fix(code-index): preserve identifiers when splitting large nodes

[0xMink]

Closes #10715

Summary

• When the code parser splits a node exceeding MAX_BLOCK_CHARS, small children (export keyword, class/function name) fall below MIN_BLOCK_CHARS and are silently discarded, causing identifiers to disappear from the search index
• Add a split-local sibling accumulator (_expandChildrenWithAccumulator) that prepends small children to the next large sibling as a synthetic QueueItem, and appends trailing small children to the last large sibling
• Large children with no pending small siblings are pushed as raw Nodes to preserve recursive splitting of deeply nested structures
• Uses space-separated joins to preserve readable whitespace between accumulated tokens
• Trailing small siblings after a raw Node are dropped (rather than converting the Node to a QueueItem that would block recursion) since they are typically closing delimiters with no semantic value

Test plan

• Existing: 36 tests pass (vitest run services/code-index/processors/__tests__/parser.spec.ts)
• New: should preserve class/function identifier when splitting a large node
• New: should not contaminate across unrelated parent nodes
• New: should allow recursive splitting of large children with no pending parts
• New: should preserve recursive splitting even with trailing small siblings
• New: should append trailing small siblings to the last large sibling
• Total: 41

[roomote]

Rooviewer   See task

Both previously flagged issues are resolved in this revision. No new issues found.

• Wrapping large children as QueueItem when no pending parts exist prevents recursive splitting of deeply nested nodes (behavioral regression)
• pendingParts.join("") concatenates sibling texts without whitespace separators, producing garbled output like exportclassTestParser

Previous reviews

• 4db7f10: Review #1
• 19f6478: Review #2
• 6b105e3: Review #3

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

[roomote]

Bug: when a large child has no pending small siblings, it gets wrapped as a QueueItem here. Back in the main loop (line 212), !isQueueItem(currentItem) is false for QueueItems, so the child's own tree-sitter children are never examined for recursive splitting -- it falls straight through to line-based chunking. This means a class_body node containing multiple methods will no longer have those methods extracted as individual semantic blocks; instead they get chunked by arbitrary line boundaries. Consider pushing the raw Node when there are no pending parts, so the main loop can still recurse into its children.

_{Fix it with Roo Code or mention @roomote and request a fix.}

[roomote]

Bug: pendingParts.join("") concatenates sibling texts with no separator. Tree-sitter child nodes don't include inter-node whitespace, so accumulating ["export", "class", "TestParser"] produces "exportclassTestParser" rather than "export class TestParser". The same issue applies to join("") on line 344 (tail handling). The existing test doesn't catch this because it only checks includes("TestParser"), which matches as a substring of the garbled concatenation. Consider using join(" ") or join("\n") to preserve readable separation.

Suggested change

	text: pendingParts.join("") + text,
	text: pendingParts.join(" ") + " " + text,

_{Fix it with Roo Code or mention @roomote and request a fix.}

[0xMink]

Fixed. Large children with no pending small siblings are now pushed as raw Node objects so the main loop can recurse into their children. When trailing small siblings follow a raw Node, they are dropped rather than converting the Node to a QueueItem (which would block recursion). Added test: "should preserve recursive splitting even with trailing small siblings".

[0xMink]

Fixed. Changed pendingParts.join("") to join(" ") with " " + separators at all concatenation points. Test 1 now explicitly asserts "export class TestParser" (proper spacing) and rejects "exportclassTestParser" (garbled).