fix: point pdd-change Step 6 dependency context at <pdd-dependency>, not <include>

[agarwal-ishaan]

Summary

• pdd change Step 6 builds its module-dependency context by scanning <include> tags (build_dependency_graph()), but <include> tags are LLM context, not declared architectural dependencies -- so the graph both misses real <pdd-dependency>-declared dependents and fabricates false ones from stylistic includes.
• Updates the source prompts to specify a new sync_order.build_dependency_graph_from_architecture() that reads architecture.json's dependencies field (already populated from <pdd-dependency> tags by architecture_sync.py), and wires Step 6's _build_dependency_context() to use it instead.
• Prompts only in this PR. The corresponding code regeneration (pdd generate --incremental on pdd/sync_order.py and pdd/agentic_change_orchestrator.py) is pending in a follow-up PR -- Anthropic API access is currently blocked by a billing issue, and the fallback Gemini model hits its output-token ceiling trying to patch the ~4000-line orchestrator file. Once resolved, the code PR will implement exactly what's specified here.

Fixes #1807

Test plan

• Follow-up PR regenerates pdd/sync_order.py and pdd/agentic_change_orchestrator.py from these prompts via pdd generate --incremental
• Unit tests for build_dependency_graph_from_architecture() and the updated _build_dependency_context()
• Manual repro: a module declaring <pdd-dependency> on another without <include>-ing it shows up correctly as a dependent; a module <include>-ing another purely for style context does not

[agarwal-ishaan]

Consolidating into a single PR per request -- superseded by #2376, which contains this PR's commits plus the rest of the fix. Closing without merging (nothing is lost; #2376's branch includes these exact commits).