Testing the peoplereadme CLI

All testing is shell-only (no GUI). Use a clean clone + uv sync, then uv run peoplereadme ....

Basic checks

• uv run ruff check src tests and uv run pytest -q (all tests offline, no API keys needed).
• uv run peoplereadme validate must exit 0 (riley-walz must stay schema-valid).

Synthetic persona fixture (for ingest/trace/eval testing)

1. uv run peoplereadme init me --class self
2. Build a synthetic X export zip: data/account.js + data/tweets.js with window.YTD.account.part0 = [...] / window.YTD.tweets.part0 = [...] prefixes; tweets are {"tweet": {"id_str", "full_text", "created_at": "%a %b %d %H:%M:%S +0000 %Y", ...}}. Self-threads (replies to own tweets via in_reply_to_status_id_str) produce traces with reconstruction_quality 0.8; standalone posts get 0.2 (both pairwise-eligible; the harness excludes < 0.2). Use ~40 tweets so the 15% test split has ≥5 eligible traces.
3. uv run peoplereadme ingest me --source x-archive=archive.zip then uv run peoplereadme trace me.

Harness (M2) live testing

• Requires OPENAI_API_KEY (available as a saved Devin secret). Use --model openai/gpt-4o-mini and small --n-pairs (e.g. 6) to keep cost < $1; a 6-pair run with baseline takes ~70s.
• uv run peoplereadme eval me --model openai/gpt-4o-mini --n-pairs 6 --seed 7 writes personas/me/evals/fidelity/<date>.json, auto-creates evals/rubrics/v1.json, and exports a blind calibration batch to evals/fidelity/calibration/<date>.jsonl (+ private <date>.answers.json key).
• Judge calls are disk-cached in ~/.peoplereadme/cache/llm (override root with PEOPLEREADME_CACHE); delete to force fresh judging.
• Calibration roundtrip: fill human_pick ("A"/"B") in the blind JSONL, then uv run peoplereadme calibrate me --batch <date> --ratings rated.jsonl. Expect exit 1 with "INVALID report card" whenever kappa < 0.4 or fewer than 50 rated pairs — that hard floor firing is correct behavior, not a bug.
• Expected diagnostics gotcha: with a skeleton persona package (no real context files) the judge often scores below chance and generated outputs are far longer than real tweets — check diagnostics in fidelity.json (position_bias, mean_real_len, mean_generated_len).

Failure modes (should all be clean errors, exit 1, no tracebacks)

• eval on a persona without traces/traces.jsonl
• calibrate with a missing batch/answers key
• eval with an invalid API key (LMError path)