feat(context): name the tests a call edge reaches, as a ready run_task filter

[atlas-from-plumb]

What

context_for_task's affected-tests estimate named only packages, e.g. ./internal/tools/... — 1808 test(s). In a directed-use study every agent then wrote its own -run filter or ran the whole package. Test runs were the slowest step of their loop, about 20 per six runs.

Since #640/#646 resolve same-package calls, a test that calls a seed is reached by a real call edge. That tells it apart from a test that only sits in the package. Under a package row, one line now names up to 8 of those tests, as a ready run_task call:

  ./internal/tools/... — 1808 test(s); holds a seed
    run first: run_task(slot:"test", target:"./internal/tools/...", run:"^TestMoveSymbol_(AllowsMoveBetweenSameGOOSFiles|…|RefusesCommentVsFilenameAsymmetry)$") — 8 of 16 test(s) that call a seed or a declaration next to one, directly or through callers (index call edges); 8 more are not named

• What is named: only Test/Example/Fuzz functions that go test -run selects, and only tests reached by an edge. Co-located tests are counted, never named.
• Order: tests that run a seed come first, then tests reached only through a callee or sibling of it (marked Aside). Within each group, fewest hops first, then name.
  • Neighbours walk one hop out, callers of a seed first, so the order does not depend on the pack's ranking.
• The filter: it writes the common prefix once and is cut to what validateRunFilter accepts. The rest is counted.
• Stale index: names none. The package rows stay.
• Pack budget: the line is classAffected at prioAffected, so it never displaces a body or a top related entry.
• Unchanged: the rows themselves, and topology_affected's own rendering (fields only).

Measured

On the #571 task (base 8ac6b43, seed buildMovePlans, intent change, 32k):

• The line names 8 of 16 tests, all TestMoveSymbol_*. That is exactly what the study's agent picked by hand (-run 'MoveSymbol').
• The first cut named tests of a callee first (TestApplyTextEdits_*). That led to the Aside ranking.
• The pack grows from 27,162 to 27,623 bytes, and a diff of the two packs shows only the added line.

Tests

• C07 (the oracle case pinning "one row per package"): the rows are unchanged, plus the exact run-first line.
• Edge-only naming: C01's cart tests reach the seed only through a name match, so none are named. C02's pricing test is reached by an edge and is named.
• No Go target: the line gives the names only.
• Unit rules: runnable names, dedupe, caps, stale, order; the filter fit against validateRunFilter; runFilterFor checked with Go's regexp in both directions, including a rune-split prefix; betterPath; inwardHops; the neighbours' walk order.
• End to end: a 6-test fixture over a real index, whose names make each wiring decision reorder the line when it is wrong.

Review and verification

• Review: an independent peer review (warm-cobra) returned CONTINUE, with no blocking findings. Its three spot mutants were all killed. Its NIT ("through one caller") is folded in.
• Mutation (plumb mutation_test): every valid mutant is killed except one, the stale-index argument at the call site. No fixture drives a live store to "stopped"; the stale rule itself is unit-pinned. Survivors on earlier passes led to the end-to-end fixture and the walk-order test.
• GOWORK=off make verify-full: green except internal/quality/golangcilint. That package collided with another session's golangci-lint ("parallel golangci-lint is running"), and it passes alone on a re-run. A clean re-run of verify-full is noted in the comments.

Part of PLAN-543 (directed use on real bugs), following the arm D finding that the pack was paid for twice.

🤖 Generated with Claude Code

[golimpio]

Approved under the owner's standing delegation. Bar met on head 2bcfa1d: independent peer review (warm-cobra) CONTINUE with no BLOCKING and its spot mutants killed; plumb mutation_test with every valid mutant killed (one disclosed residual); GOWORK=off make verify-full exit 0 read from its exit file, not dirty; CI 9/9 SUCCESS.