Port preprint-af exactly to native Go

[santoshkumarradha]

What changed

• rewrote the AgentField node in native Go with the same 19 reasoners, prompts, schemas, orchestration, parallel fan-outs, compile gate, and critique/repair loop
• removed the Python AgentField implementation and made the Go stack the only application runtime
• replaced Python-dependent parity tests with immutable prompt, persona, and discovery-schema goldens captured from the final reference implementation
• made the default docker compose up --build path launch the Go node
• bundled TeX Live, OpenCode, Matplotlib, NumPy, and Pandas for self-contained paper and figure generation
• updated the README and architecture documentation for the Go-only layout
• included the verified DeepSeek v4 Flash output PDF at artifacts/deepseek-v4-flash/preprint-af-go.pdf

Why

The earlier Go implementation had drifted from the Python behavior and the repository shipped two runtime paths. This branch restores byte-level prompt and schema parity, preserves the dynamic AgentField call graph, verifies the complete pipeline, and then removes the duplicate Python node so there is one production implementation.

Python remains in the container only as the execution runtime for generated Matplotlib figure scripts. It is not an AgentField node.

Validation

• go test ./...
• go vet ./...
• stripped static build: 8.1 MB
• docker compose config --quiet
• complete Docker build with Matplotlib PNG/PDF and LaTeX compilation smoke tests
• installed package starts and discovers exactly 19 reasoners
• full AgentField DeepSeek v4 Flash run completed with 49 tracked executions and a compiled 11-page PDF
• git diff --check

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[AbirAbbas]

Took this over since it had been sitting for ~7 weeks. No rebase was needed — the branch was still on top of main. Pushed 6 commits on top of yours; all of it is parity work against the Python being deleted, nothing redesigned.

The ones worth knowing about:

• Every str | None reasoner parameter was published as {"type":"object"}. The Python SDK resolves annotations through typing.get_type_hints, which normalises str | None to Optional[str], so _type_to_json_schema unwraps it to {"type":"string"}. target_venue, field_hint and every model argument were wrong, and examples/payload.json did not validate against the published write_paper schema.
• scan_novelty returned the harness error. Python wraps that call in try/except and degrades to safe_ai_fallback(NoveltyScan) — a crashed literature scout was taking the whole write_paper run with it.
• HarnessConfig.Model fell back to AI_MODEL; Python falls back to DEFAULT_MODEL, so setting AI_MODEL was silently retargeting the OpenCode workers too.
• DID/VC were off. The Python Agent constructor defaults enable_did=True, vc_enabled=True.
• Fidelity overlay re-scanned the list it was appending to, so the \cite{smith2024} finding suppressed the one for \cite{smith} and the repair planner never heard about the second missing key.
• Concurrent unbuildable figures both read-modify-wrote TODO.md and the later write dropped the other's bullet; append_todos is synchronous in Python so the event loop can't interleave it.
• round4 double-rounded — round(0.00005, 4) is 0.0001 in CPython, was 0.0 here.
• Unicode/JSON: truncation counted bytes where Python counts code points, and STATE.json used Go's encoder instead of json.dumps(indent=2).
• Added .github/workflows/ci.yml (go build/vet/test). The workflow removed in 3a0d5d2 gated exactly the files this PR deletes, which would have left a 5k-line rewrite with no gate at all. It's green.

Two things I left for you:

1. RunLatexmk requires a zero latexmk exit. Python accepted a non-zero exit when the PDF's mtime had changed (helpers.py:313). latex_test.go encodes the stricter rule deliberately so I didn't touch it, but it means a run Python would have shipped a PDF for can now end with an empty pdf_path — I updated the README/architecture text to say that.
2. Three things the pinned Go SDK has no equivalent for: the node description (only reachable as CLI metadata), dev_mode=True, and HarnessConfig.max_budget_usd (passed per call instead, and the OpenCode adapter doesn't enforce it in either SDK).

Not verified here: the Docker image build, registration against a live control plane, and a real LaTeX run — no Docker/WSL integration, no OpenRouter key, no TeX on this box.