llmtrim 0.1.13

Added

• llmtrim discover finds where compressible tokens still escape compression. A
  read-only scan over the before/after capture corpus (written when LLMTRIM_CAPTURE_DIR
  is set) that re-buckets each request's token surface by block kind (system, user,
  assistant, tool_result, tool_call_args, document, tool_schema) and, with --by-tool, by
  the tool behind each tool_result. Each row reports the residual still in the compressed
  request, its share of the corpus-wide residual, and how much compression already removed
  (before→after) — so the next compression target is chosen from real traffic instead of
  guesswork. --json for the machine-readable report, --dir/--limit to scope the scan.
• llmtrim wrap <agent> convenience launcher. A one-command way to run a coding agent
  through the interceptor: llmtrim wrap claude, llmtrim wrap codex -- --model …, or any
  binary on PATH. It is sugar over setup plus a subprocess launch (no per-agent config and
  no base-URL rewriting) and it refuses to launch when HTTPS_PROXY isn't pointing at
  llmtrim in the current shell, so a wrapped agent can't silently bypass compression. Starts
  the daemon for you if the environment is wired but the interceptor is down.

Fixed

• setup now sets NO_PROXY so localhost and LAN traffic bypasses the interceptor. The
  managed shell-profile block (and HKCU\Environment on Windows) wired HTTPS_PROXY for the
  whole user, which made every proxy-aware program — not just LLM tools — funnel its local
  and LAN calls at 127.0.0.1:<port>, where they failed with "couldn't connect" whenever the
  interceptor was down or on a stale port (e.g. Plex device discovery, localhost dev servers).
  The block now also exports NO_PROXY/no_proxy (both casings — curl and Go only read the
  lowercase form): localhost, 127.0.0.1, ::1, *.local, plus the private LAN CIDR ranges.
  The literal hosts and *.local are honored by nearly every client; CIDR-based LAN bypass is
  best-effort (curl and Node/undici match only exact host or domain suffix, not CIDR). Existing
  installs self-heal: the daemon rewrites a pre-NO_PROXY block in place when it starts at
  login (reusing the wired port), so no setup re-run is needed. Already-running apps still
  need a one-time restart to pick up the new environment.
• Token-F1 bench scorer now char-tokenizes CJK, fixing degenerate scores on
  Chinese/Japanese/Korean. The scorer split on whitespace, which CJK text doesn't use, so a
  whole CJK answer collapsed to one or two "tokens" and the reported F1 was meaningless. It now
  character-tokenizes CJK runs, so quality benchmarks over CJK corpora report a real F1.
• tool_trim stage now appears as "tool_trim" in capture stages lists (was "tools").
  When description trimming is active (agent/aggressive presets), the ToolStage stage name
  is now "tool_trim" instead of "tools". This lets QA auditors correctly identify
  description-trimming runs as lossy and not flag them as category-4 bugs (lossless-only runs
  that dropped content). Lossless-only uses (selection + schema minification without trimming)
  continue to appear as "tools".