ci(013): install TeX Live + fonts in real-call runner; un-skip publisher SC-006 (#210)

Real-call runner now installs TeX Live (lualatex/bibtex) + the house fonts so the publisher's llmxive.cls full compile + Zenodo Sandbox publish runs for real (SC-006 reaches 'posted'). Reverts the silent skip-on-failure: the test now fails loudly with the agent's outcome/reason if it doesn't post. CI verified: publisher first-publication + versioning both PASSED (not skipped).

Author: jeremymanning

.github/workflows/llmxive-real-call-tests.yml

@@ -51,6 +51,38 @@ jobs:
       - name: Install package and dev deps
         run: pip install -e ".[dev]"
 
+      # The publisher real-call test (SC-006) and any compile-dependent
+      # e2e test need a real TeX toolchain: the llmxive.cls full compile
+      # runs lualatex + bibtex + fontspec. Without TeX Live the publisher's
+      # `_compile_full` fails, the deposit never happens, and the test can't
+      # reach `posted`. Install lualatex + the package set llmxive.cls pulls
+      # in (fontspec, geometry, xcolor, microtype, adjustbox, tabularray,
+      # tcolorbox, caption, booktabs, listings, natbib, hyperref, fancyhdr,
+      # fontawesome5, tikz, algorithms, …) via the texlive-* meta-packages.
+      - name: Install TeX Live (lualatex + bibtex) for paper compiles
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            texlive-luatex texlive-latex-recommended texlive-latex-extra \
+            texlive-fonts-recommended texlive-fonts-extra texlive-science \
+            texlive-bibtex-extra texlive-pictures texlive-plain-generic \
+            texlive-lang-english fonts-noto-core
+          lualatex --version | head -1
+          bibtex --version | head -1
+
+      # Install the house fonts (Fraunces serif + JetBrains Mono) so
+      # fontspec resolves them and the real-call compile renders the true
+      # llmXive style. The class falls back to TeX Gyre Pagella / Inconsolata
+      # via \IfFontExistsTF when these are absent, but the published PDF
+      # should match production.
+      - name: Install llmXive house fonts
+        run: |
+          mkdir -p "$HOME/.fonts"
+          cp papers/.style/fonts/*.ttf "$HOME/.fonts/"
+          fc-cache -f >/dev/null
+          fc-list | grep -iE "fraunces|jetbrains" | head || \
+            echo "WARN: house fonts not registered (class will fall back)"
+
       # FR-031 / T114: Per-PR pre-checks BEFORE running real-call tests.
       # If any of these fail, the gate fails fast — saving real-LLM minutes.
       - name: Prompt existence check (every prompt referenced by registry exists and parses)

tests/real_call/test_publisher_zenodo_sandbox.py

@@ -137,29 +137,18 @@ def test_publisher_sandbox_e2e_first_publication() -> None:
         )
         result = agent.run(ctx)
 
-        # Stage advanced.
+        # Stage advanced to `posted`. The CI runner installs TeX Live + the
+        # house fonts (see the workflow) so the publisher's full llmxive.cls
+        # compile + real Zenodo Sandbox publish runs for real here, exactly
+        # as it does locally. If it does NOT reach `posted`, that's a genuine
+        # failure — surface the agent's outcome + reason in the message
+        # rather than skipping.
         proj = project_state.load(pid, repo_root=_REPO)
         assert proj is not None
-        # The publisher is deterministic and verified end-to-end LOCALLY
-        # (~10s: full llmxive.cls compile + real Zenodo Sandbox publish).
-        # When the real publish can't complete in THIS environment, the
-        # agent records a FAILED outcome and leaves the stage at
-        # paper_accepted. The two environmental gaps that hit the CI
-        # real-call runner are: (1) no TeX Live / house fonts for the
-        # llmxive.cls full compile (the runner installs neither — the
-        # implementer e2e fixture deliberately uses \documentclass{article}
-        # to avoid this), and (2) Zenodo Sandbox being transiently
-        # unreachable. Mirror the missing-creds skip above and treat an
-        # un-posted result as a skip-with-diagnostic rather than a hard
-        # failure; publisher LOGIC is covered by tests/unit/test_publisher.py
-        # and the real sandbox path is exercised locally.
-        if proj.current_stage != Stage.POSTED:
-            pytest.skip(
-                "sandbox publish could not complete in this environment "
-                f"(outcome={result.outcome.value!r}, "
-                f"reason={result.failure_reason!r}) — verified locally."
-            )
-        assert proj.current_stage == Stage.POSTED
+        assert proj.current_stage == Stage.POSTED, (
+            f"publisher did not reach 'posted' "
+            f"(outcome={result.outcome.value!r}, reason={result.failure_reason!r})"
+        )
 
         # Publication metadata written; DOI is sandbox-prefixed.
         pub = pub_state.load(pid, repo_root=_REPO)