docs: add experimental roadmap and evidence context to vision

[abrichr]

Summary

• Add 2×2 experimental matrix (retrieval × fine-tuning) to Core Thesis in vision.md
• Add evidence context to benchmark table: note it's an internal synthetic benchmark that validates the pipeline, not real-world performance

Why

Honest scoping of claims. The benchmark results are real but should be framed as "pipeline validated" not "production-proven." Links to openadapt-evals for ongoing WAA/OSWorld evaluation.

🤖 Generated with Claude Code

[abrichr]

Closing — these changes are already on main.

The 2x2 experimental matrix and evidence context additions to vision.md were incorporated into main via PR #34 (commit 339e5d3). The branch chore/reframe-positioning has diverged (2 ahead, 7 behind main), but the actual diff (experimental roadmap table, benchmark context note) is already present in the current main version of docs/vision.md.

Content review notes (for the record):

• The 2x2 matrix correctly references 46.7% -> 100% first-action accuracy
• The "internal synthetic login benchmark" framing is honest and appropriate
• The note about WAA/OSWorld evaluation being ongoing is accurate
• Consistent with "Trajectory-Conditioned Disambiguation" framing used in the meta-package README
• No overclaiming — well-calibrated positioning

All 4 CI checks passed (ubuntu/macos, Python 3.10/3.11).

No action needed — content is already merged.