fix(go-kernel): accept pre-normalized ActionContext in evaluate command

[jpleva91]

Closes #957

Implementation Summary

What changed:

• Extracted parseActionInput() helper in go/cmd/agentguard/main.go that detects input format by checking for an action field (pre-normalized ActionContext) vs tool field (raw Claude Code tool call)
• Pre-normalized ActionContext payloads (e.g., {"action":"file.write","target":"foo.ts"}) are passed directly to the evaluator — no re-normalization
• Raw tool call payloads ({"tool":"Write","input":{...}}) continue to be normalized (existing behavior preserved)
• Added doc comment on runEvaluate documenting both input formats and the pack: limitation
• Added go/cmd/agentguard/main_test.go with 6 unit tests covering both formats and the normalize | evaluate pipeline
• Updated go/test/smoke.sh with 2 new tests: pre-normalized format and the pipeline

How to verify:

# Repro from issue — now returns allowed instead of default deny:
echo '{"action":"file.write","target":"foo.ts"}' | go/bin/agentguard evaluate --policy agentguard.yaml

# normalize | evaluate pipeline:
echo '{"tool":"Write","input":{"file_path":"src/main.ts"}}' | go/bin/agentguard normalize | go/bin/agentguard evaluate --policy agentguard.yaml

# Go tests:
cd go && go test ./cmd/agentguard/...

Tier C scope check:

• Files changed: 3 (limit: 5)
• Lines changed: ~85 (limit: 300)
• Breaking changes: None (existing raw tool call format behavior preserved)

---

Tier C implementation by copilot-cli — AgentGuard three-tier governance

[jpleva91]

🟢 LOW — Audit Receipt

Audit: Merged PR Audit Report 2026-03-27
Agent: claude-code:opus:reviewer (audit-merged-prs)
Timestamp: 2026-03-27T02:50:00Z

Risk Score: LOW

All CI checks passed. No protected file changes. No immediate action required.

See consolidated audit issue: #1066