docs(evals): publish bounded OpenAdapt performance evidence

[abrichr]

What this measures

• Publishes row-level retained model-free evidence for Frappe Lending and OpenEMR: compiled GUI versus direct API, baseline and cosmetic drift, three trials per cell.
• Separately summarizes the fixed historical WAA core-four comparison: legacy demo-conditioned computer use versus zero-shot, three retained trials per task and condition.
• Adds strict validation and a clean-checkout byte-for-byte regeneration test using committed compact, non-sensitive inputs.

Results

• Retained model-free subset: compiled 12/12 at 32.28s mean; API 12/12 at 1.63s mean; zero measured silent incorrect success, over-halt, model calls, or model cost.
• Historical WAA: DC 3/12 and ZS 3/12 (0 pp reliability delta); 6 versus 5 silent incorrect completion claims.

Provenance boundaries

• The model-free summaries came from a worktree later committed as 84c7a94, but the summaries do not embed a runtime revision. The report records this only as a post-hoc association, not exact Flow source binding.
• Exact Flow 1.12.1 requalification and current Flow-versus-zero-shot remain unmeasured.
• The paid agent arm was omitted, and the source matrices remain publication_ready=false and full_matrix_complete=false.
• Historical model ID, runner revision, token usage, and cost are unavailable and remain explicitly unknown/null.
• The compact legacy input excludes screenshots, action payloads, and model content while retaining task instructions, normalized rows, and hashes for all 48 raw summary/execution inputs.
• No new GUI, cloud, or model runs were performed.

Verification

• 5 passed
• Ruff lint and format pass
• Clean-checkout committed-input regeneration is byte-for-byte deterministic
• Exact retained input and raw-manifest SHA-256 checks pass
• Credential, local-identity, and secret-pattern scans pass
• git diff --check passes