feat(isolate-review): add experimental Think-based code reviews

[eshurakov]

Summary

• Add a standalone, production-excluded GitHub reviewer using Cloudflare Think and a filesystem-only Computer workspace in a Durable Object, without a container or shell.
• Add development-only personal and organization APIs that prepare canonical review settings, prompts, scoped authentication, and explicit full or incremental analysis.
• Keep repository investigation read-only and publication parent-only, with dry-run defaults, pinned snapshots, ownership proofs, recovery/deadline safeguards, and bounded reads, history, and grep output.
• Add local fixture/comparison tooling, usage diagnostics, regression coverage, and shared review-summary cleanup.

Verification

No new manual end-to-end review was run while preparing this draft. Historical fixture/live pilot procedures, evidence, and limitations are documented in services/isolate-review/E2E.md, README.md, and DESIGN.md; they do not establish broad model-quality parity or production readiness.

Visual Changes

N/A

Reviewer Notes

• This remains an experimental proof of concept. It is excluded from Worker deployment automation and does not change production review routing.
• Fresh automated checks passed: pnpm lint, pnpm format:check, non-emitting/non-incremental typechecks for web, isolate-review, worker-utils, and web-env, and 1,905 selected tests: 1,107 Worker, 676 web, 48 shared auth/summary, and 74 local tooling tests. This is targeted verification, not full monorepo validation.
• Sensitive-data preflight inspected both commits and all bundled fixture objects/history. No committed credentials or private operator/customer captures were found. The Commander.js fixture retains existing public upstream author metadata; both bundled refs and the flagged historical identity were verified against the public upstream repository. No raw local logs, credentials, or private evaluation artifacts are attached.
• Review focus: admission and publication fences, incremental baseline authority, interrupted parent/child recovery, provider tool schemas, and resource budgets.

CI follow-up (2026-08-31)

All required checks passed. The only failing check is Secret Scanning: two unverified URI matches in services/isolate-review/scripts/compare-reviews.test.ts:324 and services/isolate-review/scripts/compare-reviews.test.ts:909. These are intentional dummy authenticated URLs used to test credential rejection and redaction, targeting loopback/reserved test domains, not live credentials. The scanner reports zero verified secrets. No scanner exclusions, code changes, or history rewrites were applied.

No review comments or unresolved threads were present when checks finished. The PR remains a draft; automatic Kilo review skips drafts.