Optimize recruiter-facing Voice2Task project page

[Raidriar7170]

Summary

• Rewrite the Chinese and English project homepages for recruiter and ML-reader scanability.
• Surface the evidence-backed project value, synchronized result table, architecture, validation trail, and repository navigation.
• Preserve explicit non-claim boundaries alongside the measured result.

Scope

This is a docs-only stacked PR. It targets codex/materialize-manifest-bound-train-only-sft-v1 and changes only README.md, README_en.md, and the committed recruiter-facing design/implementation planning documents.

There are no runtime, source, test, config, training, inference, evaluation, DPO, GRPO, data, metric, or evidence mutations. This PR does not merge, release, deploy, or publish a checkpoint or adapter.

Verification

• PYTHONPATH=src pytest -q: 1485 passed
• PYTHONPATH=src ruff check .: passed
• OPENSPEC_TELEMETRY=0 openspec validate --all --strict: 15/15 passed
• OpenSpec active-state check: zero active changes
• Truth-surface and repository-link checks: passed
• README source-binding, leak, reviewed-claim, and external-image checks: passed
• git diff --check origin/codex/materialize-manifest-bound-train-only-sft-v1...HEAD: passed
• Final scope check: docs-only four-file stacked diff; worktree clean before publication

Result boundary

The Final SFT arm scored higher on semantic/channel aggregate metrics, but strict exact fell from 1.67% to 0.83%; therefore, this PR makes no overall model improvement claim.

No training, inference, evaluation, DPO, GRPO, data, metric, or evidence mutation was performed. No merge, release, deploy, checkpoint publication, or adapter publication was performed.