Task: Implement Hotfix: the site build aborts on a POSIX class that reads as a link

Make the site build publish again, and leave a check that fails the defect at pull-request time rather than on main after the merge.

• doc/agile/versions/v0/sprint_26/clean-dq/task_clean_dq.org publishes, and the evidence command in it renders unchanged.
• The full site build completes.
• ores-check-org-links.el fails on the text this task fixes, and passes on the tree; the doc-lint job runs it on every documentation change.
• compass lint, validate_docs.sh and the skill checks stay green.

1. Reproduce the failure on the single page, then reduce it: the export aborts on the page, and a probe shows a table cell is the context that matters.
2. Establish which delimiters and which inline forms survive. An inline source block survives in a paragraph but not in a table cell; tilde verbatim survives in both, and preserves the command text.
3. Fix the two spans in the page, publish it, and read the rendered HTML to confirm the command is intact.
4. Write the check, prove it fails on the original text and passes on the tree, and wire it into the doc-lint job.
5. Run the full site build and the documentation gates, then raise the pull request.

Every number this task claims about a tree, with the command that produced it and the commit it was measured at.

• The mechanism, reproduced minimally. In a table cell, the probe command of two bracketed classes delimited with tildes exports; the same command delimited with equals signs aborts with the same :space: error. The equals sign inside the command closes the span, so the character class that follows is parsed as a link. Tilde delimiters cannot be closed by an equals sign, and an inline source block is not protected inside a table cell, which is why the fix is the delimiters rather than the form.
• The check is exact rather than heuristic. No org link type begins with a colon, so a parsed link whose target does is always a mistake; the check reports the file, the line and the target, and it does not need the id map, so it is cheap enough for the doc-lint job.
• Two defects of this family have now reached main: the file: link with a line-number search into a Python file, fixed in PR #2262, and this one. The check this task adds catches the second shape; a general "every link resolves" check would need a full parse of the corpus, which the site build already does on every push to main.
• A second, unrelated finding, carried here because it surfaced while this task ran the generators. doc/recipes/http/currencies/currency.org is generated output (it carries the #+generated: marker and the http_recipe.org.mustache template) that the tree did not contain: the other HTTP resources' pages are committed, this one was not. Nothing reported it, because the codegen drift check reports drift with git diff and an untracked generated file is invisible to it. The page is committed on this branch; the gate hole it exposes – a generated file the tree never received – is the same class as the unrun CMake file list check captured separately.

doc/agile/versions/v0/sprint_26/clean-dq/task_clean_dq.org publishes, and the command in its two evidence cells renders as the exact command that produced the number. The full site build completes, so the section that follows the dq page is no longer aborted.

projects/ores.lisp/src/ores-check-org-links.el is the structural half: it scans every org file the site publishes, parses only those whose text carries a candidate, and fails when a parsed link's target begins with a colon. On this tree it reports 6020 files scanned, 2 parsed and 0 unresolved; against the pre-fix text it reports both links and exits 1. The .github/workflows/doc-lint.yml job runs it, so the shape fails a pull request instead of main.