Add Cloud Agent development environment

[abimaelmartell]

Summary

Adds a repo-managed Cloud Agent environment for pdf-inspector (primary) with pdf-evals as a repository dependency, so a fresh Cloud Agent boots ready to build the binaries and run the regression/scoring suite end to end.

• .cursor/environment.json — default base image + install script, with github.com/firecrawl/pdf-evals declared under repositoryDependencies.
• scripts/cloud-agent-install.sh — idempotent, non-interactive setup:
  • installs the small missing system packages (python3.12-venv, qpdf);
  • builds the release binaries (pdf2md, detect-pdf) using the pinned rust-toolchain.toml channel (1.98.0), which cargo auto-installs;
  • creates the pdf-evals .venv and installs only the core benchmark/scoring deps (the section of requirements.txt above # optional provider dependencies), skipping the heavy cloud-OCR providers (docling/torch, pymupdf4llm, markitdown, llama-cloud) that the local dev flow does not need;
  • writes pdf-evals/config.yaml pointing the local provider at the freshly built pdf2md (git-ignored), so bench.py works without any extra env var.

No start/terminals: pdf-inspector is a CLI/library with no long-running service.

Validation (on the VM, from a clean state)

Check	Result
Pinned toolchain	rustc 1.98.0 auto-installed from rust-toolchain.toml
cargo fmt --check	OK
cargo clippy -- -D warnings	OK (zero warnings)
cargo test	1083 + 172 + 3 unit/integration + 2 doctests passed, 0 failed
Release binaries	pdf2md / detect-pdf build and extract a sample PDF
bench.py doctor	194 PDFs, 0 errors
bench.py test -q	20/20 PASS at 100% vs committed snapshots (binary resolved via generated config.yaml)
bench.py score -q	20/20 scored, composite mean 0.58 (TEDS/MHS via apted+lxml)
Install idempotence	Clean run + two reruns all succeed

Note on build-test tooling

This setup was validated by running the exact install script from a clean state on the VM (fresh .venv/config.yaml, re-run for idempotence). The prebuilt-image build-and-verify path was not exercised because this agent run is anchored to the pdf-evals repo, so the build tooling cannot check out pdf-inspector as the primary repo to test its committed .cursor/environment.json. Once merged, the committed file is the highest-precedence source for Cloud Agents started on pdf-inspector.

  

---

Summary by cubic

Adds a repo-managed Cloud Agent environment for pdf-inspector so a fresh agent boots ready to build the release binaries and run the pdf-evals regression suite.

• The install script is idempotent and non-interactive, installing only the core benchmark dependencies (skipping heavy OCR providers).
• It writes a git-ignored config.yaml in pdf-evals to point at the freshly built pdf2md binary.

^{Written for commit c853394. Summary will update on new commits.}