nudge full-body Gortex reads toward compress_bodies (#40)

[zzet]

Issue #40: Claude Code burns context reading whole files via Gortex's own read tools.

The PreToolUse hook redirects native Read/Grep/Glob to graph tools, but its matcher (Read|Grep|Glob|Task|Bash|Edit|Write) never matched mcp__gortex__read_file — once the agent was inside a Gortex read tool, nothing guarded how it was called, and read_file defaults to compress_bodies:false. So the expensive path was the default path with no forcing function.

Close the gap:

• Add mcp__gortex__read_file and mcp__gortex__get_editing_context to the PreToolUse matcher; existing installs upgrade in place (the previous matcher joins the legacy set).
• enrichGortexRead nudges a full-body read on a source file toward search_text (locate sites, no bodies) or compress_bodies:true (+keep). Soft additionalContext by default; hard deny behind GORTEX_HOOK_FORCE_COMPRESS, mirroring GORTEX_HOOK_BLOCK_EDIT. The decision is made purely from the tool input — no daemon round-trip — so the hook stays sub-millisecond. Already-economical calls (compressed, size-capped, non-source) pass silently. The nudge also rides along with the permissive-mode auto-approve as soft context.
• Derive the marketplace plugin's hooks.json matcher from the constant so it can't drift from gortex init.

Adds bench/issue40_context_repro.py, which drives the live daemon to measure the full-vs-compressed read gap (~2x)

bench/issue40_context_repro.py — drives the real tools through the running daemon (gortex mcp --proxy) and measures actual wire bytes for the three access patterns.

python3 bench/issue40_context_repro.py

What it shows (4 mid-size Go files ≈ reporter's 16–20KB C++ files)

┌────────────────────────────────┬──────────────┬────────────────────────────────────────────────────────────────────┐
│            Pattern             │     Cost     │                                Note                                │
├────────────────────────────────┼──────────────┼────────────────────────────────────────────────────────────────────┤
│ read_file (full)               │ ~17.3k tok   │ the "eats context" path                                            │
├────────────────────────────────┼──────────────┼────────────────────────────────────────────────────────────────────┤
│ read_file compress_bodies:true │ ~8.6k tok    │ 50% less aggregate, up to 75% per file — same signatures/structure │
├────────────────────────────────┼──────────────┼────────────────────────────────────────────────────────────────────┤
│ search_text (locate sites)     │ ~41 tok/site │ line-precise file:line, zero bodies read                           │
└────────────────────────────────┴──────────────┴────────────────────────────────────────────────────────────────────┘