docs(a2a-demo): refresh narration + lineage for ADK 1.33 + AI.GENERATE rewrite

[caohy1988]

Follow-up to #147. The two narrative docs (DEMO_NARRATION.md, DATA_LINEAGE.md) still described the pre-#147 data model — remote_agent_invocations in particular still showed the old FROM <CALLER>.agent_events WHERE event_type='A2A_INTERACTION' join that #147 replaced with the supervisor_a2a_invocations mapping read. This PR aligns both docs with the merged code.

DATA_LINEAGE.md (factual fixes)

• Layer 1 note — add the ADK 1.33 caller-side telemetry shape paragraph: the A2A_INTERACTION row lives in a sibling caller session (agent='audience_risk_reviewer'), not under the supervisor; user_id/app_name shared, no FK link.
• Layer 2 table — add the supervisor_a2a_invocations row with full schema, per-tool-call time-window pairing rule (supervisor_ts <= a2a_ts < next_supervisor_ts partitioned by user_id), and the G1.5 gate it backstops.
• Layer 3 section — add the receiver-side dual-path writer paragraph: primary AI.GENERATE path now uses SQL-style output_schema => 'decisions ARRAY<STRUCT<...>>' with deterministic model_params (temperature=0.0, topP=0.1); fallback parser in build_org_graphs._repair_receiver_extraction_from_prompt_contract re-stores via store_decision_points, which deletes existing decision rows for the parsed session_ids and appends replacements via a BigQuery load job.
• Layer 4 table — fix the remote_agent_invocations source-projection SQL to match build_joint_graph.py: reads from <CALLER>.supervisor_a2a_invocations and joins through caller_campaign_runs. The pre-fix(a2a-demo): adapt joint graph to ADK 1.33 RemoteA2aAgent sub-session telemetry #147 SQL was a real factual error.

DEMO_NARRATION.md (preserves the 5-min flow)

• Beat 1 — single optional follow-up paragraph that names the ADK 1.33 split-session shape and the bridge table. Gated by a technical Q&A audience asking.
• Beat 2 — single optional follow-up paragraph that names the AI.GENERATE typed output_schema rewrite and the fallback parser. Same gating.
• New 'Presenter aside — robustness' section between Close and Questions: names both design choices (ADK 1.33 bridge + receiver fallback) as deliberate resilience to runtime/model shifts.
• Questions To Invite — add 'How does this hold up when ADK or the model changes shape?' and refresh the closing answer.

The 5-minute spoken talk track itself is unchanged. All additions are presenter-aside material.

Verification

• build_joint_graph.py:123-137 confirms the remote_agent_invocations CTAS reads from <CALLER>.supervisor_a2a_invocations and joins through caller_campaign_runs (matches the updated Layer 4 SQL in DATA_LINEAGE.md).
• run_caller_agent.py:126 confirms the supervisor_a2a_invocations table is written by run_caller_agent.py (matches the new Layer 2 row).
• build_org_graphs.py:249 confirms _repair_receiver_extraction_from_prompt_contract is the fallback writer (matches the new Layer 3 paragraph).
• context_graph.py:2147-2224 confirms store_decision_points is delete-by-session plus load-job append, not table truncation.

Test plan

• Layer 4 source-projection SQL re-checked against the actual code in build_joint_graph.py
• Layer 2 row matches the CTAS template in run_caller_agent.py
• Layer 3 fallback write contract re-checked against ContextGraphManager.store_decision_points
• DEMO_NARRATION.md 5-min flow unchanged (no edits inside any existing beat's quoted talk-track)
• git diff --check origin/main...HEAD
• No code changes; docs-only PR