stack-bench: team-chat cross-backend agentic benchmark

[cloutiertyler]

Summary

Adds team-chat, the main stack-bench task: a cross-backend agentic benchmark where an AI coding agent must build a full team-chat backend (rooms/owners/membership with permissions, presence, server-maintained unread counters, message edit + tombstone delete, idempotent sends, per-room gapless sequence numbers under concurrency, atomic credit tips), graded purely behaviorally through each backend's real client SDK.

Built in Harbor format (the Terminal-Bench 2.0 harness): fixed prompts (instruction.md), Dockerized environments, oracle solutions, machine verification. This is separate from the one-shot model leaderboard in tools/xtask-llm-benchmark/ — stack-bench fixes the agent+model and varies the backend.

Verifier

One backend-neutral TypeScript scenario (_shared/harness/src/teamChat/) against an AppClient contract; per-backend adapters (~200 lines each). 46 checks in 4 weighted metric groups, emitted as reward.txt + multi-metric reward.json + per-check result.json (the feedback payload for multi-step agent loops):

• correctness 0.40 — validation/permission edges, idempotent resends, gapless seq with 3 concurrent writers, exact unread counters (multi-row atomicity), tip conservation under concurrent transfers, monotone mark-read, tombstone privacy, τ-bench-style full goal-state comparison
• realtime 0.30 — exactly-once in-order fan-out to 3 subscribers, server-side cross-room isolation, late-joiner history-then-live with no boundary dupes/gaps, live propagation of edits/deletes/presence/membership/balances
• durability 0.20 — Jepsen-style backend process kill + restart mid-scenario via each environment's /opt/stack-bench/backendctl restart; all state must survive, seq counters must continue exactly, dedupe records must persist
• perf 0.10 — delivery latency p95 + 120-message concurrent burst; raw p50/p95/throughput always reported

Both environments are single-container (SpacetimeDB: rust:1.93 + CLI; Convex: the official convex-backend image with Node added), so the backend shares the agent's resource budget and process-level restart works uniformly.

Validation

Oracle reward 1.000 through Harbor on both backends (46/46 checks, 0 exceptions):

	SpacetimeDB	Convex
Oracle reward	1.0000	1.0000
p50 / p95 latency	4ms / 7ms	18ms / 27ms
Burst throughput	952 msg/s	~90 msg/s

Also in this PR

• xtask reworked for the tasks/<task>/<backend> layout (--task flag, default team-chat); cargo stack-bench alias + workspace member wired up
• The original realtime-chat spike kept as a smoke test

Usage

cargo stack-bench list
cargo stack-bench oracle spacetimedb        # harness self-test, expect 1.0
cargo stack-bench agent convex --model anthropic/claude-fable-5
cargo stack-bench all --agent claude-code --model ...   # comparison table

Next steps

Real-agent runs (needs ANTHROPIC_API_KEY / CLAUDE_CODE_OAUTH_TOKEN), pass^k trials via -- -k N, effort metrics in the comparison table, more backends (Postgres, Mongo, Supabase, Firebase), pre-baked deps for hermeticity.