fix(map-plan): add normalize_blueprint repair pass for decomposer drift (#168)

[azalio]

Summary

Fixes #168. task-decomposer (Step 5) routinely emits a blueprint.json that fails the framework's own validate_blueprint_contract (Step 5.6), forcing manual JSON surgery before planning can proceed. Two mechanical self-consistency drifts recur:

1. Forward-dependency ordering — a subtask declared before the dependency it declares. The runtime walker consumes subtasks[] in declaration order (the validator's deadlock guard), so this is rejected even though the dependency DAG is fine; only the array position is wrong.
2. coverage_map ↔ bracket-tag mismatch — a coverage_map[req] = owner whose owner subtask's validation_criteria does not cite [req].

What this does

Adds a deterministic normalize_blueprint repair pass (runner function + CLI + /map-plan Step 5.55, before validation) so the flow is self-serve — decompose → normalize → validate → proceed — with no hand-editing:

• Stable topological sort of subtasks[] so every dependency is declared before its dependents. Independent subtasks keep their relative order (minimal diff); a true cycle is left untouched so validate_blueprint_contract still reports it.
• Bracket-tag injection: for every coverage_map[req] = owner whose owner lacks [req], appends a [req]-tagged validation_criteria entry.

Why a deterministic repair pass rather than loosening the validator or re-prompting the agent: the runtime walker relies on declaration order (loosening would risk a dependent running before its dependency), and a prompt fix is unreliable — the issue notes it fails "routinely" despite the decomposer being prompted with the contract. A topo-sort is safe under either scheduling model.

Conservative by design: never invents coverage_map ownership, never rewrites dependency edges, never touches a soft constraint that relies on tradeoff_rationale. Genuine semantic gaps (a hard constraint missing from coverage_map, an unknown/cyclic dependency) still fail Step 5.6. Idempotent — changed: false and writes nothing on already-normalized input.

python3 .map/scripts/map_step_runner.py normalize_blueprint [<path>] [--check]

Tests

New regression tests in tests/test_map_step_runner.py:

• test_normalize_blueprint_fixes_issue_168_drift — reproduces the exact task-decomposer emits blueprints that fail validate_blueprint_contract (forward-dep array order + missing coverage bracket-tags) #168 scenario (both drifts at once) and asserts validate_blueprint_contract flips false → true after normalize.
• idempotency, stable-order preservation for independent subtasks, cycle pass-through, and CLI --check/in-place paths.

Invariants honored

• Edited the .jinja single source and re-rendered (make render-templates); all generated trees and the live .map/scripts/ copy stay byte-identical (make check-render).
• make check green: ruff + mypy + pyright clean, 2291 passed, 3 skipped.

🤖 Generated with Claude Code