security: await audit writes, add per-turn AssistantPrompt audit (#55)

- Auth.js authorize/signOut: await createAuditEntry on every branch so auth
  events land in the audit store before the JWT is minted / session closes.
  Previously these were fire-and-forget and would be dropped under DB load.
- Assistant chat: make appendPromptInjectionGuard async and await its own
  SecurityAlert audit; add auditUserPromptTurn() that records every user
  message turn as an AssistantPrompt entry containing conversationId, length,
  SHA-256 fingerprint, pageContext and whether the injection guard fired.
  Raw prompt text is intentionally not stored — the hash lets a responder
  correlate a chat transcript with a forensic request without the audit
  store accumulating a plain-text corpus of everything users typed.
- Replace bare crypto.* with explicit node:crypto imports.
- Document the retention posture in docs/security-architecture.md §6.

Fixes gitea #55.

docs/security-architecture.md

@@ -131,11 +131,24 @@ injection attempts and to surface them as audit-log entries.
 
 ### Activity History System
 
-- Centralized `createAuditEntry()` function (fire-and-forget, never blocks)
+- Centralized `createAuditEntry()` function. Security-critical callers (auth, assistant
+  prompts, admin mutations) `await` the write so the entry is durable before the
+  user-visible effect completes; non-critical callers may fire-and-forget
 - Covers 29+ of 36 tRPC routers
 - Logged fields: `entityType`, `entityId`, `action`, `userId`, `changes` (JSONB with before/after/diff), `source`, `summary`
 - Authentication events: login success/failure, logout, rate limiting, MFA failures
 
+### Assistant prompt audit
+
+Each user turn through the AI assistant writes an `AssistantPrompt` audit row
+with conversation ID, prompt length, SHA-256 fingerprint, current page context,
+and whether the prompt-injection guard flagged the input. Raw prompt text is
+**not** retained by default — the hash + length fingerprint is enough for a
+responder to correlate an audit row with a later forensic export if the user
+retains their chat transcript, but the audit store itself does not accumulate a
+plain-text corpus of everything users typed into the assistant. This balances
+GDPR Art. 30 (records of processing) against data-minimisation.
+
 ### External API Call Logging
 
 - All OpenAI/Azure/Gemini API calls logged via `loggedAiCall()` wrapper