docs: supervisor prompt + agent-profile ideas + supervisor-lab port plan

[drewstone]

Adds docs/supervisor-profiles.md — a design doc with four AgentProfile sketches for the research/knowledge domain, informed by emilkowalski/skills and the existing Tangle supervisor stack.

What's in it

Four profile sketches (system prompt + tools + when-to-use), each naming the real primitive it composes (so building is "assemble", not "invent"):

1. grounded-researcher (leaf worker) — already exists twice (researcherProfile here, grounded-researcher.ts in supervisor-lab); the doc is the canonical shape both converge on.
2. verifier-driver (the gate, not the judge) — composes agent-runtime's verify({implement, verifier}) combinator; selector ≠ judge.
3. research-supervisor (ground → fan-out → gate → fuse → recurse) — the research-specialized merge of supervisor-lab's existing driver archetypes.
4. dedup-verifier (thin, cheap, mostly deterministic) — the emilkowalski dedup lesson as a profile: the "taste" lives in the deterministic validator (createResearcherValidator, the lint/validate CLI), the verifier is a thin lens that runs the cheap check first and escalates to an LLM only for the one semantic residual.

What we borrowed from the skills repo

Tight two-field frontmatter, terse non-negotiable numbered rules over prose, "default to flagging; approval is earned" (verifier posture), disable-model-invocation (verifier is routed, not auto-invoked), and cheap-because-narrow (generate vs. judge separation — the reason the judge can be thin).

supervisor-lab status

It exists (tangle-network/supervisor-lab) and the two-agent loop is already ported as bench/supervise-topology.ts (supervisor mounts the coordination MCP over a live Scope/Supervisor, authors worker profiles as code, spawns/drains/steers, grades on a real judge; OFFLINE $0 + LIVE paths). The doc records this and gives a concrete plan to land the two new profiles + wire the thin verifier gate into the existing harness (new BENCH=research arm grading via createResearcherValidator), rather than rebuilding anything.

Docs-only. CI=true pnpm install clean; merges cleanly into main. Do not merge — for review.

[tangletools]

✅ Auto-approved PR — 292a41db

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-25T01:18:57Z}