fix(pipeline): register command-step output_artifacts for downstream injection

[nextlevelshit]

Closes #1490.

Problem

type: command steps that wrote files declared in output_artifacts never populated execution.ArtifactPaths[step.ID+\":\"+art.Name]. The persona/adapter dispatch path called writeOutputArtifacts after the adapter returned (executor.go:3572-3583); the command dispatch path returned directly from executeCommandStep to validateStepContracts without ever registering the path.

Downstream memory.inject_artifacts lookup (executor.go:4555) missed on every command-step source, fell through to the stdout-fallback branch, and (since command scripts almost never echo the artifact bytes to stdout) wrote a 0-byte file under .agents/artifacts/<as-name>. The receiving persona then read empty input and short-circuited. The pipeline reported success.

Real-world hit

ops-pr-respond 1472 (run ops-pr-respond-20260428-182026-1f67) posted "No findings" to PR #1472 — even though filter-scope (a command step) had produced 15 valid security findings. The planner persona received empty scoped_findings. Pipeline status: green. Deliverable: wrong.

This is exactly the failure mode "green ≠ correct" — masked from CI, masked from web UI, only caught by reading the actual JSON outputs after the run.

Fix

Call writeOutputArtifacts(execution, step, workspacePath, nil) after executeCommandStep returns successfully on both dispatch paths:

• executeStep at ~1786 (the linear walker)
• the executeGraphPipeline callback at ~1318 (the DAG walker)

writeOutputArtifacts already handles the "script wrote the file, don't overwrite" case via os.Stat(artPath) == nil (line 4715) — for command steps this is always the branch taken.

Test

TestCommandStepOutputArtifactsRegisteredForInjection reproduces the bug shape: a command step writes JSON to a declared output_artifact, then a downstream persona consumes it via inject_artifacts. Without the executor.go change the injected file is 0 bytes; with the change it contains the JSON the command produced.

I confirmed locally: git stash the executor.go fix → test fails ("0 is not greater than 0"). Restore → passes.

Test plan

• go test ./internal/pipeline/... — green
• New regression test passes with fix, fails without
• Re-run ops-pr-respond 1472 after merge — expect substantive PR comment listing real findings