docs(readme): architecture diagram, runnable demo, and the over-built note

[JumpTechCode]

Presentation pass so the README lands as an architecture showcase (no new subsystems — diagram, runnable demo, claim integrity).

Changes

• Visual architecture diagram. Replaced the ASCII box-art with a GitHub-rendered mermaid flowchart of the webhook → outbox → Kafka → enrichment → diagnosis flow (with pgvector memory and embed-on-resolve).
• "See it run — webhook to diagnosis". A copy-paste path: make compose-up → make migrate → an HMAC-signed curl to /webhooks/generic → poll /incidents/<id>. It produces a real diagnosis. The bare-stack output deliberately shows the evidence-citation gate working — with no real integrations wired there's nothing to cite, so it strips the invented refs, flags hallucinated_evidence, and caps confidence — and points to the eval results for the full-context numbers. Prereqs (bootstrap, .env, the generic secret, jq/openssl) are called out so it runs top-to-bottom.
• "Isn't this over-built?" note backed by new ADR 0015 — names the trade-off (deliberate production patterns for a portfolio) and the honest simplification (collapse Kafka into the outbox-as-queue for a real single-tenant deploy).

Verified

• Every command was run live against the stack: signed webhook → 202, diagnosis produced and served via the API.
• The illustrative diagnosis JSON matches the real API response field-for-field (likely_category, confidence, hypothesis, hallucinated_evidence, evidence, suggested_actions[].{description,command,risk,rationale,requires_human_approval}).
• Mermaid syntax is conservative (quoted labels, standard shapes) and renders on GitHub.
• Senior doc review (claim-integrity focused): the two Important findings — demo prerequisites and jq — are fixed in this PR.

Pairs with the foundational fixes landed alongside it (#79 temperature, #81 observability wiring, #80 multistage image, #82 prompt packaging) so every claim the README makes is now backed by the code.

🤖 Generated with Claude Code