refactor(bin): cut per-invocation load and repair broken skill references

diagnose read its feedback-loops reference unconditionally, so every
invocation paid for guidance most runs never used; the read is conditional
again and the per-invocation cost drops from 1,278 to 831 words. Its HITL
template moves to assets/ because it is copied out, not read as reference.
prototype's two branch flows move into references/ for the same reason —
only one branch is ever taken.

research could not search the codebase it was asked to research without Grep
and Glob. caveman's description had grown into a paragraph where one sentence
carries the trigger. Four references pointed at things that do not exist: a
to-prd skill, a /setup-matt-pocock-skills command, two cross-skill ../ links
that only resolve in the source tree, and two places calling this project's
Gitea host GitHub.

Addresses #114.

plugins/bin/skills/caveman/SKILL.md

@@ -2,17 +2,15 @@
 name: caveman
 disable-model-invocation: true
 description: >
-  Ultra-compressed output mode: drops articles, filler and pleasantries while
-  keeping technical substance exact. Cuts token usage by roughly 75%. Hand-invoked
-  only — type /caveman to turn it on, "stop caveman" or "normal mode" to turn it
-  off. Stays active across turns until you do.
+  Ultra-compressed output mode that drops articles, filler and pleasantries while
+  keeping technical substance exact, cutting token usage by roughly 75%.
 ---
 
 Respond terse like smart caveman. All technical substance stay. Only fluff die.
 
 ## Persistence
 
-ACTIVE EVERY RESPONSE once triggered. No revert after many turns. No filler drift. Still active if unsure. Off only when user says "stop caveman" or "normal mode".
+ACTIVE EVERY RESPONSE once user type `/caveman`. No revert after many turns. No filler drift. Still active if unsure. Off only when user says "stop caveman" or "normal mode".
 
 ## Rules