autosetup: anchor on the project whose artifacts these actually are

[shellygr]

Follow-up to the artifact-directory work in #169, from the same sweep that produced #186.

What goes wrong

find_build_config_dir walks up from the contract and stops at the first directory that holds a
build config and has its artifact directory on disk. That test is satisfiable by a directory that
never built anything, because two configs can name the same physical directory:

foundry.toml            src = 'pkg/src'   out = 'pkg/out'     <- this one ran
pkg/foundry.toml        src = 'src'       out = 'out'         <- this one did not

pkg/out exists and pkg/foundry.toml is a config, so the walk stops at pkg/. Everything
downstream then reads those artifacts in the wrong frame: the extractor builds its project-file set
relative to pkg/, while the artifacts record pkg/src/Widget.sol relative to the root. No source
path matches, the scope filter drops every artifact, and the run dies with "No project-local logic
contracts found in FoundryManager build output" — from a build that had no errors at all.

In the sweep this cost one project 110 matching artifacts, and the class it lands in is the run's
largest.

The change

The artifacts say which project wrote them: Foundry records metadata.settings.compilationTarget,
Hardhat records sourceName, and both are relative to the project the build ran in. A candidate
directory now has to pass that test — at least one recorded source path has to resolve inside it —
so a directory holding somebody else's artifacts is skipped and the walk continues to the one they
belong to.

This extends #169's own principle rather than reversing it: decide from what is on disk, not from
the directory layout. A genuinely nested project that built its own artifacts records paths relative
to itself, passes, and still anchors where #169 put it. Artifacts with no recorded source (older
Foundry, stripped metadata) answer yes, so nothing moves for them.

Sampling is capped at 20 artifacts per candidate, so the check costs a few JSON reads on a path that
already walks the tree.

Tests

tests/test_project_dir.py gains three cases: the shared-out-dir layout above (fails without the
change), a nested project that really did write its own artifacts (must keep anchoring there), and
artifacts with no metadata (must not move). 24 pass in that file, 111 across the parser and
workaround suites.

The repo's full run fails 14 for me and 15 on a clean tree, all in integration tests that vary
between runs; none of them is unique to this change.

Draft, and worth reviewing alongside #186: both come out of the same sweep, and this one touches the
path #169 introduced.

🤖 Generated with Claude Code