F1 + F2 follow-ups: fail-closed demo mode, computed cost fallback, first tests, CI

[sgharlow]

Follow-ups from the PR #1 review, resolved for maximum credibility rather than minimum patch:

F1 — undisclosed demo output → structurally impossible

• DEMO_MODE !== 'true' → the PR_ANALYSIS job fails with an explicit reason and posts nothing. A default deployment can no longer present fabricated findings — safety by structure, not by label.
• With demo mode on: the summary comment opens with a DEMO OUTPUT banner, the heading says "(demo sample)", and every inline comment self-identifies.
• Proven by test: an exploding-proxy GitHub client (any property access throws) passes untouched through the disabled path.

F2 — the fail-closed pricing claim is now computed, not asserted

• Fallback = max over the pricing table (was hardcoded gpt-4 while gpt-4-32k costs 2×). Test asserts the fallback ≥ every listed model, so the guarantee can't silently rot when prices are edited.

First tests + CI (closes deferred Sprint-5.1)

• node:test over compiled dist — zero new dependencies; 3/3 green locally.
• CI: build + tests on Node 20/22, Docker build validation. This repo went from never-compiled to compile-gated in one day; this PR makes regression impossible to miss.

README

• Banner updated: repaired state, benchmark finding-Repair: make the project compile (restore never-committed modules, fix dep type drift) #1 case-study pointer (the flaw is now the exhibit for why the gates exist), honest claim level (built, not live-proven).