Switch Stage-2 review harness from Claude Code to pi

[bockets]

What & why

Stage 2 of the BiggiePockets review (verify & synthesize) ran on anthropics/claude-code-action. Claude Code translates OpenRouter's usage into its own schema on the way to the terminal result message, so the cost OpenRouter reported regularly did not survive to the Datadog span — that's the pricing data loss this PR fixes.

The stage now runs pi (@earendil-works/pi-coding-agent), a plain npm CLI: no GitHub app, no OIDC, no per-repo install. It's authenticated to OpenRouter with the same OPENROUTER_API_KEY the Codex stage uses (env var — nothing to configure).

Verified live before shipping:

• pi resolves deepseek/deepseek-v4.1-flash in a fresh config dir (pinned via models.json — the model is absent from the catalog pi ships and the live catalog refresh is a background fetch).
• Headless --mode json pass executes read/grep/glob/bash/write tools, writes verdict.json, and exits cleanly.
• Usage is recorded on every pass: token counts + cost.total, computed from the model's OpenRouter list rates — the same rates OpenRouter bills against (confirmed: 17,515 input tokens x $0.019/M = $0.0003328, matching OpenRouter exactly). That's the cost that used to die in Claude Code.

What changed

• Workflow (.github/workflows/biggiepockets-review.yml): claude-code-action step -> Install pi + Configure pi (copies scripts/pi/models.json into PI_CODING_AGENT_DIR) + one headless pi --mode json pass with the same read,grep,glob,bash,write allowlist and --no-context-files --no-skills --no-extensions for hermetic running. The step fails when verdict.json is missing (pi can exit 0 on a retried-but-failed session), so Re-run failed jobs still retries just this pass.
• New scripts/pi/models.json: pins the OpenRouter provider + Stage-2 model and its OpenRouter list rates (fetched from OpenRouter's API at pin time: input $0.15/M, output $0.60/M, cache-read $0.003/M).
• scripts/llm-usage.py: new pi pass reading the last finished assistant message's usage from pi-output.jsonl (input/output/cacheRead/cacheWrite -> input_tokens/etc., cost.total -> total_cost). Legacy claude/codex parsers kept.
• scripts/review-diagnostics.py: parses pi's event stream (terminal message only — tool calls/args/results never published, secrets redacted) and attaches a redacted stderr tail.
• Telemetry renamed with the harness: span claude.synthesize -> pi.synthesize, CLAUDE_MODEL -> PI_MODEL, tag claude_model -> pi_model. If you filter Datadog on the old span name, this is the change to know about.
• README: "Install the Claude GitHub app" -> "Install nothing" (the workflow installs the npm package on the runner); workflow-permissions section drops the OIDC requirement (restrictive "Read repository contents" is now fine); cost-tracking section documents the pi pass and the pinned model config.
• Tests: +9 (pi usage parsing, pi diagnostics, stderr redaction); 38 total, all passing. actionlint clean.

Things to know

• Stage-2 model roll = edit PI_MODEL and scripts/pi/models.json together (id + rates).
• Cost reporting now trusts pi's recorded usage rather than whatever survived Claude Code's schema translation — that's the fix.
• One semantic note: cost.total comes from the pinned OpenRouter list rates. Wiring exact reported usage.cost through a pi provider extension can be added later if you ever want billed amounts for catalog-less models, but standard usage bills at list rates, so this matches.

Verification

• python3 -m unittest discover -s tests -> 38 OK
• actionlint .github/workflows/biggiepockets-review.yml -> clean
• resolve-prompts.sh deterministic per PR (untouched logic, comment-only registry change)
• Live pi runs on this machine against the org's OpenRouter key: exact command shape from the workflow, verdict written, usage+cost recorded.