check: incremental checking for Python, C/C++ and every build tool

[rs545837]

sem check now skips or narrows work for every language it detects, and in every case we measured the verdict and errors match a full run.

What changed

• Shared layer. A checker carries the base's passing verdict when nothing it reads changed, such as docs, CI config or another language's sources. A TypeScript edit no longer reruns cargo or Gradle.
• Python. pyright rechecks only the files whose imports can reach the change and keeps the rest of the base's errors. An edit inside the body of a function with a return annotation rechecks only that file. mypy carries.
• pytest. Reruns only the test files that can load a change, keeping the other files' results, failures included.
• C/C++. Compiles (-fsyntax-only) only the compile_commands.json units that include a changed file.
• Cargo, Gradle/Maven, dotnet, SwiftPM. Carry their verdict. Cargo treats every file as an input when the crate has a build script, a proc macro or include_str!.
• Configured commands. Carry when none of their inputs changed.

When 70% or more of a large scope is affected, pyright runs in full instead, since one full run is faster.

Measured

Project	Change	Full	Incremental
rich (213 files)	body edit in a function 33 files import	3.45s	0.94s, 1 file, identical 336 errors
rich	mid-level module	3.8s	1.3s, 11 files
rich	docs edit	~7.9s	0.47s
zlib (34 units)	broken zutil.h	34 units	18 units, same errors
zlib	inflate.c edit	34 units	2 units

Not covered yet

• Gradle, Maven, dotnet and SwiftPM were tested only with a stand-in command, since the test machine has no JVM or .NET.
• Java, Kotlin, C# and Swift carry or run in full, with no per-file scoping.
• Early cutoff covers module-level annotated functions only, not methods.
• C/C++ headers generated outside the repo aren't tracked, so run --full after regenerating them.
• Set foreignInputs: true for a build that reads other languages' sources.

Tests

16 unit tests and 14 check_cli tests pass, including 4 new end-to-end tests for command inputs, cargo carry, C units and pytest selection.

[inspect-review]

inspect review

Triage: 112 entities analyzed | 0 critical, 0 high, 46 medium, 66 low
Verdict: standard_review

Findings (0)

---

Reviewed by inspect | Entity-level triage found 0 high-risk changes