feat: configurable per-tool build/test commands + opt-in go-test runner

[DatScreamer]

What & why

Every tool runner hard-codes its argv (go build ./..., golangci-lint run ...), there is no
config override, and there is no full-suite test tool. A Go project whose dev server
rebuilds with build tags (e.g. -tags 'dev devaccounts') cannot have Interlinked run that
exact command — so Interlinked's checks compile into a separate Go build cache, recompiling
per surface (measured ~30–84 s cold vs ~6–21 s warm on the motivating Go workspace). This
branch lets a project declare its build/lint/test commands once and have every Interlinked
surface honor them.

Changes

1. New two-tier config — .interlinked/tool-commands.json (committed) +
   .interlinked/tool-commands.local.json (gitignored), keyed by check/config names
   (go_build, go_test, golangci_lint, …). A full command array wins; otherwise
   base_args replace the runner's default scope (./...) so Go flag ordering stays correct.
   Trust split mirrors merge.ts: the committed team tier may set base_args +
   timeout_ms (flags for a fixed binary / a bounded cap); command/env stay personal-tier
   (interlinked doctor reports violations). No shell interpolation — argv is written as
   executed (-tags 'dev devaccounts' is ONE token ["-tags","dev devaccounts"]).
2. Runners honor the override — go-build, golangci-lint, and the new go-test tool
   spawn the configured argv, sync + async variants (src/harness/check-engine/).
3. New go-test catalog tool — project-wide, opt-in (auto-runs only when the project
   configures go_test, or on explicit --only go-test / --tools go-test /
   verify --only go-test), generic + Go test parsers producing [proven] per-test findings.
   Wired into check --only/--tools/--report, TOOL_IDS, and the ALL_TOOL_IDS drift guard.
4. affected_tests carries the tags — PostToolUse on a .go edit runs
   go test <tags> -count=1 ./<pkg>, replacing the full-suite ./... scope with the touched
   package (src/harness/quality-checks/).
5. Skills + docs updated (interlinked-verify, interlinked-setup, router; design doc
   docs/interlinked-test-runners-plan.md).

Verification

• npm run typecheck — all feature files clean; only the pre-existing WIP error
  (go-halstead-walk.ts, untracked) remains on this checkout.
• Affected suites (check-engine, verify, check, quality-checks, dispatchers): 104 files /
  2796 tests pass; two mutation-kill suites pass; new unit/integration tests for the
  resolver, runner argv+parser, and verify-tools resolution.
• Full npm test: one regression this change introduced (a check-command available-checks
  list pin) was fixed before opening the PR; the failing files on a full run fail identically
  on pristine origin/main (Node 20.19.2 lacks node:sqlite for the mutation-journal
  suites, plus known metrics/compiler-gate env-timing flakes).

Acceptance (isolated GOCACHE, re-confirmed after rebase)

Measurement	Command	Cache state	Wall
Before (stock install)	check --only go-build	untagged, no shared cache	8.7 s
After (branch)	check --only go-build	tagged, dev-warmed reuse	6.0–6.8 s (clean)
After (branch)	check --only go-test	tagged, dev-warmed reuse	21–26 s (per-test findings)
Reference	go build -tags 'dev devaccounts' ./...	fresh (no reuse)	29.3 s
Reference	go test -tags 'dev devaccounts' ./...	fresh (no reuse)	84.4 s

After-runs hug the warmed-tagged-cache baseline (direct warm repeats 5.5 s build / 23 s
suite), proving the tagged command reuses the dev cache instead of recompiling.

Known v1 gaps (deferred, documented in the plan)

• verify --json has no per-runner go-test section yet (fixed-shape schema).
• Coverage-gate resolver wiring (js/ts/python only).
• PostToolUse honors the config after interlinked harness restart in existing repos.

Closes #7