Formatter: fix attribute spacing lost on elements nested in ERB blocks - #2148

Summary

Fixes #2142.

When the formatter encounters a content-preserving element (e.g. <pre>) whose body contains an ERB block or conditional, it falls back to IdentityPrinter to reconstruct that ERB-wrapped subtree byte-for-byte (javascript/packages/formatter/src/format-printer.ts → visitContentPreservingBody).

IdentityPrinter.visitHTMLOpenTagNode (javascript/packages/printer/src/identity-printer.ts) wrote the tag name and then its attribute children back-to-back with no separator. Without the parser's track_whitespace: true option (which the formatter does not pass), the AST simply has no node for the whitespace between a tag name and its first attribute, or between attributes — that gap isn't represented at all. So any HTML element with attributes nested inside an ERB block lost its attribute spacing:

<pre><% if condition %><span class="x">x</span><% end %></pre>

formatted to:

<pre><% if condition %><spanclass="x">x</span><% end %></pre>

This isn't just a cosmetic diff — <spanclass="x"> is a different (bogus custom) element than <span class="x">, so the formatter silently changes rendered output.

Root cause

Confirmed by inspecting the parsed AST directly: Herb.parse(source) (default options, no track_whitespace) returns open_tag.children containing only the HTMLAttributeNodes — there is no node covering the single space between the tag name and the first attribute, or between attributes. IdentityPrinter assumed children could be printed contiguously, which only holds when track_whitespace: true is used (in that mode an explicit WhitespaceNode fills the gap).

Fix

IdentityPrinter.visitHTMLOpenTagNode now tracks the end position of the last thing it wrote and compares it against each child's start position. If there's a gap (no node covers it — the track_whitespace: false case), it writes a single separating space. If the positions already line up (an explicit WhitespaceNode already covers the gap, i.e. track_whitespace: true), nothing extra is written.

How I found and verified this

• Reproduced against the published @herb-tools/formatter@0.10.3 + @herb-tools/node-wasm@0.10.3 packages with the exact example from Formatter: Attribute spacing is lost inside ERB blocks in pre elements #2142 — confirmed the <spanclass="x"> corruption.
• Traced the parsed AST (with and without track_whitespace) to find the actual gap in node coverage described above.
• Applied the same fix to the bundled/published printer code and re-ran the repro — output now round-trips correctly, including multi-attribute elements, ERB attribute splats (<div <%= attrs %>>), and interpolated attribute values, and formatting is idempotent (format(format(x)) === format(x)) in every case I tried.
• Applied the source fix to this branch and ran the actual test suites locally:
  • javascript/packages/printer: vitest run — 116/116 real assertions pass (unchanged from baseline; the 45 failing test files are pre-existing, unrelated to this change — they fail to import in this environment because ERBEndNode.build requires the generated nodes.ts from the Ruby/C codegen pipeline, which isn't available without the full C toolchain).
  • javascript/packages/formatter: vitest run — 1372/1480 passing, 2 more than baseline (the 2 new regression tests below). The 101 remaining failures are pre-existing and unrelated (CLI-binary tests needing a built executable, Tailwind class-sorter tests needing a Tailwind config) — verified identical failure set with and without this change.
• Added two regression tests to javascript/packages/formatter/test/html/content-preserving-tags.test.ts: the exact case from Formatter: Attribute spacing is lost inside ERB blocks in pre elements #2142, plus a multi-attribute variant.

Test plan

• cd javascript/packages/printer && vitest run — no regressions (116/116 real assertions pass, matching baseline)
• cd javascript/packages/formatter && vitest run test/html/content-preserving-tags.test.ts test/erb/content-preserving-blocks.test.ts test/erb/whitespace-preservation.test.ts — 94/94 pass, including the 2 new regression tests
• cd javascript/packages/formatter && vitest run (full suite) — no new failures vs. baseline