chameleon 4.4.0

Added

• Per-call response_format ("concise" | "detailed", default detailed) on the
  three fattest read tools, per the tool-writing guidance's response-format
  pattern: search_codebase concise keeps name/file/line per row (-62%
  measured), describe_codebase concise keeps each archetype's
  name/size/witness + top 5 god symbols (-54%), and
  get_duplication_candidates concise names candidates without reading any
  body excerpt (-61%). An unknown value falls back to detailed with a note --
  a typo never fails a read.
• search_codebase pagination: offset pages the same deterministic ranking
  (clamped to COMPREHEND_SEARCH_MAX_OFFSET, default 500); while more matches
  remain the response carries next_offset and the truncation note names the
  exact re-call. A past-the-end page returns empty results, found stays true,
  and the note says the ranking ended (never the false "No symbol matched").
• Wire-size regression guard (tests/unit/test_wire_budgets.py): pinned
  ceilings (~1.5x measured) on fixture responses, the help outputs, every
  tool description vs the 2KB truncation ceiling, and the total tools/list
  schema -- an accidental pretty-print or payload-duplication regression now
  fails CI instead of silently re-inflating every session.

Fixed

• The unknown-response_format echo is bounded to 40 chars and sanitized
  before reflection into the response note (a model-supplied megastring or
  tag-boundary token no longer reaches the payload verbatim), and that note
  survives alongside the empty-result guidance instead of being overwritten.