feat: Add 4 Aurora engine-behavior skills (parameter, replication, upgrade, lifecycle)

[kiranmam]

Four standalone skills covering ~15,000 cases/yr of Aurora MySQL + PostgreSQL engine-behavior gaps. All use AWS CLI + CloudWatch only — no MCP, no Data API. Covers parameter-group misconfigs (#14/#151), replication health (#135), upgrade readiness (#37/#137), and lifecycle/right-sizing (#42/#34).

[chetanandikanti]

Review — Aurora engine-behavior skills (4 skills)

Reviewed at PR head b37839b. Technically these are strong — the engine behavior is accurate (Aurora MySQL auto-manages innodb_buffer_pool_size; PostgreSQL shared_buffers = {DBInstanceClassMemory/32768} ≈ 25%; in-region storage-level replication is typically sub-second; there is no direct Serverless v1→v2 path; optimizer statistics reset requires a post-upgrade ANALYZE). Every skill's scoring dimensions sum to 100, the skills cross-reference each other cleanly, and the upgrade-advisor honestly flags the in-database checks it can't perform control-plane-only. A few items before this can merge:

Blocking

1. Required files missing (all 4 skills). Each ships SKILL.md only. Per the Agent Skills specification and CONTRIBUTING, every skill needs a README.md (with the non-production disclaimer and IAM prerequisites — rds:Describe*, cloudwatch:GetMetricStatistics, application-autoscaling:*, relevant ec2:Describe*), a CHANGELOG.md, and an evals/ directory (evals.json + eval_queries.json) with a passing eval run.
2. Frontmatter. All four use top-level version: / author:. Nest version and author under a metadata: block per the spec, and use the author alias (e.g. kiranmam) to match the merged database-rds-devops skill.
3. **Unclosed yaml fence (all 4).** `## DETECTION RULES` opens a ` yaml block that is never closed — verified: each file contains exactly one fence marker and zero closing fence. As a resultASSESSMENT COMMANDS, ASSESSMENT SCORING MATRIX, REMEDIATION PLAYBOOK TEMPLATES, and REPORT OUTPUT FORMATall render inside one code block and lose their##` headings. Close the fence and restore the headings in every file.

Content / consistency

• database-aurora-lifecycle-advisor: PART A … (#42) / PART B … (#34) and RS-05 "matches Aadd reference to demo video on The Keys to AWS Optimization #34 pattern" — in this repo #42 / #34 render as links to PRs/issues. These read as stray internal references; please strip them.
• database-aurora-replication-health: RL-01 describes lag as "elevated (>20 ms sustained; >100 ms = WARNING)," but the Thresholds and detection-rule sections use >100 ms = WARNING and >1000 ms = CRITICAL. Reconcile the "elevated" threshold.

Structure / scope

• Files are 443 / 411 / 365 / 381 lines — under the 500-line guideline but heavy; consider moving the catalogs and playbook templates into references/ (see eks-operation-review).
• Four skills in one PR — each still needs its own README / CHANGELOG / evals. Consider one PR per skill, or one PR with all four fully packaged. Also worth confirming distinctness vs the merged database-rds-devops skill.

Happy to re-review once these are addressed. Thanks for the strong technical content here.

[chetanandikanti]

Addendum — how this was reviewed (methodology, reasoning, next steps)

Steps

1. Fetched the PR head (git fetch origin pull/58/head) and diffed against main — confirmed the change set is exactly four new SKILL.md files.
2. Read all four skills in full and checked the engine-behavior claims against known Aurora behavior.
3. Ran objective checks instead of eyeballing: counted triple-backtick fences per file (each has 1 opening / 0 closing → unclosed block), counted lines (443 / 411 / 365 / 381), and listed each skill directory (only SKILL.md present).
4. Verified each skill's four scoring dimensions arithmetically sum to 100.
5. Cross-checked the two internal-consistency items (lifecycle #42 / #34; replication-health RL-01 threshold vs the Thresholds table).

Reasoning — why these are blocking

• Missing README / CHANGELOG / evals and the frontmatter shape are hard requirements in the Agent Skills specification and CONTRIBUTING; review/CI will reject without them.
• The unclosed fence isn't cosmetic — it hides the operational half of each skill (assessment commands, scoring matrix, playbooks, report format) inside one code block, so both a human reader and the agent's own section parsing lose those ## sections.
• The two content items are correctness bugs (a stray in-repo issue link, and a threshold that contradicts the thresholds table) that would mislead users and the agent.

Recommended next steps (fastest path to green)

1. Close the ```yaml fence and restore the ## headings in all four files (quick fix; immediately restores readability).
2. Add README / CHANGELOG / evals per skill and run the eval to a pass.
3. Fix frontmatter — nest version/author under metadata: and use the author alias.
4. Fix the two content items (lifecycle stray refs; replication threshold).
5. Consider splitting into one PR per skill (or fully package all four).

Re-ping me once addressed and I'll re-review.