Add token-aware completion guidance for concise outputs

[apenab]

Problem

Completion responses can be longer than needed, causing avoidable token spend.

Proposal

Introduce token-aware completion guidance:

• Instruct responses to stay concise (e.g., max words/compact format).
• Keep factual completeness and citation/grounding requirements.

Note: current RLM config already sets max_tokens=4096 per subcall; this change targets actual verbosity, not just hard caps.

Expected Impact

• Estimated token reduction: ~5-10% (likely marginal in practice)

Risk

• Low

Acceptance Criteria

• Conciseness guidance is configurable and documented.
• Measurable reduction in average completion length/tokens.
• No meaningful regression in correctness/grounding.
• Report includes token savings vs baseline.