[CaffeinatedCoder]

Closes #268

Problem

Under decision 113 every lane built and tested the whole solution unless the change was prose throughout. #266, one
tool, its tests, docs and a line of .gitignore, ran 12m55s on Linux and 13m32s on Windows in run 41215127177, and
packed and smoke-tested because .gitignore was unlisted. The tests that could notice it were the BashGuard tests and
the conventions.

Change

• tools/CiLanes: a third suite, reach. When every changed path is prose, .gitignore, or inside a solution project
  under tools/ or test/, select answers suite=reach and projects, the projects the change reaches: a project
  reaches its own directory, what its items and imports name outside it (a linked source, a ProjectReference), and,
  through each reference, what that project reaches; prose adds test/Prose.slnf. Anything else, or any file it
  cannot read (the solution, a project file, the prose filter), runs the whole solution. select always writes
  projects ([] outside reach). New ci-lanes suite <path>... answers each path alone.
• .gitignore gets the class ignore list: Linux alone, narrowed to the conventions. No test reads it (the
  conventions list the index), and dotnet pack does not. .gitattributes stays unlisted.
• The gate accepts reach only beside no pack and no macOS leg, with one or more distinct tools//test/ project
  files; it refuses projects beside prose/full and a missing or malformed projects.
• dotnet.yml: the changes job exports projects; the test legs write it as test/Reach.slnf (gitignored) and
  restore, build and test that when the suite is reach.
• A run-time read by path is declared with <ReadsByPath Include="..." />: the conventions declare the root, the
  oracle tests test/ (they find each claim's test among every test project's sources). Measured: no other test
  project reads another by path.
• Conventions: TestReachTests evaluates every tools/ and test/ project with MSBuild and fails a build input that
  would not run its test project; scans test sources for literals naming another project, or a scan of test//tools/,
  that the reader does not reach; and fails a narrowed suite without the conventions. ReleaseWiringTests holds the
  new workflow step.
• design/adr/0127, amending 0113; AGENTS.md lists ci-lanes suite.

This PR changes tools/CiLanes and the workflow, so its own run is the whole solution on Linux and Windows with the
pack, as before.

Tests

• SelectionTests: rows for a tool-only change, a test-project-only change, a linked source, transitive reach through a
  referenced test project, the build: run parallel work in worktrunk worktrees beside the repository #266 shape (tool, tests, docs, .gitignore), .gitignore alone, tools/Shared, a mix
  with src/, paths outside every solution project, an unreadable project file, and suite. Each failed against
  113's selection first.
• GateTests: a reached suite passes on Linux and on Linux and Windows; fails cancelled, skipped or failed legs, the
  pack, macOS, empty/out-of-place/duplicate projects, projects beside the other suites, and a malformed or missing
  projects. The new rows failed against the old gate first.
• Seeded and caught: ci-lanes skipping linked sources (TestReachTests and SelectionTests fail), skipping reference
  recursion, dropping the ignore-list class; removing the oracle tests' or the conventions' ReadsByPath; a read of
  tools/PrHygiene added to the BashGuard tests. Each failed the test that claims it, then passed when restored.
• Locally: test/Reach.slnf written as the leg writes it, from build: run parallel work in worktrunk worktrees beside the repository #266's paths, builds. The whole solution: 24128
  passed; the 12506 failures are all the SQL Server host (SqlServer.cs:84, SqlServerFactHost.cs:29), whose
  container does not start on this arm64 Mac without TYPEKIT_MSSQL_IMAGE; CI's Linux leg runs them.

🤖 Generated with Claude Code