🐛 FIX: topmatter config merging, and source lines in included files

[chrisjsewell]

Two independent fixes. Closes #1165, closes #1167.

⚠️ One judgment call for review first

Enforcing global_only in topmatter (below) removes two undocumented but previously functioning behaviours, found by the review's old-vs-new probes:

• inventories in topmatter worked per-document in docutils-standalone (inv: links resolved against it). Now: a myst.topmatter warning and the field is ignored (the documented routes — docutils.conf myst-inventories, CLI flag — are unaffected). Sphinx was never affected (intersphinx).
• update_mathjax: false in topmatter half-worked under sphinx (suppressed the tex2jax_ignore classes on that page, though never the MathJax config override). Now warned-and-ignored.

Both fields are already marked global-only in the generated configuration docs, the break is loud rather than silent, and blessing them as per-document would create sphinx/docutils divergence — so this PR enforces. If you'd rather deprecate first (apply + warn "will be ignored in future"), that's a two-line change; say the word. Note also that any document carrying a global-only key in topmatter now emits a warning where it was previously silent, so strict (-W) builds may surface them (suppressible via myst.topmatter).

Topmatter config merging (#1167)

merge_file_level re-assigned the raw topmatter value after validation, clobbering values that converting validators had set:

• heading_slug_func: github in topmatter crashed at every heading ('str' object is not callable) — now it warns once as global-only.
• enable_extensions survived as a raw list instead of the validated set; url_schemes in list form crashed the publish outright, and its dict form skipped URL templating (raw refuri="wiki:Hobbit"). All now keep their validated/converted values.

Fields with global_only metadata are rejected in topmatter with a warning naming the field; everything else keeps the validator's converted value via a before/after identity check (with a documented contract note for future validator authors).

Source lines in included files (#1165)

Every warning inside an {include}d markdown file was attributed one line too low (startline + 1 double-counted the renderer's 0-based→1-based conversion). Also fixed while there: start-after advanced the offset by characters rather than lines, and negative start-line passed through unresolved. All shapes now map exactly — plain, start-line (including negative), start-after (including a multi-line marker and marker-at-EOF), combined with end-before, and nested includes — with mock_include_errors.md's regenerated expectation hand-verified against the included file. The :literal:/:code: branches are untouched (docutils line parity from #1164 preserved).

Review

Adversarial + behavior-regression pass with old-vs-new probes against master: a 7-page sphinx include-project renders byte-identical HTML with only line-number corrections in warnings/gettext locations (each hand-verified against the physical file); all five global-only fields probed individually in both docutils and sphinx paths (results above); adversarial include options (negative/beyond-EOF start-line, CRLF, multi-line markers) crash-free with exact lines. Two pre-existing items noted, not touched: sphinx console warnings show the parent page's path with the included file's line (location=(env.docname, line) — warning-taxonomy territory), and a few other nested_render_text callers share the same off-by-one family (substitutions already carry a TODO).

Full suite: 1247 passed.

---

Generated by Claude Code