fix(convert): coalesce a bold/italic span Confluence's editor split around a link

[willkg]

Summary

• Confluence's ADF stores formatting as marks per text run rather than nested elements. MdToConfluence always publishes a bold/italic span containing a link nested (<strong>text <a>link</a></strong>), which round-trips fine on its own — but once that page is opened and saved in Confluence's editor (even with no actual edit), it re-serializes through ADF and comes back split: <strong>text </strong><a><strong>link</strong></a>.
• storage_to_md.go rendered those two nodes independently, producing **text **[**link**](url). The closing ** is preceded by a space, so CommonMark's flanking rule refuses to treat it as emphasis at all — the exported markdown came back with literal ** characters, not just missing bold styling.
• Verified live 2026-08-30 by PUTting a test page's own unmodified atlas_doc_format body back at it via the API — exactly what the editor does on every save — and confirming the storage came back split.

Fixes #111.

Fix

• Added coalesceSplitMarks in storage_to_md.go: detects adjacent inline nodes sharing a formatting mark (including a link whose entire content carries that mark) and hoists the mark to wrap the whole run before rendering — the only shape markdown can actually express.
• A link only partly marked is left untouched, since there's nothing correct to hoist.
• Updated CLAUDE.md's architecture notes to describe the repair.

Test plan

• Added TestStorageToMarkdownCoalescesSplitMarks (bold/italic, both orderings, plain adjacent-run merging, partially-marked link left alone)
• make check passes (vet, fmt-check, test, build, lint)
• Reproduced the bug live against a throwaway test page (deleted afterward) via a direct atlas_doc_format PUT mimicking an editor save

[willkg]

Code review found two real gaps in the original fix, both addressed in 1831df2:

1. The repair only matched plain <a> links. The identical editor-induced split also happens for an internal Confluence page/space link (<ac:link>) — verified live the same way, by PUTting a bold internal link's own unmodified atlas_doc_format back at the page. A naive name-check addition wouldn't have worked, since <ac:link>'s visible text sits one level deeper (ac:link-body) than <a>'s. Added isLinkNode/linkTextBody/withLinkTextBody to cover both shapes, plus a dedicated test (TestACLinkCoalescesSplitBoldMark) and a golden fixture (testdata/storage2md/split-marks) for the <a> cases.
2. Merging two adjacent same-tag mark elements kept only one side's attrs map, silently dropping the other's — currently unobservable since no mark-tag renderer reads attrs, but a latent trap. Fixed with mergeAttrs.

The other two findings (case duplication between the mark-before-link/mark-after-link branches, and the unconditional same-tag merge being broader than strictly needed) were judged acceptable as-is: the duplication is a few lines and adding an abstraction over it would cost more clarity than it buys, and the unconditional merge is necessary for a third adjacent run to fold into an already-repaired node and is documented as such.