refactor(benchmarks): retire Tau3 Banking from Kepler (2/2)

[ayush-or]

TL;DR

Removes only Tau3 Banking's duplicate TypeScript implementation after its native Harbor replacement is validated and available. Draft / blocked on cutover; do not merge yet. MMMU Pro Vision and GPQA stay in Kepler.

What changed?

• Removes tau3_bench_banking from the registry, metadata, config union, option schemas, and CLI dispatch, along with its simulator, tools, retrieval, evaluator, and tests.
• Preserves MMMU Pro Vision, its original media manifest and mirror tooling, --image-detail, and all other remaining Kepler benchmarks. MMMU files and tooling are unchanged from the base branch.
• Adds regression coverage that Tau3 is no longer registered and MMMU, GPQA, MMLU Pro, Tau Airline, and VGI remain available.
• Does not change deployed routing, regenerate MMMU media, delete historical results, or publish datasets.

Stack and rollout

1. #113 removes the original six Harbor-backed implementations; base c6351231d8666e99b25c8f3b6e386bbb0283d4ff.
2. This PR removes Tau3 Banking only.

Native adapter prerequisite: Harbor #3595, which extends the existing Tau3 adapter for revision-pinned banking generation and BM25+grep.

Merge gates: finish native model/runtime validation, publish the pinned dataset, support Tau3's agent and runtime service in the deployed Harbor runner/catalog, and drain existing Kepler Tau3 jobs. Then sync the harness subtree and update monorepo registrations/callers together. Merging removal first would break existing callers; generation alone does not satisfy these gates.

How to test

At bdec2cf:

• Passed locally: bun run format:check, bun run check (existing warnings only), bun run typecheck, bun test (977 passed, 0 failed), and bun run build.
• Passed: CI validation.
• The first narrowing commit failed formatting on an extra blank line in CLI coverage; bdec2cf fixes it and all checks were rerun.
• Native smoke passed, not statistical parity: the same task_001 at the pinned source and BM25+grep scored 1/1 in Kepler (a3f44415) and 1/1 in Harbor (fea9e195), using openai/gpt-5.4-mini for agent/user with medium effort. Harbor also passed its Docker oracle. Step/token ceilings and simulator API/seed behavior differ; broader parity and production integration remain pending.

To reproduce, run bun install --frozen-lockfile, then the five validation commands above. Confirm the migration regression tests retain MMMU/GPQA and reject Tau3.

Reviewer focus

Upstream Tau3 defaults to BM25 without grep; Kepler uses BM25+grep. Match the source revision, retrieval, prompts, user model/effort, grading, and seeds before comparing scores. This PR deliberately remains blocked; no production benchmark is removed by opening it.

Security and privacy

No new credentials, model calls, logs, or ePHI flows in the diff. Deletes Tau3 implementation code and synthetic fixtures; preserves historical result/storage infrastructure.