workflow: convert dense simulation setup prose to numbered steps (#1948)

Author: Copilot

.github/workflows/workshop-student-simulator.lock.yml

@@ -255,7 +255,11 @@ The profiles are generated from `.github/skills/micro-environment-simulator/work
 
 For each student in the generated cohort, simulate their experience step-by-step using the following rules:
 
-First read the baseline Monte Carlo output that was already written to `/tmp/gh-aw/agent/sim/data/monte-carlo-replay.json`, then read both `curriculum.json` and `curriculum-quality-metrics.json`. Build a candidate set from the five highest-dropout steps and the five lowest-scoring curriculum steps. Semantically score and rank every candidate using the `stateReadiness`, `pathClarity`, and `recoverySupport` dimensions defined below, then select the three highest-risk steps for detailed analysis. Inspect every workshop page mapped to each selected step, plus the preceding activities that produce its required state, before writing `/tmp/gh-aw/agent/sim/data/agent-step-insights.json`. A simulated step may map to several pages; do not treat a file limit as a step limit. For example, inspect the learner's Step 7 authoring path (Terminal, GitHub UI, or GitHub Copilot) and the shared Step 7d model-access activity before adjusting Step 8.
+1. Read the baseline Monte Carlo output that was already written to `/tmp/gh-aw/agent/sim/data/monte-carlo-replay.json`, then read both `curriculum.json` and `curriculum-quality-metrics.json`.
+2. Build a candidate set from the five highest-dropout steps and the five lowest-scoring curriculum steps.
+3. Semantically score and rank every candidate using the `stateReadiness`, `pathClarity`, and `recoverySupport` dimensions defined below, then select the three highest-risk steps for detailed analysis.
+4. Inspect every workshop page mapped to each selected step, plus the preceding activities that produce its required state, before writing `/tmp/gh-aw/agent/sim/data/agent-step-insights.json`.
+5. A simulated step may map to several pages; do not treat a file limit as a step limit. For example, inspect the learner's Step 7 authoring path (Terminal, GitHub UI, or GitHub Copilot) and the shared Step 7d model-access activity before adjusting Step 8.
 
 Evaluate content assumptions semantically instead of inferring state from keyword counts. Use one focused yes/no question per assumption, answer only `YES`, `NO`, or `UNKNOWN`, and cite `file:line` evidence. Copy each step's `contentHash` from the baseline output into `evaluatedContentHash`. The simulator ignores evaluations when that hash does not match the current mapped page content, so page edits require fresh evaluations.