docs(plans): apply review feedback to init/increment perf plan

5-lens subagent review of the plan found the PR-P1/P2 boundary was
architecturally wrong: the graph write helpers are SHARED between the full
and incremental paths, so a "full-path-only" split is impossible.
- Verified call graph: _write_edges/_write_routes_and_exposes/_write_nodes_impl/
  _write_meta are each called by BOTH paths; _write_clients_producers_and_calls
  is incremental-only (global pass5/6).
- Re-split by write-FUNCTION: PR-P1 = _bulk_copy + _write_edges (the ~250s
  prize, accelerates both paths); PR-P2 = _write_nodes_impl +
  _write_routes_and_exposes + _write_clients_producers_and_calls; PR-P3 = ignore
  cache (independent).
- GraphMeta (_write_meta) left on MERGE (shared, one row) — reverses Open Q1.
- Fixed all binding sentinel greps: PR-P1 zeros the edge _CREATE_* only;
  PR-P2 zeros node/route/client constants + _MERGE_SYMBOL only after both
  routes functions convert; PR-P3 sentinel narrowed to
  LayeredIgnore(project_root).is_ignored (the bare-constructor grep wrongly
  matched once-per-run sites :177/:569, which are correctly left alone).
- Load-order §1f corrected (UnresolvedCallSite before UNRESOLVED_AT;
  Route/Client/Producer before their edges). Test files qualified
  (test_brownfield_routes / test_mcp_v2_compose / test_vectors_progress /
  test_path_filtering). PR-P2 tests placed in TestIncrementalOrchestrator.
  Baseline flagged as equivalence anchor, not production invariant.
  PR-P1 DoD lists the four test names.

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

Author: HumanBean17