[LifeSci Skill] Add bounded target-evidence research

[phorward-deployed]

Summary

Add research-target-evidence-skill, a bounded PubMed and ClinicalTrials.gov workflow for target biology, therapeutic-program, and safety briefs and comparisons.

The skill:

• accepts one or more targets in a single invocation with shared source pacing
• preserves landmark-program and human/preclinical evidence quotas
• emits compact paper and trial cards under a deterministic aggregate output budget
• ranks target-relevant trial records while preserving registry status and safety caveats
• reports per-target partial failures and omitted-card counts
• exposes aggregate retrieval telemetry, with per-request details available only for diagnostics

The helper is intentionally stateless: each invocation executes the same bounded request plan against the public sources. The research router selects this skill for focused target-evidence reviews, and the plugin version moves to 1.1.0.

Why

Broad target-evidence prompts can otherwise expand into repeated source calls, and independent per-target invocations can duplicate work or exceed source and output limits. This provides one coordinated, size-bounded path for single-target briefs and cross-target comparisons while retaining explicit limitations and source links.

Validation

• python -m unittest test_research_target_evidence.py (14 tests)
• ruff check research_target_evidence.py test_research_target_evidence.py
• black --check research_target_evidence.py test_research_target_evidence.py
• python -m py_compile research_target_evidence.py
• quick_validate.py plugins/life-science-research/skills/research-target-evidence-skill
• git diff --check
• single-target DLL3 smoke test: 4.17 seconds, 8 public API requests, 13.9 KB result, no retries or rate-limit responses
• three-target GPC3/CLDN18.2/TROP2 smoke test: 12.96 seconds, 24 public API requests, 28.5 KB result, no retries or rate-limit responses

Limitations

This is a bounded, relevance-ranked research pass rather than a systematic review. Human/preclinical classification is heuristic, and registry adverse-event counts are not treated as automatically treatment-attributed.