Chameleon

A from-zero whole-plugin QA round for a production release. Every supported
language and framework was bootstrapped fresh on the current engine (all
healthy, no degraded profiles), the upgrade path was re-verified (a schema-stale
profile re-derives to current on refresh), and the tool, hook, and daemon
surface was driven through real usage across the matrix. Six bugs surfaced and
were fixed, each pinned red-green. The highest-impact one was caught by
chameleon flagging its own large source files.

Fixed

• Spurious "removed-export" violations on any file over 100KB. The
  PostToolUse verify hook caps an oversized file to its first 100KB before
  handing it to lint_file, but that truncation fact was lost at the boundary:
  lint_file inferred truncation from the length of the content it received,
  which already sat at the cap, so it read the prefix as the whole file and
  every export defined past the cap looked removed. Editing a 486KB module
  produced 27 false "you broke N importers" violations. The hook now derives
  truncation from the file bytes it read and threads content_truncated through
  the daemon dispatch into lint_file, which skips the removed-export check on a
  known prefix. A genuine removed export in a file under the cap is still
  flagged. (tools.py, hook_helper.py, daemon.py)

• posttool_verify crashed on a non-string tool_name. A malformed payload
  whose tool_name was a list or dict is unhashable, so the tool_name in _EDIT_TOOLS membership check raised TypeError before the existing file_path
  guard could run. The hook still failed open at its outer wrapper but logged a
  traceback that /chameleon-doctor then surfaced as a spurious degraded signal.
  It now guards the type the same way the adjacent file_path check does.
  (hook_helper.py)

• Production-branch detection claimed origin backing on a repo with no
  remote. git remote remove (and a pruned remote default) leaves the
  origin/HEAD symref behind, dangling: git symbolic-ref still resolves the
  branch name even though the tracking ref is gone and no remote is configured.
  detect_production_branch trusted that stale name and reported
  source=origin_head / from_origin=True, which auto-locked a production_ref on
  an effectively local-only repo and pinned derivation to a branch that can
  never be fetched. It now requires the resolved head to have a real tracking
  ref before claiming origin backing; a dangling symref falls through to local
  detection. (production_ref.py)

• Call-graph tools returned a silent found:false across workspaces in a
  monorepo. In a nested-workspace repo, search_codebase surfaces a symbol from
  the coordinator index, but get_callers, get_blast_radius, get_callees,
  query_symbol_importers, and get_duplication_candidates re-home the file to its
  own sub-workspace profile and reject the coordinator repo arg. They returned a
  bare found:false with no reason, so the natural search-then-query flow read as
  "no data" rather than "wrong repo arg." They now return
  reason=repo-arg-mismatch so the caller can re-query with the file's own
  workspace. (tools.py)

• daemon_status reported a phantom last request. _DaemonState seeded
  last_request_at with the process start time, so /chameleon-status showed a
  last-request timestamp before the daemon had served anything, contradicting
  the field's documented "None until a request is served" contract. It now
  starts as None and is set only by a real request; the idle-reap decision runs
  off a separate monotonic clock and is unchanged. (daemon.py)

• daemon_status reported alive when the socket was gone. daemon_info judged
  liveness from the PID alone, so a daemon whose socket a /tmp reaper had
  unlinked while the process still lived was reported alive even though the fast
  path was unreachable, diverging from is_daemon_alive (the real gate).
  daemon_info now requires a connectable socket, matching the gate. (daemon.py)