feat: auto-merge sibling hits into semantic parents

[Lucas-Bur]

Parent

#83

What to build

Collapse multiple high-ranking child hits from the same semantic parent into one parent result. Auto-merging runs after ranking and filtering, preserves the best score and provenance, and avoids spending the output budget on overlapping fragments from one function or class.

Default policy: two or more selected children with the same parent ID collapse into the parent. The policy must remain deterministic when top-K or output budgets are applied.

Acceptance criteria

• Two or more selected children from one parent collapse into one parent result
• A single child remains a child result
• Merged results preserve best score, rank provenance, file, and parent line range
• Top-K and output-budget handling remain deterministic before and after merging
• Results from different parents never merge
• JSON output identifies merged results and contributing child IDs
• Tests cover threshold, ordering, deduplication, and budget boundaries

Blocked by

• feat: parent-child retrieval for AST chunks #146
• feat: lazy text loading at query time #96

[Lucas-Bur]

Auto-merge must remain a retrieval/output operation, not a workaround for embedding oversize; see #164.

Child vectors must already be bounded and lossless before ranking. When multiple child hits collapse into a parent result, do not re-embed the full parent if it exceeds the model limit. Keep the parent source range/provenance for display, but treat its output/context budget separately from the embedding budget.

Please add an invariant that merging siblings cannot create a new oversized embedding request or hide that some source content was never represented by a bounded child. Deterministic budget handling should distinguish retrieval merging, embedding token limits, and final response character/token limits.