Loop records cost_usd: 0.0 — PiBackend uses text mode, parser expects JSON

[Svaag]

Problem

Every production loop run on loop records cost_usd: 0.0 (e.g. canary #235 → PR #239), so token-capital/cost accounting is effectively blind.

Root cause

• PiBackend.default_command = ("pi", "--print", "{prompt}") → text mode.
• SubprocessBackend._parse_harness_output does json.loads(stdout) and returns {} on non-JSON, so total_cost_usd / usage.{input,output}_tokens / num_turns are never extracted.
• The saved prod pi transcript is plain markdown — no JSON, no cost fields.
• pi does support structured output: --mode <text|json|rpc> (confirmed via pi --help). ClaudeCodeBackend already uses --output-format json.

Fix direction

1. Switch the pi command to JSON mode (pi --print --mode json {prompt}), either in PiBackend.default_command or model-policy.production.yml backends.definitions.
2. Verify pi's JSON schema matches the parser's expected keys (total_cost_usd, usage.input_tokens, usage.output_tokens, num_turns, is_error, result); add a small field-mapping in _parse_harness_output if pi differs.
3. Confirm the worktree diff/PR-result flow is unaffected (diff is captured from the worktree, not stdout — should be safe).
4. Add a backend test asserting cost/usage parse from a representative pi JSON payload.

Impact

Blocks the token-capital baseline (private evals phase). Not a safety issue — runs still complete and the daily ledger still caps run count; only the cost ledger reads 0.