feat(tools): spill tool output as a tail-only window

[Aaronontheweb]

What

Change tool-output spill bounding from a symmetric head+tail window to tail-only. When a tool result exceeds the inline budget, Netclaw keeps only the last budget chars inline and drops the head; the full result still spills to the session file, and the model continues through tool_output_read (opaque CallId + Start/Limit) to pull the dropped front back.

Why

The head+tail split is confusing and assumes useful content is evenly distributed. For command-like output (shell, logs, stack traces) the useful content is at the tail. Tail-only is simpler to reason about and matches how agents read command output. tool_output_read already recovers the dropped leading/middle content, so nothing is lost.

What changed

• BoundedOutputReader.Window: now returns Separator + last budget chars when over budget.
• BoundedOutputAccumulator: rewritten to a single tail ring; removed the _head buffer; Finish() still reports Truncated.
• ToolOutputSpill: doc/comment updates only (no behavior change in Compose; it still names the opaque CallId).
• Updated BoundedOutputReaderTests and ToolOutputSpillTests to assert tail-only behavior.

Verification

• Build: Netclaw.Actors and Netclaw.Actors.Tests succeed, 0 warnings.
• Tests: BoundedOutputReaderTests + ToolOutputSpillTests + ToolOutputSpillFreshSessionTests = 45 passed / 0 failed. ShellToolTests = 41 passed / 1 skipped (Windows host).
• dotnet slopwatch analyze: 0 issues.
• Copyright headers verified intact on changed files.