fix(code-agent): scope the act-with-tools directive to coding requests

[lucifer78907]

Type of Change

• Bug fix
• New feature
• Enhancement / refactor
• New skill
• Documentation
• CI / build

Description

Split out of #<forge_bugs PR number> at reviewer request — this is an
LLM-behavior change and belongs on its own.

codeAgentDirective opened with an unconditional instruction:

Every response MUST include tool calls. NEVER respond with only text.

The system prompt is built once at startup and never sees the incoming
message, so that rule also governed conversational turns. A plain "Hi" left
the model no legal move: it had to produce a tool call, so it discharged the
obligation through the nearest write tool and saved its greeting to a
file.

Summary

• The mandate was unconditional but the prompt is message-agnostic, so a
  coding rule was applied to greetings and small talk.
• With no text-only response permitted, the model satisfied the constraint
  via a write tool — turning "Hi" into a file on disk.
• Fix: scope the acting-over-narrating pressure to coding requests, and give
  conversation an explicit, legal text-only path.

Changes

• forge-cli/runtime/runner.go — codeAgentDirective gains a CONVERSATION
  carve-out; FORBIDDEN / REQUIRED sections are now explicitly qualified
  "(on coding requests)". A comment records why the carve-out is
  load-bearing, so a future edit doesn't restore the unconditional form.
• forge-skills/local/embedded/code-agent/SKILL.md — mirrors the same
  distinction, with a "CONVERSATION vs CODING WORK (read this first)"
  section, and qualifies absolute rule 1.

The acting-over-narrating behaviour on actual coding work is unchanged:
"Let me patch that" with no tool call is still forbidden, scaffold →
write → run is still required in one turn, and the exploration rules are
untouched.