feat(brain): runtime embedding config + incremental per-symbol embedding

[ajianaz]

Summary

Add runtime-configurable embedding backend selection and incremental per-symbol re-embedding to Brain Mode.

Currently, the embedding backend (hashing-trick 256d vs nomic pretrained 768d) is selected at compile time via #[cfg(feature = "pretrained-embed")]. This means users cannot switch backends without recompiling. Additionally, embed_project() re-embeds all symbols in a project on every index run, even when only 1 file changed.

Problem

1. No runtime config — embedding backend is compile-time only, no .cora.yaml control
2. Full re-embed on every index — embed_project() queries SELECT ... FROM symbols WHERE project_id = ? with no changed-symbol filter
3. No brain section in .cora.yaml config schema

Proposed Changes

1. BrainConfig in .cora.yaml

brain:
  embedding: auto    # auto | hashing | pretrained

Value	Behavior
auto (default)	Best available: pretrained (if compiled) → hashing fallback
hashing	Force 256d hashing trick (zero dependency)
pretrained	Force nomic 768d (requires --features pretrained-embed)

New struct in src/config/schema.rs (~15 lines).

2. Runtime Embedding Dispatch

Refactor src/embed/mod.rs from compile-time #[cfg] to runtime enum dispatch:

pub enum Backend {
    Hashing,       // always available
    Pretrained,    // cfg(feature = "pretrained-embed")
}

pub fn resolve_backend(config: &BrainConfig) -> Backend { ... }

The embed_code_dispatch() function reads the resolved backend instead of compile-time cfg.

3. Incremental Per-Symbol Embedding

Migration v7: Add embed_fingerprint TEXT column to symbols table.

ALTER TABLE symbols ADD COLUMN embed_fingerprint TEXT;

brain.rs: embed_project() only embeds symbols whose name + signature hash differs from stored fingerprint:

// Before: embed ALL symbols
SELECT id, name, kind, signature FROM symbols WHERE project_id = ?1

// After: embed ONLY changed symbols
SELECT id, name, kind, signature FROM symbols
WHERE project_id = ?1 AND (
  embed_fingerprint IS NULL
  OR embed_fingerprint != ?
)

Performance Impact

Scenario	Before (full re-embed)	After (incremental)
Full index (cold, 1106 symbols)	0.9s	0.9s (same — all new)
1 file changed (~89 symbols)	0.6s	~0.05s
cora watch auto-reindex	0.6s per change	~0.05s per change

Benchmarked on Uteke (1106 symbols, 147 files) with nomic 768d backend.

Files to Change

File	Change	Est. Lines
src/config/schema.rs	Add BrainConfig struct	+15
src/config/loader.rs	Add brain defaults	+5
src/embed/mod.rs	Runtime dispatch + Backend enum	+50 (refactor)
src/index/schema.rs	Migration v7: embed_fingerprint column	+15
src/index/brain.rs	Incremental embed in embed_project()	+30 (refactor)
src/index/mod.rs	Pass brain config to embed call	+3

Total: ~120 lines changed across 6 files.

Testing

• cargo test --features tree-sitter — all 903+ tests pass
• cargo test --features tree-sitter,pretrained-embed — pretrained path tested
• cargo clippy --all-targets --features tree-sitter -- -D warnings
• cargo clippy --all-targets --features tree-sitter,pretrained-embed -- -D warnings
• Manual test on Uteke (1106 symbols) — incremental embed is faster
• Manual test on Corin (537 symbols) — brain search returns vector signal
• Config test: .cora.yaml with brain: { embedding: hashing } forces 256d
• Config test: .cora.yaml with brain: { embedding: pretrained } forces 768d
• Config test: no brain section → auto default works
• Migration test: existing DB upgrades cleanly to v7

Related

• Builds on PR fix(watch): use global index DB with schema migrations #498 (cora watch fix — uses embed_project on file change)
• Prepares for ONNX integration in v0.14.0 (milestone Add GitHub Action for cora in CI/CD #17)