CL-6440: longevity campaign harness (@corbits/longevity-sim)

[TheGreatAxios]

Adds @corbits/longevity-sim, the longevity simulation harness for CL-6440 (references the ticket; does not close it — the campaign findings land there).

What it does

• Boots a scratch stack (hub + sidecar + dedicated Postgres, free ports) and provisions a simulated 10-person sales team via real sign-ups over the public HTTP APIs.
• Deploys real-inference agents through the workflow deployment freeze path (asset -> source push -> POST /workflows/deployments -> invite), spread across multiple Ollama-served models via openai-compatible catalog providers — no Anthropic, no stub inference; hub/sidecar children inherit a whitelisted env, so no API key can leak in.
• Drives a seeded, deterministic campaign plan: paced persona chatter, thread replies, concurrent bursts, fire-and-forget and measured agent turns, time-compressed routines (scheduler-claimed via next_fire_at advance + manual runs), hub restarts mid-conversation, provider switches, and mid-campaign skill-edit / agent-spawn self-improvement checks.
• Checkpoints degradation metrics by message count (send/turn latency percentiles, DB size, message-page and workbench-list timings at depth, hub/sidecar RSS, collector-failure log signatures) and renders a markdown campaign report with knee detection and a verdict.

Notes

• Pure core (personas, plan, metrics, knees, report) is fully unit-tested (51 tests); the stack/engine layer drives only public HTTP APIs plus the shared e2e boot glue.
• Verified live end-to-end against a Mac Studio Ollama fleet (4 models): boot, 44-message smoke campaign, restart recovery, provider switch, spawn-at-depth all green; skill-probe failure correctly attributed to the known tools/history blocker on the openai-compatible path.

References CL-6440.