Test/lint/typecheck gate runner with structured per-repo results

[Ryan-Atkinson87]

Spec: §4.2 step d (algorithmic gate), §16.3 repos.commands, §16.3 limits.test_fix_attempts.

The deterministic gate that runs a repo's full test suite, lint, and typecheck and returns a structured result the engine feeds back to bounded AI fix attempts. Pure engine logic — no AI.

Acceptance criteria

• Reads each repo's commands.test / commands.lint / commands.typecheck from config (project.yaml Pydantic schema and fail-fast loader #4) and runs them in the repo working directory.
• An omitted/empty command is treated as skipped, not failed (§16.3); a not-yet-established frontend command is reported as "not established", not a silent permanent skip (per the CLAUDE.md tech-stack note).
• Returns a structured result per gate: pass / fail / skipped, exit code, and captured stdout+stderr, suitable for feeding back to an AI fix attempt.
• Each command has an explicit failure path — both a timeout and a non-zero exit produce a result; never a hang or a silent pass.
• Does not itself implement the retry loop — it exposes results; the bounded limits.test_fix_attempts loop is wave-loop orchestration (Phase 3). The result shape must carry enough detail for that loop.
• Unit tests cover a passing gate, a failing gate (captured output), a timeout, and a skipped/omitted command, using fake commands.
• Lint, format, and typecheck pass.

Depends on: #1, #4

Ordering: independent of the branch-lifecycle chain (#9–#11); can be built in parallel.