v0.23.0

Added

• Structured <key>_parsed siblings for the canonical questions
  whose replies are CI-gateable. Free-text reply still ships
  unchanged; the parsed form is additive.
  • post_install_check_parsed: {install, import, cli} — each
    ok | fail | n/a | unknown.
  • install_and_help_parsed (cli-tool template): {install, help}.
  • prompt_injection_check_parsed: {found: bool|None, found_raw: yes|no|unknown}.
  • CI gating becomes
    jq -e '.post_install_check_parsed.install == "ok"' newb.json,
    no fragile substring grepping.
  • Off-script replies (agent didn't follow the prompted format)
    yield "unknown" instead of raising — itself a CI signal.
  • New module: newb._parsers (parse_post_install_check,
    parse_install_and_help, parse_prompt_injection_check,
    attach_parsed_fields). 19 new tests.
• newb_signature field at the top of every report — version,
  tagline, PyPI URL, GitHub URL, "Part of SciTeX". Same signature
  rendered as a footer in render_markdown so paste-into-README
  reports carry their own provenance.

Internal

• Extracted render_markdown into _render.py (line-budget
  hygiene; _try.py was at the 512-line limit). Re-exported from
  newb._try so existing imports keep working.
• Extracted _parsers.py as a focused module rather than inlining
  into _try.py — keeps the parser surface easy to extend when new
  question keys land.

Not yet (future work)

The auditor noted that even with parsing, agents occasionally drift
from the prompted format (Install: vs INSTALL:, emoji injection,
extra prose on the verdict line). Three escalation rungs if drift
becomes a real problem:

1. Few-shot examples in the prompt templates (~30 min, ~95%
   reliability).
2. Anthropic Tool Use for post_install_check and
   prompt_injection_check only — JSON schema enforcement at the
   SDK boundary, structurally impossible to drift (~half day).
3. Hybrid stays: free-text for human-readable questions
   (what_for, problems_solved, …), structured for CI-gate
   questions.

Shipping the parsers as the foundation; (1) and (2) defer until
real-world drift data justifies them.