docs(skills): add the eval→learn→improve feedback loop (TDD evals + observability step 8)

[HMAKT99]

TL;DR

Two small, purely additive edits to existing skills that together add the one thing the pack is missing for LLM/agent work: the eval → learn → improve feedback loop. No new skill directory, no README/manifest/count changes, fully revert-safe. node scripts/validate-skills.js → 0 errors, 0 warnings.

File	What's added
skills/test-driven-development/SKILL.md	Evals: Testing Non-Deterministic (LLM-Backed) Behavior — golden set + rubric-judge grading, trajectory assertions, and growing the suite from production incidents
skills/observability-and-instrumentation/SKILL.md	Step 8: Feedback loops — learn from production runs — route failures to triage → distill recurring modes into eval cases + prompt/tool fixes → re-eval before shipping

The gap this fills

The pack's lifecycle runs forward: spec → plan → build → verify → review → ship. Once an LLM/agent feature is live, there's no documented operate-and-learn arc — no workflow for turning what production reveals back into durable improvements. That's the across-runs loop this PR adds, and it slots into the two skills that already own its two ends (verification and observability) rather than inventing a new scope.

Why this is not a duplicate (per CONTRIBUTING)

CONTRIBUTING asks contributors to justify the gap and prefer referencing over duplicating. This PR is built to that rule:

Adjacent work	What it owns	What this PR does instead
#286 evaluating-llm-output (proposed)	How to build an eval suite	Keeps eval mechanics lean and homes them in test-driven-development; references, doesn't re-teach
#285 reliable-agent-loops (proposed)	Single-run loop control (bounding, idempotent retries, resume)	Describes per-run reliability generically and leaves it to that domain
#253 harness-engineering (proposed)	Repo-local guardrails/governance for coding agents	This is about operating an LLM/agent product in production, framed in the observability skill

The new content is only the connective tissue none of them own: capture → triage → distill → re-eval, across runs.

Why this form (extensions, not a new skill)

CONTRIBUTING: "If your idea is a refinement of an existing skill, prefer a focused edit to that skill over a new directory." These are exactly that — additive sections that strengthen two existing skills, matching their structure, tone, and table format. Each follows the standard anatomy additions: a process section, rationalization rows, red flags, and verification checkboxes.

Validation

• node scripts/validate-skills.js → PASSED (0 errors, 0 warnings) — all 24 skills, required sections intact
• Frontmatter valid; TDD description 332 chars (< 1024 limit)
• Cross-references resolve to existing skills only (no dead refs)
• Diff touches only the two SKILL.md files — no README, manifest, or count edits

[HMAKT99]

Flagging this proactively since evals and agent-loops are a crowded area right now: this PR is intentionally not another skill in that cluster. It adds zero new directories and instead strengthens two skills that already exist.

The distinction in one line: the open proposals cover building an eval suite (#286) and making one agent run reliable (#285). Neither covers what happens across runs over time — feeding production failures back into eval cases and confirmed fixes. That across-runs loop is the only thing this PR adds; everything else is referenced, not re-explained.

I kept it deliberately easy to merge:

• Fully additive — no edits to existing prose, no README/manifest/count churn, trivially revert-safe.
• Passes CI — node scripts/validate-skills.js → 0 errors, 0 warnings; all required sections preserved; no dead cross-refs.
• Matches the house style — each section ships the standard process + rationalizations + red-flags + verification, in the same table format as the surrounding skill.

If you'd rather these land as a single section in one skill, route the eval mechanics elsewhere, or word the #286/#285 references differently, I'm happy to adjust — the content is structured so it can move without rework. Thanks for maintaining this pack.