fix(feedback): add transfer evidence cards

[rodboev]

Closes #38

Summary

Add a per-generation transfer_evidence.json artifact that separates reusable improvements from task-specific residue before the next feedback handoff consumes generation output. The new artifact becomes the structured reuse boundary for feedback-context and context.md, while improvement.md stays the human-authored analysis file.
Tighten the transfer-evidence heuristic so higher-is-better metrics only count as reusable when they strictly improve, and percent-formatted scores are parsed instead of silently dropping back to unscored reuse decisions.

Why

Current origin/main only carries forward raw execution and evaluation context plus a truncated carryover from improvement.md. That forces later generations to infer what was actually reusable, which can flatten score gains, formatting wins, task-specific residue, and unsupported claims into the same channel.

Scope

• write transfer_evidence.json next to results.json after each generation
• derive a conservative first-pass card from evaluation status, score delta when available, and bounded improvement.md bullets
• render a deterministic TRANSFER EVIDENCE section in feedback context
• project only bounded reusable guidance plus explicit cautions into context.md
• update the harness feedback prompt and the focused goldens and tests that lock these text surfaces
• keep the change narrow to the transfer-evidence contract; PR #36 may require a rebase because it touches the same files, but this PR does not expand into that separate leak-fix scope

Test plan

• python -m pytest tests/test_orchestrator_helpers.py tests/test_feedback_context_golden.py tests/test_context_manager.py tests/test_context_golden.py tests/test_prompts_snapshot.py tests/test_generation_loop.py -q, passed, 31 passed in 0.17s; pytest emitted the recurring Windows temp-cleanup warning for C:\Users\Rod\AppData\Local\Temp\pytest-of-Rod\pytest-current after the successful exit
• python -m ruff check sia/layout.py sia/results.py sia/orchestrator.py sia/context_manager.py sia/prompts.py tests/test_orchestrator_helpers.py tests/test_feedback_context_golden.py tests/test_context_manager.py tests/test_context_golden.py tests/test_prompts_snapshot.py tests/test_generation_loop.py, passed
• python -m ruff format sia/layout.py sia/results.py sia/orchestrator.py sia/context_manager.py sia/prompts.py tests/test_orchestrator_helpers.py tests/test_feedback_context_golden.py tests/test_context_manager.py tests/test_context_golden.py tests/test_prompts_snapshot.py tests/test_generation_loop.py --check, passed, 11 files already formatted
• ty check sia/, exited cleanly with pre-existing unresolved-import warnings in optional openhands and pydantic_ai integrations outside this change