feat: parent-child retrieval for AST chunks

[Lucas-Bur]

Parent

#83

What to build

Add parent-child retrieval on top of AST-aware chunking. Fine-grained child entries are embedded for precise matching, while query results return their complete semantic parent by source range. Parent records are deduplicated by stable parent identity and remain available without embedding every parent body.

Index v2 should reserve the parent identity, source range, and entry-level metadata needed by this slice.

Acceptance criteria

• Searchable child entries carry a stable parent ID and exact source range
• A child hit resolves to the complete parent function, class, or declaration
• Parent bodies are stored or loaded once and deduplicated by parent ID
• Multiple child hits from one parent do not duplicate returned parent content
• Unsupported languages retain normal chunk retrieval
• JSON and human output preserve accurate file and line metadata
• Tests cover child matching, parent resolution, deduplication, and missing parents

Blocked by

• feat: incremental indexing via mtime cache #84
• feat: embedding cache for incremental reindex #95
• feat: lazy text loading at query time #96

[Lucas-Bur]

This slice needs an explicit token-window invariant; see the dedicated safety issue #164.

The current AST chunker packs by lines and deliberately keeps a single oversized node intact. The production Transformers.js pipeline then silently truncates inputs over the active tokenizer limit. A child vector can therefore represent only the prefix of its declared source range, while identifiers in the suffix become unretrievable.

Please make parent-child indexing model-aware: every embedded child must be measured with the active tokenizer and remain below the hard/operational limit, with recursive AST-child or line/token splitting for oversized parents. A complete parent may be returned as source context without being embedded itself. If a node cannot be split safely, fail with an actionable typed error rather than create a truncated vector.

Suggested acceptance addition: no child embedding is produced from silently truncated input, and oversized parents preserve full source coverage through bounded child entries.