docs: add payments case study #1 (success theater)

[MarouaBoud]

The tool has been build-complete and CI-green for a while but had never been pointed at a target in a written-up, reproducible way. This is the first case study — the launch-track "prove it works in the wild" deliverable.

What's here

• docs/case-studies/01-payments-success-theater.md — a 60-agent run against the payments world, graded B, written up entirely from the real report. Headline: 100% task success, 24% misuse — 42% of agents told to charge_customer called list_charges instead (and 37% the reverse, which charges the card). Exactly-once held under an 11-fault chaos storm; the 9-agent adversarial cohort was contained (0 destructive tools reached).
• docs/case-studies/payments.stampede.yaml — the exact config, so the whole thing is reproducible with one command.
• README.md — a "See it in action" section linking the study, replacing the old hypothetical misuse example with the real one.

Honesty notes (called out in the doc)

• This is the deterministic --dry-run path, so all dollar figures are $0 — the "4.7× cost spread" is a token spread and the denial-of-wallet amplification is structural, not monetary. The misuse / persona / exactly-once / containment findings are real and brain-independent. The doc states this plainly per the toolkit's "honest over impressive" rule.
• Every number is read straight from the run's JSON report; the reproduce commands were run verbatim and produce byte-identical output. The run_id is intentionally not quoted as a reproducibility anchor because it hashes the full config (output paths included) — the findings are the stable guarantee.

Docs-only; no code touched.