graph: draw the existing graph, and every rebuild keeps the languages the index chose (#1480)

[swapnilpaliwal-sd]

Refs #1480. A broken graph is repaired on the path that issue describes (ax_contract.ensure_graph). After this change the repair keeps the index's languages. It does not fix that issue's silent wait.

What was wrong

• axiomcode graph and the MCP axiomcode_graph tool always ran axiomcode-graph build. That ran axiomcode-build with no flags, so it detected every language in the tree.
• On a repository indexed with --lang python (or with --src / --library), that request never matched the recorded graph. So every call rebuilt the whole graph, in every language present.
  • A few stray .js or .ts files were enough to start a JavaScript or TypeScript Souffle compile.
  • The rebuild left a graph behind for each extra language. Later queries answered from it: on a C# project, a stray TypeScript graph took over context answers.
• The output was hard for an agent to use:
  • a raw Python dict of counts
  • a ../../../ page path, relative to the caller's working directory
  • advice to run axiomcode-graph build . --library, a command agents do not have
• axiomcode help graph documented axiomcode-graph build/export. The dispatcher help said the verb was "current by construction".
• The page drew only the main language's graph, even when the repository had been indexed in several languages.
• The damage outlived the call. A build with no language recorded the languages it detected as the index's own choice (lang_auto: true in the file table). After one graph, every background refresh re-detected them too: on a C# project indexed with --lang csharp, an edit plus a commit then solved csharp, typescript and javascript. Other paths reached axiomcode-build with no language in the same way: a bare index re-run "to make sure", and a query repairing a broken graph pointer (ax_contract.ensure_graph).

The change

• axiomcode graph now reads the file table the index recorded: its languages, --src and --library.
  • Graph up to date (no source file differs): the page is drawn from the existing graph. The engine does not run.
  • Graph out of date: it is rebuilt with the flags it was indexed with, before the page is drawn. The environment comes from the same code the background refresh uses (ax_fresh.rebuild_env, moved out of worker), so the two cannot drift. Only the languages the index chose are solved, and no graph appears for a language it left out.
  • No graph yet, or flags passed explicitly: the verb indexes first, as before.
• One rule for every rebuild, in axiomcode-build. When a build names no language and the graph there was indexed with an explicit --lang (lang_auto: false), the build keeps that --lang, and its --src unless one is given. ax_fresh.py chosen reads them from the file table. This covers the refresh, --fresh, a broken-graph repair, a bare index and graph. A bare index prints that it kept them, and --lang or AXIOMCODE_REINDEX=1 chooses again. A graph whose languages were detected is still detected again, so a language added later still gets its graph.
• A repository indexed in several languages is drawn in all of them, on one page. The page reads .axiomcode/lang/<lang> beside the main graph. Their ids carry the language, so none collide.
• The output is now prose:
  • how the graph was obtained (up to date, rebuilt and with which flags, or first index)
  • the counts
  • the axiomcode index … --library advice, when it applies
  • the page's absolute path, last
• These now describe the same behaviour: the axiomcode help graph docstring, axiomcode help index, the dispatcher help, SKILL.md (and its copy), the MCP axiomcode_graph description, and the README row. build and export are still accepted as older spellings.

Tests

tests/graph_verb.py has 21 checks, all passing locally. It builds real graphs.

• Indexed repository: Python under app/ with a stray .js file, indexed with --lang python --src app.
  • Current graph: graph leaves the graph file untouched (same file, same mtime), writes no .axiomcode/lang and prints prose with the absolute page path. The page holds nothing from outside --lang or --src.
  • After an edit: the graph is rebuilt with --lang python --src app, the page shows the new function, and no JavaScript graph appears.
• Near-miss control: a Java and Python repository indexed with no --lang. graph draws both languages on one page, keeps both graphs and rebuilds neither. Its detected languages are not pinned (chosen prints nothing), so a rebuild still detects them.
• Every rebuild path: C# with a stray TypeScript front end, indexed with --lang csharp. Each of these solves C# alone, with no TypeScript graph and a csharp stamp:
  • an edit plus a commit, then the background refresh
  • a bare index (which says it kept --lang csharp)
  • a query that repairs a broken graph pointer
• Surfaces: axiomcode help graph and the MCP description say the same.

Other suites run locally:

• tests/refresh.py passes for --lang csharp and --lang java, and for --lang python on its first run. Moving the refresher's rebuild environment into rebuild_env did not change its behaviour. A second Python run failed one timing check: it expects the rebuild reason "found by the timer", but under load the MCP server's own start-up kick saw the edit first. The rebuild itself ran.
• tests/surfaces.py, tests/manifests.py and packaging/copies.py --check pass.

Smoke

One project per language, each indexed fresh with --lang on 9d626cc. graph was then run with the installed build ("before") and with this branch ("after"). JavaScript and TypeScript compiles were capped at 150 s locally, so each "before" run that started one was stopped there.

project	before: time, languages solved, stray graph left	after, graph current	after one source edit
Python, 653 files	167 s (stopped): python + javascript	1 s, no engine run, no stray graph	python only: 86 s, no stray graph
Java, 116 files	2 s, up to date (one language); dict output and relative path	1 s, prose, absolute path	java only: 9 s
C#, 110 files plus a small front end	166 s (stopped): csharp + typescript + javascript, left a stray lang/typescript graph	1 s, no engine run, no stray graph	csharp only: 10 s, no stray graph

Commit on the C# project: an edit plus a commit, then the background refresh.

• Before: on the project the old graph had run on, the refresh solved csharp + typescript + javascript (JavaScript was stopped).
• After, with this branch: csharp only, 6 s, no stray graph.

The file table after graph:

• Before: csharp,typescript,javascript, lang_auto: true (Python project: python,javascript).
• After: csharp, lang_auto: false.

A table already rewritten by the old graph still reads as detected, and is not repaired by this change. Re-indexing with --lang fixes it.

Not yet measured on the full corpus or on held-out projects; the smoke numbers above are the only corpus numbers.