Ethics guardrails: rejection hard-constraint + unified/extended anti-exhaustion (audit P4)

[BeamusWayne]

Follow-up to #4 (P1–P3). This is the ethics/behavioral phase of the audit, kept in its own PR for focused review. Prompt/SKILL only — no Python.

Part A — clear rejection is a non-overridable hard constraint

• SKILL.md run-rule 4: once the other person clearly rejects (不喜欢你 / 只是朋友 / 把你当朋友 / 喜欢别人), message/confess/crisis/progress must NOT output any pursuit advice — overrides every signal, score, stage, and rejection subtype.
• crisis_handler.md C-1:
  • "我把你当朋友 / 我们只是朋友" reclassified as clear rejection (was "温和型 → 可能有余地"), aligning with SKILL.md's "一句明确的'我们只是朋友'覆盖所有绿灯".
  • Removed the "不能表现出来" (hide-your-intent) line.
  • Default post-rejection action is graduated disengagement.
  • "继续追" downgraded from a default branch to a tightly-gated exception requiring the other person to re-initiate explicit mutual signals, with mandatory frequency/intensity decay. Removed the "用时间冲淡被拒标签" framing.

Part B — anti-exhaustion precheck unified and extended

• New prompts/burnout_precheck.md = single source of truth: unified triggers (scope = all /simp, OR logic), 劝歇 output, escalation policy, no-profile session-level fallback.
• Extended the precheck to /simp message, /simp analyze, /simp crisis — the real anxiety-peak entry points — not just daily/progress.
• daily_coach Step 0 and progress_tracker risk row now reference the single source instead of carrying divergent thresholds (was OR vs AND, different counting scopes).
• Escalation strengthened: on a double-hit (high frequency AND anxiety words) the tool holds firm and refuses pursuit advice instead of yielding on the second ask.

Why separate from #4

#4 was mechanical/low-risk (test reliability, data integrity, naming consistency). This PR changes how the product behaves around rejection and user anxiety — it encodes ethical stances and benefits from focused review.

Test plan

• python3 -m pytest -q → 225 passed (prompt-only change, no regression)
• grep clean: no "不能表现出来" / "让时间冲淡…标签" left
• burnout_precheck wired into SKILL.md + daily/progress/message/analyze/crisis

Still pending (not in this PR)

• P5 — data-consent product positioning (whether/how to analyze a non-consenting person's private data); needs a product decision.