follow the skill spec: references/ not refs/, and namespaced subskill names

- refs/ -> references/, the folder name the Agent Skills spec uses and the one
  Hermes skips when it walks for nested skills.
- rl and pinn declared name: rl and name: pinn, which are global names in a flat
  skill namespace. Now ml-debug-rl and ml-debug-pinn. They also called themselves
  sub-skills of 'ml-debugging', which is not this skill's name.
- Drop the dead link to SKILL_old.md. It moved into gitignored slop/, so the link
  was broken for anyone who cloned.
- Route references/llm_judge_litreview.md, the one reference SKILL.md never named.
- Description leads with the trigger situations. Hermes truncates it to 57 chars.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

docs/candidate_quotes.md

@@ -1,7 +1,7 @@
 # Unused quotes from the ml-debug evidence cache
 
 Mined from `/home/wassname/.agents/skills/ml-debug/docs/evidence/` (about 40 cached sources) and
-`/home/wassname/.agents/skills/ml-debug/refs/`. Every quote here was checked against
+`/home/wassname/.agents/skills/ml-debug/references/`. Every quote here was checked against
 `/home/wassname/.agents/skills/ml-debug/README.md` and is not used there. Line numbers were
 verified by grep on a distinctive substring; long source lines are single wrapped paragraphs, so
 one line number can hold a long quote.
@@ -82,7 +82,7 @@ Why it lands: seed noise alone can clear a significance bar. So one A-versus-B g
 Why it lands: turns "am I overconfident" into one answerable question with a calibration target, and points the next action at the least reliable step rather than the most interesting one.
 
 ## My Model of the Research Process (shared draft), as quoted in the skill's own topic note -- Neel Nanda
-- file: /home/wassname/.agents/skills/ml-debug/refs/research_taste.md:134
+- file: /home/wassname/.agents/skills/ml-debug/references/research_taste.md:134
 - failure modes: 1, 3
 - epistemic context: quoted from an unpublished Google Doc draft, so weaker provenance than the published posts by the same author.
 
@@ -224,7 +224,7 @@ Why it lands: two modes at once. Hypotheses 2 and 3 can be hypothesis 1 wearing
 Why it lands: a symptom-to-cause table where every symptom has two or three candidates and only one of them is a learning rate. It is a ready-made hypothesis-2-and-3 generator for the moment the agent reaches for the knob.
 
 ## My Model of the Research Process (shared draft), as quoted in the skill's own topic note -- Neel Nanda
-- file: /home/wassname/.agents/skills/ml-debug/refs/research_taste.md:120
+- file: /home/wassname/.agents/skills/ml-debug/references/research_taste.md:120
 - failure modes: 3
 - epistemic context: unpublished draft quoted in a local topic note; weaker provenance than the published posts.
 

docs/evidence/llm_judge_biases.md

@@ -12,7 +12,7 @@ summarizer produced from a web page, and nobody has read the paper. On
 2026-08-15 I re-pulled the five [ID] entries the litreview depends on and two
 of the five carried a wrong number, so treat the remaining 11 as roughly 2-in-5
 wrong until each is checked against raw text. Do not promote an [ID] number
-into SKILL.md or refs/ without re-pulling the paper first.
+into SKILL.md or references/ without re-pulling the paper first.
 
 ## "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" — Zheng et al. (LMSYS), NeurIPS 2023 — https://arxiv.org/pdf/2306.05685
 

docs/evidence/miller_2024_error_bars_evals.md

@@ -2,7 +2,7 @@ Source: https://arxiv.org/pdf/2411.00640 (Evan Miller, Anthropic, Nov 2024) + ht
 Title: Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations
 Fetched-via: r.jina.ai on the arXiv PDF and the Anthropic post, 2026-08-16
 Fetch-status: verbatim from full PDF text (math notation mangled by the PDF-to-markdown pass; prose is clean)
-Used-by: refs/llm_judges.md (repeat draws, temperature, paired differences)
+Used-by: references/llm_judges.md (repeat draws, temperature, paired differences)
 
 # Adding Error Bars to Evals (excerpts)