docs(audit): tighten meeting-content-egress claim per Layer 2 findings

A full adversarial audit pass (six parallel agents against
audit/prompts/prompt-pack.md) found the claim's headline sentence
misleading on two points, even though both were already disclosed in
the caveat text:

1. The notetaker sends full transcript text to the configured BYO-LLM
   for summarization, not just recorded audio to Deepgram — a second
   real destination for meeting content that the headline omitted.
2. LiveKit is listed under "meeting links" alongside Zoom, but if an
   operator points it at LiveKit Cloud instead of self-hosting, live
   meeting audio/video transits LiveKit Cloud's infrastructure for
   the call's duration — a materially different claim than a Zoom
   join-link, which never touches Calnode's or a third party's media
   pipeline at all.

Rewrote the claim's headline to name both real destinations up front,
and gave LiveKit Cloud its own explicit paragraph in the caveat rather
than a one-line bullet next to Zoom, framing it the same way as a
self-hosted-vs-managed-database choice: identical code path, operator
draws the trust boundary.

No code changes — this is a wording-accuracy fix surfaced by the
audit, not a new vulnerability. The 9 real code-level findings from
this audit pass (SSRF gaps in CalDAV/LLM/LiveKit config, a cross-host
meeting-content leak in notetaker_http.go, a booking-creation
validation gap, an MCP consent CSRF gap, and others) are tracked
separately, pending a scoping decision on which to fix first.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: shockalotti

audit/claims.yaml

@@ -54,12 +54,19 @@ claims:
 
   - id: meeting-content-egress
     claim: >
-      Meeting CONTENT — recorded audio and transcripts — never leaves your
-      infrastructure except to Deepgram, and only when you explicitly enable the
-      optional notetaker feature.
+      Meeting CONTENT — recorded audio and transcripts — only ever leaves your
+      infrastructure to services you explicitly configure: Deepgram (audio, if you
+      enable the notetaker) and your own BYO-LLM endpoint (transcript text, for the
+      same notetaker's summary step). Nothing carries meeting content by default.
     verify:
       - "grep -rn 'api.deepgram.com' internal/ — the single outbound call site
-         that receives recorded audio (internal/stt/deepgram.go)."
+         that receives recorded audio (internal/stt/deepgram.go). Calnode hands
+         Deepgram a presigned, time-limited GET URL into your own bucket rather
+         than uploading the file itself — Deepgram's servers fetch it."
+      - "internal/handler/notetaker.go — the summarize step sends the full
+         transcript text to h.getLLM() (your configured BYO-LLM endpoint), the same
+         client used by the conversational booking assistant. See
+         internal/llm/client.go."
       - "Recordings themselves go to YOUR OWN S3-compatible bucket (LITESTREAM_*
          env vars), never to Calnode or any third party — see
          internal/handler/livekit_recording.go and migration 00043."
@@ -73,16 +80,23 @@ claims:
       that carry no meeting content:
         - Calendar sync: Google Calendar, Microsoft Graph/Teams, CalDAV providers
           (iCloud, Fastmail, Nextcloud, etc.) — internal/gcal, internal/calendar.
-        - Meeting links: Zoom (per-host OAuth), LiveKit — either self-hosted or
-          LiveKit Cloud, an operator choice, not a Calnode-run service.
+        - Meeting links: Zoom (per-host OAuth) — meeting metadata only (topic,
+          time); Zoom's own infrastructure handles Zoom-call audio, outside
+          Calnode's process entirely.
         - Payments: Stripe, for paid bookings.
         - Optional analytics: GA4/GTM, native and consent-gated (see
           cookie-consent-gated-analytics below) — off by default, and even when on,
           carries no meeting content.
-        - BYO-LLM: your configured LLM endpoint (default api.openai.com, fully
-          swappable to any OpenAI-compatible provider) receives booking constraints
-          and, for the notetaker, transcript text to summarize — never raw calendar
-          data. See llm-no-raw-calendar-data below.
+      LiveKit (built-in video) is the one integration where "leaves your
+      infrastructure" depends entirely on an operator choice made at setup, not a
+      Calnode default: if you point it at a self-hosted LiveKit server, live
+      meeting audio/video never leaves infrastructure you control. If you point it
+      at LiveKit Cloud instead, live meeting media transits LiveKit Cloud's
+      infrastructure for the duration of the call — a third party beyond Deepgram/
+      your BYO-LLM, and beyond what this claim's headline sentence alone would
+      suggest. This is exactly analogous to choosing a self-hosted vs. managed
+      database: Calnode's code path is identical either way, but the trust boundary
+      is drawn by your own infrastructure choice, not by the code.
       A skeptical reader should verify this list is exhaustive themselves — that's
       exactly what the adversarial prompt above is for.