fix(eval): strip leading slash commands and surface replay errors

[theonly1me]

What changed

• Strip leading slash commands such as /plan, /goal, and /boost from extracted prompt text while preserving Unix file paths.
• Detect synthetic local command exits in parseClaudeStream and mark isError: true so failed command invocations are not treated as valid zero-turn runs.
• Parse Claude CLI error arrays and subtype flags into errorMessage on EngineRun rather than falling back to a generic message.
• Capture process stderr in runClaudeCode when an engine subprocess exits with a non-zero exit code.
• Extract single-session replay into src/eval/replaySession.ts to keep src/eval/run.ts within line count limits.
• Log prompt snippets and failure reasons when a session is skipped, and record skippedSessions in EvalReceipt.
• Report skipped session counts in the shadowclone eval terminal summary.

Why

Prompts recorded from Antigravity frequently begin with slash commands like /plan. When piped to claude -p, Claude Code CLI treats leading slash tokens as interactive commands, printing <local-command-stdout>/plan isn't available in this environment.</local-command-stdout> and exiting with code 0. Because is_error was false, parseClaudeStream treated this as a successful run with zero actions, yielding 0% delta with an artificial 0.5 planning score. Stripping UI slash commands ensures the prompt text executes as the intended task.

In addition, replay errors were swallowed across multiple layers: runClaudeCode discarded stderr, parseClaudeStream ignored the errors string array on the result payload, and runEval printed a single generic line without the prompt or failure reason. Capturing errors across the runner and recording skippedSessions in receipts makes replay failures actionable.

How to verify

bun test src/eval/prompt.test.ts
bun test src/engine/parseClaude.test.ts
bun test
bun run check
bun run ./src/cli/index.ts eval --sessions 1 --max-budget-usd 0.50 --yes

Live validation on session 001bb52a-41b1-4c32-94aa-dee90ad403ae confirmed that /plan was stripped, Claude executed real tool calls (Bash, ToolSearch, WebFetch), and the session was scored in receipt 0b4a786f-b641-4bf5-bf80-f88910dd67f1.json.

Data handling

• Any new capture source has an opt-in flag and a README entry in this PR
• Everything that reaches the network passes redactSecrets
• No raw capture in a log line, an error message, or a test fixture
• A test exercises the real entry point, not just the redaction function