[DmitryMalishev]

Task: QVAC-27614

🎯 What problem does this PR solve?

• Fixes Issue 2 — Cached tiled-image geometry is lost. Fabric's server serializes image chunks into placeholders without preserving grid_x/grid_y. Some valid Qwen-style multi-grid images then have incorrect token counts and decoder positions, causing HTTP 500 / invalid input batch failures. Runtime/addon PR #5489 integrates this engine fix through the shared Fabric package.
• Fixes Issue 12 — Image resize rounding differs from the reference processor. Half-way dimensions round upward instead of using ties-to-even, potentially adding an image-token row or column. Runtime/addon PR #5489 delivers the corrected runtime and adds an addon-level boundary regression.
• Issue numbers refer to the reported issue list. Issue 3 — SDK completion stop reasons is addressed independently in SDK PR #5516.

📝 How does it solve it?

• The b11739 addon compatibility changes and temporary overlay are maintained in QVAC #5331. Runtime/addon Update README.md for llava ggml-org/llama.cpp#5489 needs those adaptations for end-to-end validation against this rebased engine; changing only its source pin is insufficient.
• Serialize and restore both tile-grid dimensions. Bump the chunk format from version 3 to 4 and its minimum accepted version to 4. The new base already uses version 3 for per-image suffix/any-resolution metadata; versions 2 and 3 lack the token-grid fields and must be rejected.
• Test 1/2/3/4-tile round trips, placeholder token counts and decoder positions, version-2/version-3 rejection and truncated input. Keep the C API assertions active in release builds.
• Use integer ties-to-even rounding in the shared image-sizing helper. Cover half-way and adjacent dimensions, pixel budgets, existing sizing callers and all four floating-point rounding modes.
• Target temp-11739 (upstream b11739 integration), rebased directly onto base 14ecfda0c (merged ci : drop duplicate permissions block in pre-tokenizer-hashes #480). After this merges, publish the updated vcpkg port and replace the temporary overlay in QVAC #5489 with its official dependency before merging that integration PR. Public model/addon APIs are unchanged.

🧪 How was it tested?

• Current head 97159252e is rebased directly onto temp-11739 at 14ecfda0c, including #480. Only five mtmd implementation/test files differ from the base. git diff --check passed.
• This final rebase changes only the inherited pre-tokenizer workflow relative to tested c3edc86c2; the mtmd implementation and regression files are identical. That verified source passed all three focused mtmd targets under ASan+UBSan with leak detection and halt-on-error; cache tests reject versions 2 and 3.
• Fresh current-head CI is running: CPU, server, CUDA, Vulkan, Metal, Apple and sanitizers. Pending or queued jobs are not claimed green.
• The inherited duplicate-permissions workflow error is corrected by merged ci : drop duplicate permissions block in pre-tokenizer-hashes #480. The verified archive contains one top-level permissions block; the current push has no corresponding invalid-workflow failure.
• QVAC #5331 now combines the b11739 native API adaptations with this exact engine pin and verified SHA-512 (overlay commit). All 51 local overlay/runtime policy tests passed; Fabric prebuild and C++/desktop end-to-end validation is running.
• Historical server counterfactual/replay on old-base 4fe4ded85 passed 100 sequential GPU requests, 14 batched controls, four CPU grid cases, two resize-boundary requests, slot restore and stale-cache rejection. Those results and the old addon overlay do not certify the new base. SDK stop-reason PR server : fix system prompt cli ggml-org/llama.cpp#5516 remains independent.

💥 Breaking Changes

• Serialized mtmd chunk format changes from version 3 to 4; versions 2 and 3 are rejected. Existing persisted media-chunk caches must be recreated; old chunks are rejected rather than restored with incorrect geometry.

AI assistance: Codex assisted implementation and validation; human maintainer review is required.