refactor(research)!: remove the unused web research drivers and TCloud

[drewstone]

Problem

The package says it "does not browse, call models, or run agents", but it shipped a web research stack that does: runVerifiedResearchLoop, the TCloud-backed web worker and router client, and the adaptive, collection, and driving drivers. It also shipped the claim-grounding verifier and decorator, and an investment-thesis task with a 900-line answer set compiled into src/. Nothing outside this repository imports any of it. The stack is the only user of @tangle-network/tcloud, so every Knowledge release also had to fit the TCloud and Sandbox dependency cohort.

Change

• Delete verified-research-loop, web-research-worker, adaptive-driver, collection-research-driver, research-driving-driver, investment-thesis-set, investment-thesis-task, and material-facts-metric, with their tests (tests/loops/, tests/eval/, and the driver cases in claim-persistence.test.ts).
• Trim claim-grounding.ts to groundClaimInText, which RAG answer scoring uses.
• Drop @tangle-network/tcloud from dependencies. The lockfile loses 190 lines.
• Delete scripts/prove-research-live.mjs and scripts/prove-research-transport.mjs, one-off proofs for that transport (refactor(research): published TCloud transport with compatible APIs and cancellation #222).
• verify-package.mjs no longer requires runVerifiedResearchLoop. Its obsolete two-agent "forbidden export" guards are gone because api-surface.json already records the exact export set.
• Keep the verified-research paper, docs/results/*, and the investment answer key as a historical record. Each is marked historical, and its code links now point at bc2abe0, the last commit that contains the code.
• Bump to 20.0.0 (69 exports removed; manifest dependency change) and add a changelog entry.

Kept: claim ledgers, TrackedClaim, and the KbStore claim-ledger methods. Agent Runtime's knowledge-improvement-job test calls store.putClaimLedger.

Consumer evidence

I extracted every identifier imported from @tangle-network/agent-knowledge (named, type, namespace, import(), and createRequire forms) in:

• the Mac ~/webb and ~/company checkouts and worktrees (939 code files);
• ~/code on beelink1 (4,125 files) and drew-gtr-pro (5,711 files);
• fresh default-branch clones of every repository that GitHub code search lists as depending on the package (tangle-network: agent-runtime, agent-app, agent-builder, agent-dev-container, blueprint-agent, creative-agent, discovery-lab, gtm-agent, insurance-agent, legal-agent, ph0ny, physim, redteam, relationships-agent, supervisor-lab, tangle-tools, tax-agent, tuner-agent, ai-trading-blueprint; drewstone: law-lead-agent, legal-underwriting-agent, skeletal-os), plus discovery, agent-eval, agent-sdk, and braid.

None of the 69 removed exports is imported anywhere. The only textual matches are discovery docs that describe the API, and frozen bundles in discovery-lab preregistrations. Those bundles contain their own copy and do not resolve the package.

Why this is the right long-term shape

Knowledge owns knowledge state. Model calls and browsing belong to callbacks or Agent Runtime, as the README and AGENTS.md already state. Removing TCloud takes the package out of the TCloud and Sandbox cohort.

Cost

39 files, 52 insertions, 9,277 deletions (source 4,162 lines, tests 4,688, scripts 137, lockfile 190). Major version bump: consumers on ^19 are unaffected until they move to 20. Risk: an unlisted private consumer of the research drivers. Rollback: revert this PR, or stay on 19.x.

Verification (beelink1, merged with origin/main)

• pnpm install --frozen-lockfile, pnpm lint (one existing warning in proposals.ts), pnpm typecheck, pnpm build, and pnpm run api:surface (950 exports) all pass.
• node scripts/check-version-bump.mjs passes: 19.1.5 -> 20.0.0 pays for the 69 export removals.
• pnpm test with network tests: 872 passed, 2 skipped, 1 failed. The failure was an unhandled transaction error in tests/kb-improvement/activation.test.ts under full-suite load on the busy host. That file passes alone (9/9) and does not touch the removed code.
• Every relative markdown link resolves, and no live file references a deleted module.
• pnpm run verify:package fails the same way on main and on this branch: agent-core dependency on agent-interface declares ^3.0.0, which does not admit installed 2.16.0. Agent Core 0.10.3, published 2026-10-06 07:00Z, requires Interface 3 inside Eval 0.208.2's ^0.10.2 range. That is an upstream cohort break, not this change.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[tangletools]

✅ Auto-approved PR — 89d7b468

Blanket team auto-approval is intentional. This is not a code review.
No automated review runs on this PR. This approval rests on the rule above alone.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-10-06T07:06:46Z}