fix(hooks): skip tool_result messages when counting human exchanges

[arnoldwender]

Summary

Fixes #549 — exchange count was inflated by counting tool_result messages as human turns.

In Claude Code JSONL, every tool response (Bash, Read, Grep, subagent results) is emitted with role: "user" and content as a list of {"type": "tool_result", ...} blocks. Both counting paths treated these as real human messages, so a session with 10 real turns but 50 tool calls was reported as 60 exchanges — causing premature save-hook triggers at the wrong frequency.

Changes

• mempalace/hooks_cli.py _count_human_messages(): In the elif isinstance(content, list): branch, added an early continue that fires when all blocks in the list have "type": "tool_result". The guard is placed before the existing <command-message> check, leaving that logic untouched.
• hooks/mempal_save_hook.sh: Added the identical tool_result guard to the inline Python script (also before the <command-message> check).

Both changes are purely additive — no existing counting logic was modified, only a new continue path inserted for tool_result-only content lists.

Test plan

• All 32 existing tests/test_hooks_cli.py tests pass (verified locally)
• A new test test_count_skips_tool_result_messages should be added to directly exercise the new guard — noted for follow-up
• Manual: create a JSONL transcript with 10 real human messages plus 50 tool_result entries; confirm count returns 10, not 60