v3.6.14 — Audit-hardened release

SuperLocalMemory v3.6.14 — two-stage adversarial audit hardening (Qualixar Iron Pattern Stages 8–9). All fixes validated against a green 5933-test suite under the real 3.12 runtime.

Security

• Cross-tenant cache isolation: the proxy cache now derives a per-credential tenant from the raw inbound key (pre-redaction); different API keys never share a cache entry. No-credential requests are not cached.
• Hardened the isolation against internal hook errors: a tenant-aware cache hook is never silently downgraded to the shared namespace on an internal TypeError (signature-probe instead of catch-and-retry).
• Vertex cache keys isolate by project + region.

Memory (bounded-memory lifeline)

• Per-tenant LRU caps on the semantic vector index and boundary-MLE cache (lossless — DB remains source of truth).
• Cap on the number of tenants held in memory (centroid + index shards) for long-lived shared-proxy daemons.
• Restored the SAFE-CACHE adversarial defense (fixed a 3-tuple unpack that silently disabled it on warm-start).

Reliability

• Atomic 0600 credential writes (no TOCTOU window, no torn-write corruption); fixed an UnboundLocalError that broke the credential fallback on keyring-less hosts.
• ObserveBuffer surfaces flush failures and reports an honest evaluated/captured/failed count.
• Cloud embedding reuses a pooled HTTP client.
• check_status distinguishes a corrupt settings.json from "not installed".

Packaging / platform

• Cross-platform plugin launcher (POSIX + Windows) and Windows venv bootstrap.
• Declared tomli-w so the codex IDE integration writes its TOML config (previously failed silently).
• Version reconciled to 3.6.14 across all manifests; consistency guarded by a test.