[FEATURE]: Detect Mermaid diagrams in untagged fences, not only ```mermaid

Feature hasn't been suggested before.

• I checked #39816 (closed) and #40508 (open). Both cover rendering Mermaid; neither mentions detection of untagged fences, which is what this asks for.

Describe the enhancement you want to request

decorateMermaid (from the merged, unreleased #43024) only acts on tagged fences:

if (!code.classList.contains("language-mermaid")) {
  clearMermaid(wrapper)
  return
}

Correct by the letter — but models often don't tag the fence, even when explicitly asked for a Mermaid diagram.

Measured across the stored sessions of one instance (opencode-ai@1.18.13, Nemotron 3 Ultra 550B):

Not every untagged block is a diagram. But the diagrams that were there arrived untagged more often than tagged, and the user just sees a code block — with no way to tell whether the feature is missing or the tag is.

The mechanism is already in the PR. renderMermaidSvg does:

if (!(await mermaid.parse(source, { suppressErrors: true }))) return

That alone identifies a diagram; non-diagram code doesn't parse. So the class check could be relaxed rather than replaced — attempt untagged (and text) blocks too and let parse decide, keeping the fast path for tagged ones. Cost is one suppressed parse per candidate, and it can be limited to complete blocks, which the code already distinguishes, so streaming is unaffected.

I first tried matching the first word against Mermaid's diagram keywords. It works, but ages badly — every new diagram type has to be added by hand. parse always knows.

Caveat: I haven't run your build; #43024 is merged but not in a published release (latest 1.18.18, 2026-08-13). This comes from reading the diff. The measurement is from our own sessions and stands on its own.