Preprocessor behavior for multi-directive input (clarify vs abstain)

[rlippmann]

Currently, the preprocessor assumes a single directive per message.

Open question:
How should the system handle inputs that appear to contain multiple directives?

Examples:

• "clear premise; reset policies"
• "use docker and prohibit peanuts"
• multi-line:
  clear premise
  reset policies

Options:

1. Return <NO_DIRECTIVE> (current behavior)
2. Return clarify via host UX (e.g., "Please send one directive at a time.") while preprocessor output remains <NO_DIRECTIVE>
3. Attempt to execute multiple directives (likely out of scope)

Key constraints:

• precision-first (false positives are worse than false negatives)
• whole-message matching (no partial extraction)
• deterministic behavior
• no semantic guessing

Considerations:

• usability: users may naturally batch commands
• risk: detecting multiple directives reliably without false positives
• interaction with multi-line input
• whether detection should be heuristic-only or also handled by LLM

Open questions:

• what qualifies as a "clean" multi-directive input?
• should multi-line inputs be treated differently?
• should this logic live in heuristic or LLM layer?

Non-goal:

• no changes to core engine grammar/semantics in this issue

Decision deferred for now.