feat: Voyage-4-Nano backend + project-scoped coding "walking mode" (#758)

[igun997]

Closes #758

Implements both asks from #758: (1) an optional large-context local model, and (2) a project-scoped "walking mode" that lets a coding agent navigate a codebase file-to-file with precise file:line provenance instead of fragmented summaries.

---

Before → After

Dimension	Before (develop)	After (this PR)
Embedding model	EmbeddingGemma-300M only (768d, ~2k ctx)	EmbeddingGemma or Voyage-4-Nano (2048d, 32k ctx) — local ONNX, selectable in config
Store scoping	Single global store (~/.uteke)	DB-per-repo: project .uteke/ auto-scoped by cwd; global store still used outside repos
Code awareness	Prose/markdown memories only	Code indexing: source tree → chunked memories with file:line, incremental via content hash
Chunking	Regex heuristics	tree-sitter AST (rust/go/py/ts/js/java/c/cpp) with regex fallback
Recall on code	Summary fragments, no location	Exact symbol + file:line, jump-to-definition ready
Onboarding	General memory only	Coding-vs-general usage-mode prompt; coding defaults to Voyage + per-project store
Visibility	CLI text only	uteke ui: single-binary local web UI — live chunking feed + zero-dep WebGL 3D code graph
Config	Global file	Layered: CLI args > project .uteke/uteke.toml > global > defaults
MCP	memory tools	+ uteke_index / uteke_index_status for agents
Agent flow	recall summaries	walk file-to-file, build a persistent project "mental map"

Recall, before vs after (same repo)

Before — no code in the store; a query like "how does treesitter AST chunking work" returns nothing actionable.
After — top hits are the exact functions, correctly ranked, each with file:line:

1. (1.000) function collect_chunk   [crates/uteke-core/src/chunker_ts.rs:182-219]
2. (0.984) function chunk_code_ts   [crates/uteke-core/src/chunker_ts.rs:156-176]
3. (0.968) function chunk_code      [crates/uteke-core/src/chunker.rs:303-316]

---

What's in it

Embedding (#758 part 1)

• Multi-model ONNX engine; Voyage-4-Nano (qwen3, q4, dim 2048, pre-pooled) as a local, offline alternative to Gemma — no cloud API.
• Hardened the OpenAI embedder against the 8192-token limit (char budget + shrink-and-retry).

Walking mode (#758 part 2)

• code_index engine: walks a tree, chunks each file, stores chunks as recallable memories with file:line.
• tree-sitter AST grammars with regex fallback; unsafe stays isolated in grammar crates (unsafe_code = "forbid" in consumer code).
• Incremental: indexed_files table + content hashing; only changed files re-embed; deleted files pruned.
• DB-per-repo isolation via layered config, auto-scoped by cwd.

Surface area

• CLI: uteke init --project, streaming uteke index [PATH], uteke index --status, coding-vs-general onboarding.
• uteke ui: single Rust binary, embedded HTML, zero npm/offline; live IndexProgress feed + WebGL 3D force-directed graph (orbit/zoom/pan, depth-shaded glow nodes, hover to file:line).
• MCP: uteke_index, uteke_index_status.

Verification

• Build green with --features treesitter.
• Tests: uteke-core 394 (treesitter), uteke-cli 55 — all passing.
• End-to-end recall verified on this repo (see table above).

Notes for reviewers

• Schema migration v15 -> v16 adds indexed_files (idempotent, existing DBs auto-migrate).
• Per-repo store DB/index artifacts are gitignored; the small .uteke/uteke.toml project config is intended to travel with the repo.
• Voyage scores compress into ~0.96-1.0 (qwen3 anisotropy) — ranking is correct, but absolute scores aren't a usable threshold. Per-backend dedup tuning is a sensible follow-up.
• Both features are additive; default behavior (Gemma, global store, prose memories) is unchanged unless coding mode is selected.