fix: harden Claro diagnostics and runtime guards

Preserve source locations, validate malformed JSON, and add opt-in resource limits. Include executable error-guide coverage, learner pilot materials, native Windows acceptance helpers, and updated CI/documentation.

docs/LEARNER_FEEDBACK_TEMPLATE.md

@@ -0,0 +1,32 @@
+# Learner feedback — blank observation form
+
+Status: NOT RUN. Leave observations blank until a consenting person participates.
+Do not fill this form with predictions, agent simulations or invented participants.
+
+- Anonymous session ID:
+- Consent to retain anonymous notes: yes / no
+- Build version and artifact checksum:
+- OS/device and assistive tools (only if voluntarily relevant):
+- Tasks attempted / skipped:
+- Breaks, pacing and requested adjustments:
+
+For each attempted task record:
+
+- What the learner tried or said (anonymous, no health information):
+- What worked independently:
+- What help or clarification was requested:
+- Exact confusing instruction/error and its filename/line:
+- Whether the suggested repair helped:
+- Learner's suggested change:
+
+Use plain observations, not ability rankings. Do not infer a medical diagnosis.
+
+Debrief:
+
+- Clearest part:
+- Most confusing part:
+- One small change to try:
+- Follow-up session agreed, if any:
+- Notes retention/deletion agreement:
+
+Automated example execution verifies code/output only, not these observations.