mutate: the scheduler is the package's, and held by tests

[aledbf]

What runs a change's mutations was ~330 lines of cmd/mutate/main.go, which has no tests, so the gate never broke its own concurrent half. It spreads the cap over functions, builds each package's edits into one binary or alone, asks the tests that reach each line and then the importers', and counts the results. That is now mutate.RunAll and mutate.Spread in internal/mutate/runner.go. main keeps the flags, the plan, and the cost and stale reports, and stays without tests, so the fixtures that use it as "a package with no tests" still hold.

Behaviour:

• The lines said as each outcome arrives go to the io.Writer RunAll is given (os.Stdout from main). Diagnostics stay on stderr.
• A failure to prepare an edit (MkdirAll, Overlay) was os.Exit(2) from inside a goroutine. It is now RunAll's error, and main exits 2 on it. The difference: the other edits finish first.

New tests, each on a throwaway module:

• Spread's rounds and caps.
• Every edit counted as the tests answered: refused, survived in order, a package with no tests named rather than counted as a survivor, an int64 literal built alone and not reported as slow.
• A package with no tests asked through its importers.
• A test broken regardless refuses nothing, both in the binary and alone. This is what coverage protects against.
• With no coverage, every test is asked (internal test of built).
• An edit that cannot be asked is an error.

mutate-exempt on three lines, each a cost or a stderr note rather than an outcome: whether coverage is measured when nothing is built, and two stderr diagnostics.

• task test (coverage 62.8% → 71.2%), task lint: green
• task mutate: 35 refused, 0 survived (this run is also the new code running itself)

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}