fix(dsql): address remaining PR awslabs#162 review items

- awslabs#17: Downgrade eval results to qualitative comparison, record model
  and version, note n=1 and recommend n>=3 for production confidence
- awslabs#18: SKILL.md is 281 lines (will update PR body)
- awslabs#20: Strengthen awsknowledge fallback to MUST — refuse fallback when
  recommendation depends on exact limit value
- awslabs#21: Already addressed in prior commit (reltuples staleness)
- awslabs#15: Document manual-only status and future Python converter direction
  (per anwesham-lab's suggestion for deterministic rewrites)
- awslabs#19: MCP mirror PR noted as follow-up in PR body

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Morlej

tools/evals/databases-on-aws/README.md

@@ -165,10 +165,15 @@ including type coercion index bypass, subquery unnesting, OR-to-IN, GROUP BY pus
 DSQL-specific patterns (reltuples estimate, join splitting). Includes one negative case
 (OR across different columns — agent should decline).
 
-**Evaluation method:** Manual behavioural comparison. Run `claude -p` with skill loaded vs
+**Evaluation method:** Manual qualitative comparison (n=1). Run `claude -p` with skill loaded vs
 `claude -p --bare` from a clean directory. Results in `query_plan_rewrite_eval_results.md`.
 No automated runner script — this suite is manual-only.
 
+**Future direction:** Many of these rewrites are deterministic pattern transformations. A future
+iteration SHOULD implement them as a Python SQL converter script that parses and rewrites SQL
+directly, with the reference files serving as documentation for the converter's rules. This
+would move correctness-critical rewrites out of the LLM and into deterministic code.
+
 **What it checks** (11 eval prompts):
 
 | Eval           | Focus                       | Key assertions                                               |

tools/evals/databases-on-aws/dsql/query_plan_rewrite_eval_results.md

@@ -1,7 +1,9 @@
 # Query Plan Rewrite Eval Results — With-Skill vs Baseline
 
 **Date:** 2026-05-08
-**Evaluation method:** Manual behavioral comparison — `claude -p` with skill loaded (from agent-plugins project root) vs `claude -p --bare` from clean directory. PASS/FAIL is a human assessment of transcripts against the expectations in `query_plan_rewrite_evals.json`.
+**Model:** Claude Opus 4.6 (global.anthropic.claude-opus-4-6-v1)
+**Runs per eval:** 1 (qualitative comparison, not variance-tested)
+**Evaluation method:** Manual qualitative comparison — `claude -p` with skill loaded (from agent-plugins project root) vs `claude -p --bare` from clean directory. n=1 per cell; PASS/FAIL is a single human transcript assessment. Results indicate directional improvement, not statistical significance. Re-run with n≥3 and majority vote for production confidence.
 
 ## Summary