policy: scope evidence-discipline trigger words after first R measurement

[achimdehnert]

Follow-up to PR #251.

Current trigger list includes very common words (done, fertig, claim, evidence, deployed). The inject_policies.py hook will fire on almost every session-ende, PR-completion, and "fertig?" turn — each fire injects 54 lines.

When to act

After first Signal R measurement (sibling issue #256) — not before. Broad triggers are deliberate during the falsification window so the policy gets maximum exposure to either prove or fail its own test.

Decision criteria

• R ≥ baseline (works): narrow to compound markers — done ✓, festgehalten:, deployed to, verifiziert:, pre-existing, not my code, infra smell, confabulated; keep single-word COMPLETE (caps).
• R < baseline: policy is cut per its own clause; no scoping needed.

Acceptance

• First R measurement complete (blocks on tools: implement Signal R measurement for evidence-discipline policy #256)
• Decision recorded in policies/evidence-discipline.md changelog
• If kept: triggers updated, sample of ~20 prior session prompts re-tested against narrower list — injection rate dropped ≥50% without missing target cases.

Refs: PR #251, blocks on #256.