Artifact-aware documentation generation (build, CI, container, manifest, config files)

[anhnh2002]

Problem

CodeWiki documents only what is in the dependency graph, and the graph is built from files whose extension is in CODE_EXTENSIONS. Dockerfiles, CI workflows, Makefiles, package manifests, configuration, schema and script files are dropped before a single Node exists. As a result no generated wiki describes how a system is built, packaged, shipped, configured or tested. Across seven benchmark wikis there was not one mention of GitHub Actions.

What this PR does

A second analysis pass (dependency_analyzer/analyzers/artifact.py) runs after the language analyzers over the same file tree and emits artifact nodes that flow through the existing clustering and agent machinery:

• Nodes. One node per artifact file (<path>::<file>, source = capped 16 KB head) plus unit nodes for CI jobs, Dockerfile stages, Makefile targets, package.json scripts and pyproject/setup.cfg entry points. Classes: manifest, build, container, ci, packaging, test_infra, schema, config, script (and opt-in prose).
• Edges. Artifact -> code references from COPY/ENTRYPOINT/RUN, CI run: lines, python -m, make <target>, npm run <script>, module:function entry points and compose build contexts. Only fully resolved ids are emitted (the parser falls back to name matching otherwise).
• Caps. Per-file head, 40 files per class, total token budget (default 200k) filled in class priority order. Lock files and binaries are never read. README and docs/ stay out unless --with-prose.
• File walk. ARTIFACT_WHITELIST lets .github/workflows, *.ini, bin/*.sh, *.gradle survive DEFAULT_IGNORE_PATTERNS; user --exclude and .gitignore still win. Artifact names added to DEFAULT_INCLUDE_PATTERNS.
• Leaf selection. artifact is always a valid leaf type; artifact -> code edges no longer prune code leaves at the 400-leaf reduction.
• Clustering. Artifact files are tagged (artifact: <class>) in the cluster prompt with an instruction that they are essential. ensure_artifact_module inserts a top-level Build, Deployment and Configuration module when clustering placed fewer than 80% of artifact nodes (both run() and the CLI adapter path).
• Prompts. Fence language now falls back to text instead of raising KeyError for Dockerfile, Makefile, .yml, .toml, .cfg, .ini. Artifact groups inline their capped source instead of re-reading the whole file. A compact <REPOSITORY_ARTIFACTS> index is appended to module prompts and to the repository overview prompt; USER_PROMPT itself is unchanged so the MCP prompt server keeps working.
• Tools. str_replace_editor view on a directory no longer hides .github and quotes the path.
• Config. Node.artifact_class; temp/artifact_index.json; flags --artifacts/--no-artifacts (default on), --artifact-token-budget, --with-prose, --artifact-exclude (also agent_instructions.artifact_exclude).

Test plan

• pytest -o addopts="" tests/ -q: 99 passed (60 existing + 39 new in tests/test_artifact_analyzer.py: classifier table, whitelist, nodes and units, edge resolution, caps, leaf selection and pruning, user prompt, fallback module).
• End-to-end on pypiserver (~10k LOC): 41 artifact nodes, 79 edges; clustering placed 41/41 artifacts and produced a build_deployment_and_configuration module with CI/container/tooling/manifest sub-docs; the overview gained a "How it is built and run" section. Baseline with --no-artifacts behaves as before (same graph, 0 artifact nodes). No errors in either run.

Notes for reviewers

• --no-artifacts restores the previous behaviour exactly, so the change is one flag away for ablation.
• Edge heuristics are per-format and shallow; they connect artifacts to code for clustering and cross-links but are not a call graph.
• Existing outputs resumed from a cached first_module_tree.json will not gain the fallback module; use a fresh output directory.