graph: stop the fix loop early when impl writes nothing

[alpibrupa]

Summary

• FixLoopDef.setup (SequenceNode([impl, test])) ran unconditionally — test always ran after impl, even when impl never called write or edit at all.
• Reproduced live this session: glm-5.3-flash spent its entire step budget deliberating in plain assistant text (second-guessing Lex syntax, drafting code in prose) without ever writing the target file, and the pipeline still ran test (120 steps) and verify (84 steps) against a file that was never created — 298 steps burned before the final lex check criterion in task_spec.lex caught it, the only thing that ever did.
• wrote_or_edited(events) scans a turn's trail for a cap.completed event naming the write or edit tool — same shape as verify_found_failure's existing FAIL-substring scan, reading what the dispatcher recorded rather than trusting the model's own closing text.
• run_setup runs setup's first stage with event tracking when the shape allows it (an AgentNode alone, or the head of a SequenceNode — the exact shape every current FixLoopDef.setup uses) and stops immediately, before test or the mechanical verify_program ever run, if nothing was written. Any other setup shape falls back to the previous unconditional run_node, so this only ever adds coverage, never narrows it.

Test plan

• lex check (full sweep), lex fmt --check, lex doc-sync --check, lex test (unit suite) all pass
• wrote_or_edited's own examples {} block covers write/edit/read/invoked-not-completed
• Live integration test in a throwaway worktree: a real task answerable by reading alone ("how many functions does this file define") correctly stopped after impl done — 16 steps with no test/verify stage following, versus the unguarded behavior that would have run both anyway

🤖 Generated with Claude Code