bug: `cora brain` FTS-only mode returns 0 results for common words present in code (e.g. "registry"), no stemming

[ajianaz]

Describe the bug

Without an LLM/embedding provider configured, cora brain falls back to FTS5-only mode (results show signals: ["fts"]). It returns 0 results for common words that clearly appear in filenames, import paths, and code, while other words work fine. Stemming is also not applied.

To reproduce

Project: hompimpah (SvelteKit + Phaser + Rust/Axum), 582 symbols indexed.

cora index
for q in registry game games GameMeta "get game" scene; do
  printf '%-12s -> %s results\n' "$q" "$(cora brain "$q" --json | jq length)"
done

Actual results

Query	Results	Note
registry	0	❌ filename registry.ts, present in imports/comments
game	12
games	2	❌ no stemming — differs from game
GameMeta	15	camelCase kept as single token
get game	0	phrase query, no match
scene	19

Expected

• registry should match (it is a filename and appears in import paths).
• game and games should converge (stemming or query expansion).

Likely cause

1. The FTS5 index appears to cover the symbol signature/snippet only — not file path, imports, or comments. So words that live primarily in paths/comments are unreachable.
2. No stemming enabled → game and games are two distinct tokens.
3. camelCase identifiers are indexed as a single token → space-separated or sub-word queries cannot reach them.

Suggested fix

• Index file_path and comment/doc tokens into the FTS table (or a separate lower-weighted column).
• Enable the FTS5 porter tokenizer (or unicode61 + stemming) so game ≈ games.
• Optionally split camelCase identifiers into sub-tokens for matching.

Environment

• cora 0.11.1 (released 2026-08-01)
• No provider configured (FTS fallback path)
• macOS arm64