Release v0.6.0

[github-actions]

Merging this pull request tags v0.6.0 and creates its GitHub release.

CI waits for approval: open the Checks tab and click Approve workflows to run.

Added

• FieldMap.source_type: where the records came from (hf, inspect_task or local). The scanner runner sets it, so a scanner can judge evidence by its source; it is None when a scanner is called directly.

Changed

• encoding_issues no longer flags tabs used as layout: indentation (code pasted without a fence, such as tracebacks in issue text), a tab after a list bullet or number, and tabs separating the columns of a table row. A single tab inside prose is still flagged. On a sweep of Inspect Evals this removed nearly all of SWE-bench's and b3's findings.
• A control character that swallowed the start of a LaTeX command, such as a tab followed by extsuperscript (\textsuperscript with its backslash read as a string escape), is now medium severity, and the explanation and latex_commands metadata name the command. latex_escapes only looks inside math, so these were otherwise reported as stray tabs.
• extraction_artifacts no longer reports a non-breaking space on its own in a HuggingFace or task dataset, where web text is full of them; it is still counted beside another artifact, and still reported alone in local annotation files extracted from PDFs. A zero-width joiner or non-joiner between two letters of a non-Latin script is spelling (Bengali, Telugu, Persian), not an artifact. On a sweep of Inspect Evals these took the scanner from 2,103 findings to 366, mostly APPS, DROP, MMLU and MGSM.
• inconsistent_format runs each check only when a scorer of the task is thrown by it. exact and match fold case and ignore punctuation, so under them capitalisation and trailing punctuation are no longer checked; includes folds case but compares punctuation; pattern compares everything. Length outliers are still checked under every verbatim scorer.
• answer_length and inconsistent_format's length check no longer apply when the task also scores with f1, which gives partial credit for a long answer. This reverses the decision in Answer rules ignore the task's scorer and choices #33 that exact beside f1 runs the rules as before: on SQuAD, scored with both, they reported 1,495 findings, none of which could change a score. A scanner that does not apply says which aspects the task's scorers ignore.

Removed

• The forced_choice_leakage scanner. A question that offers its answer as one of two options ("Is it an MRI or a CT scan?") is an ordinary binary question: a model that pattern-matches still has to choose between the options. On a sweep of every Inspect Evals eval it reported 1,614 findings, 1,333 of them on DROP's reading-comprehension questions ("Which happened first, X or Y?", answered from a passage), and none worth acting on. binary_question_ratio still reports datasets dominated by binary questions. --scanners forced_choice_leakage is now an unknown scanner.

Fixed

• mojibake no longer flags areas and volumes in ångströms. Ų and ų are valid UTF-8 for Ų and ų when read as Windows-1252, so the scanner reported them as mojibake. They are now weak evidence, like a Mac Roman √π: reported only beside other mojibake in the same field, or between letters. On a sweep of Inspect Evals this removed 365 false positives from SciKnowEval and changed nothing else.
• The terminal report no longer crashes when a finding quotes sample text that looks like rich markup, such as MacBench's [ANSWER]…[/ANSWER]. Explanations and sample ids are escaped; before, the scan raised MarkupError after writing its findings and exited 1.