Spike: evaluate streaming approaches for large CSV datasets

[vmvarela]

Part of #69

Description

Investigate two approaches for handling large CSV files that don't fit in memory:

1. SQLite virtual table () — streaming CSV input via a virtual table interface
2. Disk-backed temp storage (PRAGMA temp_store = FILE) — configure SQLite to spill to disk automatically

Produce a written recommendation (implementation notes, trade-offs, estimated effort) to guide sub-issues 2–5.

Acceptance Criteria

• Both approaches prototyped or evaluated with a 1GB+ test file
• Recommendation written as a comment on this issue: which approach to implement first and why
• Memory usage measured for each approach
• Known limitations documented (e.g. which SQL operations won't work)

Notes

• The disk-backed approach may deliver 80% of the value with 20% of the complexity
• Start with PRAGMA temp_store = FILE since it requires no query-semantic changes
• Timebox to 4 hours max