Fix chunking file_format placement: Markdown-aware chunking was never enabled

[claudespice]

Summary

Three recipes set file_format: md inside the chunking: block of a column embedding:

chunking:
  enabled: true
  target_chunk_size: 256
  overlap_size: 64
  file_format: md      # <-- silently ignored here

chunking: deserializes into EmbeddingChunkConfig, which has exactly four fields — enabled, target_chunk_size, overlap_size, trim_whitespace — and no file_format. The struct does not carry #[serde(deny_unknown_fields)], so the key is silently dropped rather than reported: the recipes look like they enable Markdown-aware chunking, but every one of them has been chunking with the plain text splitter.

The runtime reads file_format from the dataset's params, not from the chunking block. This PR moves the key up into params:, which is also the placement used by the file recipe (file/README.md) and by spiceai's own test spicepods (test/models/spicepod_openai.yml).

Recipes fixed:

• models/openai (spicepod.yaml + the README's copy of the same block)
• azure_openai (spicepod.yaml)
• search_github_files (spicepod.yaml, commented-out example)

Verified against

spiceai/spiceai at v2.1.1:

• crates/spicepod/src/component/embeddings.rs — EmbeddingChunkConfig has only enabled / target_chunk_size / overlap_size / trim_whitespace; no deny_unknown_fields.
• crates/runtime-search/src/embeddings/table.rs — from_spicepod_columns builds ChunkingConfig from the per-column chunk config for the three size fields, and takes file_format from a separate argument.
• crates/runtime/src/embeddings/connector.rs — that argument is dataset.params.get("file_format").
• crates/runtime/src/component/dataset/builder.rs — Dataset.params is the raw spicepod params map, so the chunker sees the key regardless of connector parameter specs.
• crates/chunking/src/lib.rs — match cfg.file_format { Some("md" | ".md" | "mdx" | ".mdx") => Splitter::Markdown(...), _ => Splitter::Text(...) }.

git log -S file_format confirms the misplacement predates #508 and #136, which only flipped chunking.enabled from false to true — so this is not reverting a deliberate choice.

One thing for reviewers to weigh

The github connector does not list file_format in its PARAMETERS spec, so it will log one line at startup:

Ignoring parameter file_format: not supported for github.

This is cosmetic — the chunker reads the raw params map, so Markdown chunking does take effect — but if you would rather not have the warning, the alternative is to drop the file_format line entirely and accept plain-text chunking (which is what these recipes have actually been doing all along). Happy to switch to that if preferred.

Evidence

Static review only — these recipes need a GITHUB_TOKEN plus OpenAI/Azure credentials, so they were not run. The config path is traced end-to-end in the source links above, and all three files were re-parsed as valid YAML after the change.