v0.0.9

Other

• Merge pull request #109 from antiguru/claude/issue-101-budget-approach-LF16A
• refresh stale header comment on renamed call_tree test
• address second-pass review nits
• address review feedback (#109)
• count serialized bytes via io::Write sink, no allocation
• linear-time trim loop via per-drop bytes_freed
• apply rustfmt to call_tree.rs
• trim oversize tool responses to fit POLLARD_MAX_OUTPUT_BYTES (#101)

Added

• output budget: top_functions, top_groups, compare_profiles,
  stacks_containing, folded_stacks, and call_tree now trim their
  responses to fit POLLARD_MAX_OUTPUT_BYTES (default ~25 KB) so the
  caller harness's MCP token cap stops dropping payloads. Trimmed
  responses surface a truncated: { dropped, dropped_pct, budget_bytes, final_bytes, still_over_budget } field; rows are popped tail-first
  and call_tree rolls dropped leaves into the parent's Omitted
  summary.

Changed

• *_pct columns serialize at one decimal and *_ms at two
  decimals — the LLM never needed full f32/f64 precision and the
  shorter rendering frees up budget for more rows.