[perNyfelt]

• Package-private Qwen3VlPromptExpander plus position planner family
  (Qwen3VlPromptTokens/ImagePlacement/RopePlan/PositionPlanner/PromptPlan),
  a faithful port of the pinned get_rope_index image path (transformers
  4.57.6 modular_qwen3_vl.py:858-955) with checked arithmetic that
  rejects overflow and past-the-prompt grids before any array is written.
• Token contract: the four real Qwen sentinels (hex-verified against the
  pinned 2B artifact); Qwen3VlPromptTokens.resolve takes each sentinel
  name to its ID through the tokenizer and cross-checks the outer
  config.json ID. No hard-coded numeric IDs in production code.
• Golden tools/hf-reference/goldens/qwen3-vl-wp3-prompt.json (8 cases x 2
  add_generation_prompt variants, generated by qwen3vl_wp3_prompt.py),
  digest-pinned in tools/hf-reference/provenance.json and enforced by
  verifyHfReferenceGoldens; WP1-consistent expanded lengths 82/-56 and
  90/-60.
• 33 pure-Java tests (Qwen3VlPromptExpanderTest) over the golden, the
  resolve contract, the hand-computed position cases and the exact
  grammar-rejection messages, plus the qwen3vl-tokens fixture.
• Findings gain section 6.8 (grammar, rejection table, jmlx-tokenizer
  three-directions finding, planner port, evidence) and correct the
  earlier template-description line; plan section 5 records WP3
  acceptance. Corrects the earlier sentinel assumption, which was an
  artifact of terminal display stripping the 0x7c byte.