docs(roadmap): audit Phase 6 against benchmarks, fix overview table (#638)

* docs(roadmap): audit Phase 6 against benchmarks, fix overview table inconsistencies

Overview table: phases 7–12 had scrambled themes that didn't match
their actual sections (e.g. Phase 7 said "Runtime & Extensibility"
but the section is "Expanded Language Support"). Fixed all anchors,
themes, and deliverables. Added missing Phase 13 row.

Dependency graph: fixed phase labels to match section names.

Phase 5: added missing ✅ on section header.

Phase 6: rewrote with honest v3.3.1 benchmark breakdown showing
native vs WASM per-phase. The 1.85× native speedup comes almost
entirely from Parse (3.5×) — the other 9 phases show negligible
native advantage because they run identical JS/SQL on both engines.
Added 4 new sub-items (6.9–6.12) to address this:

- 6.9: AST node DB write optimization (393ms native ≈ 397ms WASM)
- 6.10: CFG & dataflow DB write optimization (161ms + 125ms, Rust slower/same)
- 6.11: Native insert nodes pipeline (206ms native ≈ 201ms WASM)
- 6.12: Native roles & edge build optimization (no native advantage)

Updated 6.8 to reflect partial completion (roles 255ms→9ms done via
#622, but sub-100ms target not met at 466ms). Changed ✓ → ✅ on all
completed sub-items to match document standard.

* docs(roadmap): fix benchmark table inconsistencies from review

- Update header to clarify mixed versions (full-build v3.3.1, 1-file v3.4.0)
- Add note explaining gap between phase sums and total row
- Clarify 6.5 result text with current benchmark figures (52ms / 9ms)
- Fix intro text: twelve → thirteen phases

* docs(roadmap): clarify 9ms vs 54ms roles discrepancy in 6.8

The benchmark table shows ~54ms for 1-file roles (full roles phase),
while the incremental path achieves 9ms. Add a note explaining the
benchmark measures the full phase, not the incremental shortcut.

* docs(roadmap): update 6.4 result text to reflect current 206ms benchmark

Same pattern as the 6.5 fix: the 310ms was the improvement at time of
6.4 work, now further reduced to 206ms in v3.3.1 benchmarks.

Author: carlos-alm