#14701: Agent: enforce execute-first replies and clarify whisper modes

<!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This PR updates agent guidance and follow-up prompt construction, and adjusts skill documentation: - Adds “execute-first” tool-call guardrails to the agent system prompt (`src/agents/system-prompt.ts`) with a unit test to pin the new text. - When draining queued messages in collect mode, appends a stale-state reminder to the collect summary so the model treats queued items as potentially invalidating prior assumptions (`src/auto-reply/reply/queue/drain.ts`) with a corresponding test update. - Clarifies Whisper skill docs to distinguish local vs API transcription, and flips a couple bundled `skills/**/scripts/*.sh` files to be executable; adds a test ensuring skill shell scripts are executable on Unix. <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk. - Changes are small, localized, and covered by updated/new unit tests; the only functional change is adding an extra reminder line to an internal collect prompt summary and adding prompt guardrail strings. - No files require special attention <!-- greptile_other_comments_section --> _{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!} <!-- /greptile_comment -->