index: publish each language's graph as soon as it is solved (#1555)

[swapnilpaliwal-sd]

Fixes #1555

What was wrong

A repository in several languages is solved one language after another, and nearly all of each step is that language's Souffle compile. axiomcode-build published nothing until the whole engine run ended, so the main language's graph sat finished while another language compiled, and every query made meanwhile waited on the build lock. The issue measured a C# graph written at 131 s and first queryable at 5,304 s.

The change

• bin/axiomcode all --progress <file> appends <lang> ok <graph> or <lang> failed as each language finishes. Without --language, the largest language (by IR size) is solved first.
• axiomcode-build solves the main language first and the others by file count. It runs the engine in the background in its own process group, which is stopped along with the build. The main graph is published the moment it is solved (index, stamp, file table, baseline, pointer). Every other language's graph moves into .axiomcode/lang as it arrives. .axiomcode/out/building lists the languages still to come. An engine without --progress is published at the end, as before. If the main language fails, the build still fails and the previous graph stays.
• ax_fresh: status and query read the pending list. An answer from a current main graph comes back at once, with one graph refresh: line naming the languages it cannot see yet, or the ones whose code comes from their previous graph. build-seconds records the time to the main graph. On a clean tree with pending languages, the stale-while-revalidate query path now carries the note too.
• ax_contract.ensure_graph: a first query waits until graph.sqlite exists, not until the whole build has finished. The build writes to a log that is copied to stderr, so stdout carries only the answer.

Tests

• New tests/publish_order.py, 19 checks, all passing. The repository is C# with Python and Java beside it. A wrapped engine passes on the main language as soon as it is solved and holds back the rest. The checks cover: publish order, a query answered while the others are held (with the note), status, a Java name refused at once, the finished build, a rebuild, a failing main language, a first query with no graph, and a one-language control with no pending list and no note.
• The same test against origin/0.1.4 fails its first two checks. graph.sqlite never appears while the other languages are held (the build is still running, with no graph), and the solve order is csharp java python.
• tests/run.py: csharp 14/14, python 71/71, java 119/119. latency.py --no-engine 18/18, surfaces.py 8/8 verbs, directive.py 12/12.
• No JavaScript or TypeScript suite was run locally; CI covers them.

Smoke (real copies, fresh index)

• A 461-file C# library with 4 JavaScript files. Before, with the installed build, the C# solve finished at 11 s, but graph.sqlite appeared only at 291 s, after a JavaScript rule compile that then failed. After, the graph appeared at 25 s and impact returned 30 rows at 69 s (44 s of that was a one-time query-rule compile) with graph refresh: the javascript graph is still being built, while the JavaScript compile was still running.
• Control, a 154-file C# library in one language: graph at 14 s before and after, no note, and identical impact answers except the sampled list of untyped-receiver callers. That sample differs between any two builds (pre-existing).

Deep scan deferred: the held-out corpus measurement was not run for this change.

This supersedes the local work branches fix/1555-index-per-language, fix/1555-index-per-language-availability, fix/1555-lang-availability and fix/1555-per-lang-graphs (never pushed); it continues fix/1555-per-language-graph.