Remove duplicate fresh-index embedding pass

[unicornz-code]

Summary

• port only MinishLab/semble PR fix: Double embedding pass on fresh index MinishLab/semble#260's fresh-index correction
• reuse the already-produced ordered per-file vectors instead of embedding every cold chunk twice
• prove total encoded texts equal produced chunks and preserve vector/order identity
• add benchmark-only phase instrumentation for exact cold-index and Model2Vec boundaries
• stabilize one unrelated random-mock ranking assertion exposed by the changed encode-call schedule

Provenance

Production src/semble/index/create.py is byte-identical to upstream commit a79078007f21842eb8325c8217b3acd4d050d578. No upstream installer/version/assets/result claims were imported. Context-bound Semble 0.7.1 MCP behavior is unchanged.

Validation

• 361 tests passed, 97% coverage
• Ruff check/format passed
• pydoclint passed
• mypy passed
• all pre-commit hooks passed
• wheel and sdist built
• independent read-only review: PASS after profiler findings were remediated

The profiler explicitly labels instrumented totals, nested boundaries, derived residuals, process-lifetime RSS, and single-thread limitations; it adds no production telemetry.