On read or parse error, LoadFromYAML must leave the existing rule set untouched so a bad reload doesn't strip enforcement.

Re-loading the same YAML must not double the rule set. Before LoadFromYAML adopted replace semantics, calling it twice appended the rules twice, which

Regression: eviction via plain reslice kept the original backing array alive; cap() grew proportional to total-logged count. After the fix, cap() stay

After rollover eviction, the surviving head's PreviousHash must equal the seam hash (the hash of the last evicted entry). Forging both PreviousHash an

--- New comprehensive tests ---

---- Existing tests above, new tests below ----

--- New comprehensive tests ---

── PolicyEngine integration ─────────────────────────────────────────────────

Regression: checkRateLimit used to increment count then compare, so once over the limit the counter grew unboundedly. After the fix, the counter stays

Regression: rate-limit state was previously keyed by rule.Action only, so one agent / tenant making rapid calls could starve all others for the rest o

Regression: prior to this fix, Log() returned the same *AuditEntry it stored in al.entries; GetEntries() returned the same pointers. A caller could mu