Todo API pilot: Go, FastAPI and Express as todo backends

[rshdhere]

Why

Every load-test request used to be a health check, so traces were one span named GET /api/v1/health, logs said nothing, and per-route views had nothing to show. This turns the servers into small todo backends, starting with a pilot of three (one per language family), so traces, logs and dashboards tell a story you can reason about.

What changes

• Contract (servers/openapi.json): list (paging, done filter), create, get, update, delete. 400 for bad bodies or queries, 404 for missing or non-integer ids, {"error": ...} bodies, trimmed 1-200 character titles, a 1000-todo cap that drops the oldest. In memory, per server.
• Conformance test (servers/conformance.py, stdlib only): exercises every rule; all three pilot servers pass. CI runs it against every server image that serves the todo API, so the other ten get checked automatically as they are ported.
• Pilot servers: Go net/http, FastAPI, Express. Each logs rejected requests as one JSON line. FastAPI now runs one process (was 2), so every request sees the same todos; its numbers will drop accordingly.
• k6 workload (bench/k6/src/workload.ts): 35% list, 25% read, 15% create, 12% update, 5% delete, 3% filtered list, 3% invalid create (400), 2% missing read (404). Expected 4xx count as success, so error rates show only real failures. Requests are named by route and k6's url tag is dropped (no series per todo id). Servers without the todo API automatically get health checks.
• Dashboards: per-route table (rate, 4xx, 5xx, p50, p99 inside the server) and p99 per route and per operation on every server dashboard; per-route table on Inside the Servers; per-operation matrix on Head to Head. Beyla groups ids under /api/v1/todos/{id}.
• Safety nets: an unquoted {id} in Beyla's YAML made it crash-loop locally, and nothing caught it. Now CI parses the generated beyla.yaml, and deploy.sh fails unless Beyla is running and scraped.

Tested locally (full stack on Docker Engine)

• Conformance: all checks pass for Go, FastAPI and Express, both as local processes and as containers; the CI step falls back to a health check for Fiber (not ported yet).
• A race of the three through the runner: Beyla split every request by route and status; zero real k6 failures. Server-side p99 per route: Go 0.65-0.78 ms, Express 1.3-1.5 ms, FastAPI 1.9-2.9 ms (FastAPI's list route slowest).
• Dashboards render with the new per-route panels, route-named traces and the servers' rejected-request logs.

Next

Port the other ten servers the same way once this looks right in production; CI's conformance step enforces the contract on each.