Decision cache key should include auth mode and policy-relevant inputs

[beonde]

Problem

The guard decision cache (_decision_cache in guard.py) keys entries on (badge_jws, tool_name) only. However, evaluate_tool_access() decisions also depend on:

• params_hash
• server_origin
• policy_version
• capability_class / deny_on_unknown_class
• Auth mode (badge vs API key vs anonymous)

When badge_jws is empty (API-key or anonymous callers), all such requests share the same cache bucket, which can return incorrect decisions.

Additionally, cached results reuse evidence_id / evidence_json, skipping the per-call evidence guarantee.

Current mitigation

The 5-second TTL limits the blast radius. In practice, badge-based auth (the primary path) produces unique JWS strings per identity, so collisions are unlikely for the main use case.

Proposed fix

Either:

1. Expand the cache key to include (badge_jws or api_key or "anon", tool_name, params_hash, server_origin) and generate fresh evidence IDs on cache hits
2. Skip caching entirely when badge_jws is absent

Context

Identified during PR #32 code review (Copilot review comments on guard.py:360).