feat(gemini): update search anchors prompt for interventions and eligibility keywords

Add instructions for Gemini to generate target drug names from
biomarker-to-drug mappings and patient clinical features for
eligibility pre-filtering.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

trialpath/services/gemini_planner.py

@@ -78,6 +78,15 @@ class GeminiPlanner:
         2. Include appropriate geographic filters
         3. Consider the patient's age and performance status
         4. Set a relaxation_order for broadening search if too few results
+        5. Generate a list of target interventions (drug names) based on:
+           - Known effective drugs for the patient's biomarkers (e.g., EGFR → osimertinib, erlotinib; ALK → alectinib, crizotinib; KRAS G12C → sotorasib, adagrasib; ROS1 → crizotinib, entrectinib; BRAF V600E → dabrafenib + trametinib)
+           - Next-line therapies after the patient's current treatments
+           - Immunotherapy drugs if no targetable mutations (e.g., pembrolizumab, nivolumab, atezolizumab)
+        6. Generate eligibility_keywords from the patient's key clinical features:
+           - Performance status (e.g., "ECOG 0-1")
+           - Stage (e.g., "stage IV", "stage IIIB")
+           - Key biomarker terms as they appear in trial criteria (e.g., "EGFR mutation", "PD-L1 positive")
+           - Prior therapy line (e.g., "first-line", "second-line")
         """
 
         raw = await self._generate(prompt, SearchAnchors.model_json_schema())