — output is shorter, heading structure preserved
• After squeez setup: verify ~/.claude/settings.json PreToolUse matcher includes Agent|Task
• Verify ~/.claude/squeez/config.ini contains agent_prompt_max_tokens = 2000