Neural command suggest from chat

Turn free-form chat into ranked typed commands using a hybrid neural ranker. Does not execute unless you explicitly apply with run.

How it works

1. Hashed bag-of-words embeddings (dim 48) for chat + each catalog command
2. Cosine similarity + linear head (Hebbian-warmed exemplars) → neural score
3. CommandIntentClassifier intent prior + token path
4. SimpleML NaiveBayes (when intent corpus available)
5. Token / phrase overlap with catalog pattern + help
6. Optional CognitiveLayers re-rank of top candidates

Commands

suggest what's the weather right now
do help me calm down
cmd suggest train tensorflow in c++
suggest top 8 show me storms and earthquakes
suggest best morning briefing
suggest apply 1          # dry-run
suggest apply 1 run      # execute via SpeechCommands
suggest layers on|off
suggest min-score 0.1
suggest selftest
suggestkit <chat...>

Example output

suggest chat="what's the weather right now"
  intent=... conf=...
  #1 [0.62] weather now  (neural / weather)
      Current weather
  #2 [0.41] weather forecast 5  (hybrid / weather)
  apply: suggest apply <n> [run]

Safety

• Default is suggest only
• suggest apply N is dry-run unless you add run
• Ethics / body freeze still apply when the suggested command executes

Files

• CommandSuggest.hpp / .cpp
• Plugin suggestkit