Classify benchmark gate failures separately from orchestration failures

[chubes4]

Problem

A Lab-offloaded fixture benchmark can execute its orchestration successfully and produce complete structured evidence, then correctly fail its quality gate. The operator-facing failure path frames that outcome like remote execution failure and directs the operator toward raw output-file inspection instead of presenting the typed gate result and its evidence commands.

Run 705aad31-551f-4ab6-8db3-a6e17e903b15 demonstrates the distinction:

• Lab handoff and child recipe completed.
• Child recipe exit_code was 0.
• Child output was parsed successfully.
• The run retained benchmark results, summary, finding packets, screenshots, DOM snapshots, and visual reports.
• Benchmark status was fail because failed_fixture_count was 1, with explicit runtime_execution_failed and visual_mismatch categories.

This is an expected workload gate rejection, not a transport, runner, or orchestration failure. Treating both as one generic remote failure sends operators into manual artifact spelunking and obscures whether retrying infrastructure would help.

Context: Automattic/blocks-engine#659 and child run 705aad31-551f-4ab6-8db3-a6e17e903b15.

Expected

Homeboy reports separate typed outcomes for:

1. orchestration/transport failure,
2. workload execution failure,
3. completed workload with rejected result gate.

A rejected benchmark gate should summarize failed metrics/categories and provide direct commands for the result, summary, and finding artifacts.

Acceptance criteria

• Runner and bench envelopes preserve a distinct result_gate_failed or equivalent outcome.
• A child process exit 0 plus parsed benchmark result cannot be summarized as remote orchestration failure.
• CLI output displays failed gate metrics and top typed failure categories.
• CLI output provides resolvable homeboy runs artifact get commands for result, summary, and finding packets.
• Retry guidance is limited to transport/orchestration failures; result-gate failures recommend evidence inspection or candidate repair.
• Tests cover transport failure, nonzero workload exit, malformed result, and successful execution with a rejected result gate.

AI assistance

• AI assistance: Yes
• Model: OpenAI GPT-5.6 Sol
• Tool: OpenCode
• Used for: run inspection, outcome classification, duplicate search, and issue drafting