Sprint 003 findings + Sprint 004 pipeline decision audit plan

- split_prereq_op_dataset.py: extract accepted/rejected/schema-drift rows per gate
- gen_prereq_op_rsa_data.py: convert accepted rows to Fabricate TSV format
- 1000-run extracted dataset: 1297 accepted prereq_op rows, 24 schema-drift
- Key finding: needs_validate_intake always True; effective gate is binary
  (needs_resolve_dependencies 77%, needs_architect_review 20%)
- Key finding: run corpus text is template-driven (17 unique inputs) — not
  suitable for text-based training; RSA exists before ideal input is available
- ROADMAP: gate lifecycle framing — goal is choice of backend, not forced
  determinism; RSA operates on imperfect/mixed/incomplete input by design
- training_data_strategy.md: taskitem-as-self-contained-work-unit constraint
- SPRINT-003: updated with 1000-run results, text audit findings, design constraints
- SPRINT-004: pipeline decision audit — manual walkthrough of every decision
  from project description to generated artifact; execute next session

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs/case_studies/whetstone_dsl_training_data_strategy.md

@@ -173,6 +173,37 @@ The right pattern is:
 
 This should produce a gate-native dataset rather than a project-native dataset.
 
+## Taskitem as Self-Contained Work Unit — Design Constraint
+
+**Added 2026-04-12.**
+
+A taskitem is a concurrent work unit. By design it must include all information a
+specialist needs to make its gate decision. The specialist does not reach back to
+the original spec or the intake.
+
+**Consequence for RSA gate design:**
+
+The taskitem schema must carry every field that is discriminative for the gate.
+`generate_taskitems` populates those fields from the intake contract. The RSA gate
+reads those fields from the taskitem — no other context is available at decision time.
+
+**Consequence for training data:**
+
+The current taskitem schema does not carry discriminative fields for most RSA gates.
+The run corpus has weak text (17 unique templates) because the schema was designed for
+a different purpose — it records outputs, not inputs.
+
+RSA gates are designed to operate on whatever input actually exists — including
+mixed structured/unstructured content and incomplete directions. Waiting for ideal
+input is not the RSA model.
+
+The current corpus is the training corpus. Sprint plan step descriptions are the best
+available text signal now. What the gate learns from imperfect inputs, and when it
+abstains on insufficient inputs, is the gate's actual contract.
+
+Richer taskitem fields (if added) give the gate more signal and may shift it toward a
+deterministic backend — but that is operator choice, not a prerequisite for deployment.
+
 ## Compatibility Constraint
 
 We should not assume the new specialists must fit perfectly into the exact old