fix: remove grammar constraint to fix node-llama-cpp 3.18.1 crash

[LevinFaber]

Summary

Fixes crash during qmd query expansion phase caused by node-llama-cpp 3.18.1 regression in grammar-constrained generation.

Error: Failed to accept token in sampler: Unexpected empty grammar stack after accepting piece: 0 (15)

Root Cause

node-llama-cpp 3.18.1 (bundles llama.cpp b8390) introduced a regression where grammar-constrained generation crashes deterministically with the GBNF grammar used for query expansion. This affects both Bun and Node.js — not Bun-specific as initially assumed.

Changes

• src/llm.ts: Remove grammar parameter from expandQuery() generation call
• src/db.ts: Add missing transaction() method to Database interface (build fix)

Why Remove Grammar Instead of Downgrade?

Approach	Pros	Cons
Remove grammar (chosen)	• Stays on latest node-llama-cpp (3.18.1) • Includes CVE-2025-52566 & CVE-2026-2069 fixes • No version pinning technical debt	• Slightly less strict output format enforcement
Downgrade to 3.17.1	• Grammar constraints work	• Misses security fixes • Pins to old version indefinitely • llama.cpp b8390 had no grammar fixes planned

The grammar was redundant — manual parsing already correctly extracts lex:, vec:, hyde: lines from model output and validates types. The GBNF grammar provided no additional safety that wasn't already handled by the parser.

Test Results

Runtime	node-llama-cpp	Grammar	Result
Bun v1.3.10	3.18.1	Removed	✓ Works (18.4s)
Node.js v24.13.1	3.18.1	Removed	✓ Works
Bun v1.3.10	3.17.1	Enabled	✓ Works (verified)
Node.js v24.13.1	3.18.1	Enabled	✗ Crash confirmed

Verification

• Builds successfully (npm run build)
• Tested with Bun — query expansion works
• Tested with Node.js — query expansion works

Related

• CVE-2025-52566: Fixed in 3.18.1 (tokenizer)
• CVE-2026-2069: Fixed in 3.18.1 (grammar buffer overflow)