v0.3.0

Twenty-three field lessons merged, a new grading-and-fixtures reference split out of testing-and-verification, and a fix to the bundled client.

• New reference: references/grading-and-fixtures.md covers rubrics, test fixtures, run counts, ground-truth grading, measurement validation, and anti-hallucination design. references/testing-and-verification.md keeps staging, transports, write verification, and the platform signals that mislead diagnosis. Every long reference now opens with a section list.
• Three corrections to published text: the answerlength co-update rule does not apply to upload fields, the UI-only class is per field (descriptions save, submittext does not), and a workspace fallback can change reasoning effort as well as model.
• New platform facts: pickaxe_create defaults, chat history as a recent window, action_runs carrying no duration, image action files as public CDN URLs, the three Access denied surfaces, documentuploadtype, stale sandboxes with zero attachments, sibling-tool drift, the frame as a sentence, truncation fabricating content, a Model Reminder measuring as a null, admin group toggles destroying purchased credits, and the public completions endpoint exercised with stream, inputs, and imageUrls.
• scripts/pickaxe_client.py: run_completion now takes inputs as well as message, and unwraps the completion whether it arrives as a Python-repr string or a structured dict, which the platform now returns.