hygiene: add an escapes mode — LaTeX eaten by Python string escapes

[Jammy2211]

Turns the sweep built for the 41-file raw-stringing task (PyAutoLabs/autolens_workspace#491) into a standing Hygiene check, so this debt is caught continuously instead of re-derived each time someone notices warning noise.

Why a mode, not a CI guard

The task's own follow-up suggested a -W error::SyntaxWarning compile guard in workspace CI. That would have been the weaker fix twice over: it catches only one of the two damage classes, and it returns a vacuous pass on Python 3.11. Both limitations are handled here.

The two classes

	what it is	diagnostic
warned	\s, \l, \[ — escapes Python does not recognise	keeps them literal, but warns on every compile; slated to become a SyntaxError
silent	\t in \theta, \f in \frac, \r in \rm, \b in \beta	none at all — the value is corrupted (\theta_E becomes TAB + heta_E)

A warning-only sweep reports a clean repo over a corrupted one, so the two are counted separately and files with only silent damage are marked explicitly — those are exactly the ones such a sweep misses.

Two interpreter traps, both of which have already produced a false "clean" here

1. Invalid escapes are a SyntaxWarning only on Python 3.12+; on 3.11 they are a DeprecationWarning. A SyntaxWarning-only sweep returns zero on 3.11, indistinguishable from "already fixed". Both categories are collected, so the scan is interpreter-independent.
2. compileall needs -f, or __pycache__ suppresses recompilation and the counts silently drop. This helper compiles from source text and never consults a cache.

Shape

Follows the existing helper-backed modes exactly: a stdlib-only, read-only _hygiene_escapes.py with --summary / --json-row / human output, wired into MODE_ORDER, the delegate and kind maps, prescan dispatch, the arg parser, the json-row path, the human body, and the default ranked worklist. Delegates to /refactor; Hygiene still never edits source.

The human output states the fix explicitly — prefix r on the enclosing docstring, not doubling the backslashes, which would leak into the rendered notebook prose.

Validation

Against a known answer, not just eyeballed. Run over the six workspace repos immediately after the raw-stringing merges, it reports exactly the one residue documented in #491 — autolens_workspace/scripts/group/likelihood_function.py, 2 warned + 1 silent — and nothing else.

A synthetic corpus confirms the discrimination:

demo_workspace/scripts/silent_only.py:  0 warned, 1 silent  <- silent only
demo_workspace/scripts/warned_only.py:  1 warned, 0 silent
clean_raw.py    (already r""")        not flagged
deliberate.py   print("\nreal newline") not flagged

bash -n clean, and the default worklist still renders all 14 modes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MTjtx5mdituitiyQYFGn2E

---

Generated by Claude Code