docs(evals): bind current Flow release performance

[abrichr]

Outcome

Adds a reproducible, local-only performance readout bound to the exact published openadapt-flow v1.16.1 wheel and release tag.

• 3 trials per arm per condition across clean, theme, and label-rename surfaces
• compiled replay versus steelmanned positional and name-scoped Playwright controls
• one shared screenshot/OCR effect oracle for exact note, Triage row, intended patient, and wrong-target writes
• steady action-loop and fresh-browser end-to-end median/p95
• three record/compile setup trials
• explicit correct / silent incorrect / wrong action / over-halt / halt-error taxonomy
• zero model calls, zero model cost, no VM/provider mutations

Exact binding

• Flow tag: v1.16.1
• Flow release commit: 113ce992b491576d77236f495b983165ce7a63bd
• Published wheel SHA-256: c7073283475e7ae722db2478b499d962364851104e09edb71f254ba21c1310cd
• Evals base: 7629ba5a2447c919e499e88e9b4eedbc80c8f3ab
• Runner SHA-256: ac58c0b9a02cfc991144a1f5c7a2814c6b6b748bcba27b87f8e1027ee575970f

Result

Compiled effect oracle: clean 3/3, theme 3/3, rename 3/3. Theme is deliberately reported as 3/3 over-halt because the intended effect was present while region_stable left Replayer unsuccessful. Both selector controls pass clean/theme 3/3 and fail label rename loudly 0/3 before mutation. All cells have zero silent incorrect success, wrong action, model calls, and model cost.

This is deterministic runtime overhead/robustness evidence, not a zero-shot comparison. The report explicitly leaves current Flow versus zero-shot unmeasured and records the missing WAA evaluator/hybrid wiring needed before any paid run would be valid.

Verification

• ruff check and ruff format --check on runner/tests
• 9 focused tests passed (test_current_flow_local_benchmark.py + existing performance-report tests)
• final artifact/source assertions passed
• no screenshots, copied app source, AGPL benchmark material, or package artifact changes