feat(agent-bench): add fixture-free tau3, DABStep, and FinResearch adapters

[drewstone]

Problem

agent-lab R258/R259 shows the external finance/τ lane is not ready for policy-scored research cells:

• tau3-banking: catalog only, no runnable adapter, 0 readiness cells.
• DABStep: adapter exists, but fixture-free official checkout inspection found 5/6 required paths and 7/7 resource files present; missing dataset.csv leaves 0 official task rows.
• FinResearchBench: catalog only, no official dataset/judge path in agent-bench.

Why this belongs in agent-runtime / agent-bench

agent-lab should consume benchmark adapters through @tangle-network/agent-bench. It should not own benchmark-specific loaders, official data acquisition, or grader semantics.

Acceptance criteria

• tau3-banking adapter exists in agent-bench and fails loudly without official data/simulator/state scorer.
• tau3-banking no-spend load/preflight can return at least one official task with simulator/version metadata pinned.
• DABStep fixture-free path documents where dataset.csv comes from or changes the adapter to the official available task source without fabricating scores.
• DABStep preflight without DABSTEP_FIXTURES loads at least one official task and exposes files/ as AgentProfile resources.
• FinResearchBench adapter owns official dataset loading and the logic-tree judge path, with judge model/cost captured or explicit fail-loud status.
• All three adapters preserve the rule: fixtures test plumbing only; fixtures never count as benchmark scores.

Evidence from agent-lab

• PR: https://github.com/tangle-network/agent-lab/pull/217
• R259 report: docs/dabstep-official-readiness-r259-report-2026-07-02.md
• R258 report: docs/external-finance-tau-coverage-r258-report-2026-07-02.md

Research unblock

Once these adapters/data paths are real, agent-lab can run no-spend readiness, then calibrated one-cell smokes, then policy-vs-baseline cells for M3 external validity.

[drewstone]

R339 in agent-lab narrows the tau3-banking blocker to the consumer install boundary.

Evidence from /home/drew/code/agent-lab-k-law-r339:

• pnpm k-law:tau3-readiness-r339 -- --check passed with 0 model calls, 0 policy-scored cells, and 0 K records.
• Official checkout /home/drew/code/tau2-bench has 97 banking_knowledge tasks plus db.json.
• Fixture adapter loads: loaded 1 fixture task(s): banking-knowledge-fixture-0.
• Judge fails closed without TAU3_BENCH_DIR, which is good.
• Official preflight fails before measurement: spawn .../node_modules/@tangle-network/agent-bench/.venv/bin/python ENOENT.

The important shape: agent-lab consumes @tangle-network/agent-bench as file:../agent-runtime/bench, so pnpm copies the package into the store and the package-local .venv is not present. I also tried link:../agent-runtime/bench as a local diagnosis, and that reaches the sibling bench root, but it is not a safe consumer fix because it exposes live source/API drift to agent-lab typecheck and still lacks tau2 deps such as loguru in the shared bench venv.

Requested acceptance addendum:

• From an installed consumer such as agent-lab, TAU3_BENCH_DIR=/home/drew/code/tau2-bench pnpm exec tsx <small resolveAdapter("tau3-banking").preflight/loadTasks smoke> succeeds without symlinking live bench source.
• The adapter supports a durable tau Python environment boundary, e.g. a documented TAU3_PYTHON / TAU_BENCH_PYTHON override or substrate-managed venv, instead of assuming a package-local .venv exists after pnpm install.
• The fixture path stays plumbing-only; fixture tasks never count as benchmark scores.

Agent-lab artifact for the consumer-side receipt: runs/2026-07-07/k-law-tau3-readiness-r339.json and docs/k-law-tau3-readiness-r339-report-2026-07-07.md on branch research/k-law-tau3-readiness-r339.