What does this PR do?

_find_tail_cut_by_tokens calls len(content) to estimate message tokens, but
when content is a multimodal list (e.g. [{"type": "text", "text": "..."}, {"type": "image_url", ...}]),
len() returns the number of blocks (~2) instead of total character count.
Every multimodal message is therefore estimated as ~10 tokens regardless of
actual size, causing tail protection to expand far beyond the token budget.

The fix sums text lengths across content blocks when content is a list —
mirroring the pattern already used correctly in the sibling method
_prune_old_tool_results (line 488).

Related Issue

Fixes #16087
Related: #14694 (anti-thrashing cascade), supersedes #14395 (closed without merge)

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✅ Tests (adding or improving test coverage)

Changes Made

• agent/context_compressor.py — _find_tail_cut_by_tokens(): replace
  len(content) with sum(len(p.get("text", "")) for p in content if isinstance(p, dict)) if isinstance(content, list) else len(content)
• tests/agent/test_context_compressor.py — added
  TestTokenBudgetTailProtection::test_multimodal_list_content_counted_correctly

How to Test

1. Run the new regression test:

   uv run pytest tests/agent/test_context_compressor.py::TestTokenBudgetTailProtection::test_multimodal_list_content_counted_correctly -v