fix(LTX25-DEVICE-SEAM-SIBLING): the third question, and a leakage bucket that is a property rather than a spelling (#659, #660)

.agents/roadmap_v1.md

@@ -207,8 +207,8 @@ issue is not yet placed. Keyed record: update in place, never append.
 | [#758](https://github.com/mudler/vllm.cpp/issues/758) | `ENG-MM-INPUT-PIPELINE` | A multimodal refusal cannot distinguish a configured limit from an UNIMPLEMENTED arm. `Qwen3VLChatSupportedMmLimits()` declares the seam's honest ceiling `{"image": 1}` with video/audio absent, but the message a client gets is upstream's generic `At most 0 video(s) may be provided in one prompt.` — identical to what `--limit-mm-per-prompt '{"video": 0}'` produces. AGENTS.md requires an unimplemented arm be refused "with a message naming the missing piece"; #749 claimed the ceiling satisfies that and its review found it does not. The only present signal is by OMISSION (`ValidateNumItems` withholds the `--limit-mm-per-prompt` hint when raising the limit would not help). Not fixed in flow: naming the arm diverges from a verbatim-ported message three suites assert byte-for-byte, so it needs its own spec and fresh review. Found in the #749 review round (#607 wave L2, #686) | bug |
 | [#651](https://github.com/mudler/vllm.cpp/issues/651) | — | `test_agent_record`'s MODEL-ratchet docstring is two contradictory paragraphs spliced together, and the surviving half records a pin transition that never happened | bug |
 | [#652](https://github.com/mudler/vllm.cpp/issues/652) | — | `model-matrix.md` prose counters drifted: LTX-2.5 reached the rows and the CI-enforced rollup but none of the five sentences that count them | bug |
-| [#659](https://github.com/mudler/vllm.cpp/issues/659) | — | LTX-2.5 device select adopts M3a's platform seam but not its companion capability guard: `ltx2_video.cpp` asks `CurrentPlatform().device_type()` and `TryGetBackend(...)` but never `supports_model_architecture`, so a PARTIAL backend (Metal 15/75 ops, Tenstorrent) is handed a queue and dies in a kernel bind where it used to be refused BY NAME (found while reviewing #553 for landing) | bug |
-| [#660](https://github.com/mudler/vllm.cpp/issues/660) | — | `check-device-leakage`'s `kcuda` bucket is the token grep `\bkCUDA\b`, so `minimax_h3_video.cpp:221-226`'s `static_cast<vt::DeviceType>(device)` hardcodes CUDA as enum value 1 and counts as 0. Gate strength plus an enum-ordering hazard; the H3 video lane should ask the same seam `ltx2_video.cpp` now does (found while reviewing #553) | bug |
+| [#659](https://github.com/mudler/vllm.cpp/issues/659) | `LTX25-DEVICE-SEAM-SIBLING` | Spec [`ltx25-device-seam-sibling.md`](specs/ltx25-device-seam-sibling.md), `READY`. LTX-2.5 device select adopts M3a's platform seam but not its companion capability guard: `ltx2_video.cpp` asks `CurrentPlatform().device_type()` and `TryGetBackend(...)` but never `supports_model_architecture`, so a PARTIAL backend (Metal 15/75 ops, Tenstorrent) is handed a queue and dies in a kernel bind where it used to be refused BY NAME (found while reviewing #553 for landing). Guard added at the same site; `tests/vllm/multimodal/test_diffusion_device_seam.cpp` holds it against an injected declining platform, and `test_minimax_h3_video_fold.cpp`'s own predicate asks all THREE questions — asking two made it red on a correct refusal, i.e. on exactly the partial-backend build this issue exists to serve | bug |
+| [#660](https://github.com/mudler/vllm.cpp/issues/660) | `LTX25-DEVICE-SEAM-SIBLING` | Spec [`ltx25-device-seam-sibling.md`](specs/ltx25-device-seam-sibling.md), `READY`. `check-device-leakage`'s `kcuda` bucket is the token grep `\bkCUDA\b`, so `minimax_h3_video.cpp:221-226 @ 11cc1d589`'s `static_cast<vt::DeviceType>(device)` hardcodes CUDA as enum value 1 and counts as 0. Gate strength plus an enum-ordering hazard; the H3 video lane should ask the same seam `ltx2_video.cpp` now does (found while reviewing #553). New `dev_cast` bucket, with **29** individually-asserted RED spellings that are not in the tree (counted, not quoted: M20-M24, M29-M33, M35-M39, M41-M44) — including the literal operand `(vt::DeviceType)1`, which the first cut MISSED while catching `(vt::DeviceType)d`, and whose own discriminator test used an identifier and so could not see the gap it was written to rule out. Rounds 2-4 each found the same shape once more IN THE MESSAGE: round 2, nine spellings that wrote the target type the docstring named as its anchor; round 4, eight more — a C-style REFERENCE and POINTER target, the DECLARATION form's cv-qualifiers in three spellings (a FOURTH cv-position, while M36 asserted "all three"), `bit_cast`'s two-argument form, a cv-qualifier inside the pointer run, and `__builtin_bit_cast`. All compile-verified legal, all now RED, measured 0 new / 0 lost over the scan roots (**760** files at `79ebbce42`, **765** after the merge — the count rots on every unrelated PR, so it carries its SHA) with `\bDeviceType\b`=162 and `\bkCUDA\b`=18 as positive controls in the same pass, unchanged across the merge; `sizeof (vt::DeviceType) + 1` was a pre-existing FALSE positive and is now excluded. The docstring no longer claims the property: it enumerates the spellings it matches, states each blind spot with a reason true of that entry alone (`*(vt::DeviceType*)&x`, a character-literal operand, `memcpy`), and names the residual as a property a text checker cannot decide — [#828](https://github.com/mudler/vllm.cpp/issues/828) tracks the AST-level check. M34/M40/M45 pin the negatives; M46 pins that the declared blind spots are still blind, so message and pattern cannot drift apart silently | bug |
 | [#674](https://github.com/mudler/vllm.cpp/issues/674) | — | `main` is RED on `sanitize-cpu (address,undefined)` since `cefacd2d0`: `Ltx2LoadVaeWeights` (`ltx2_loader.cpp:1325`) reinterpret_casts the safetensors mmap to `const uint16_t*`, and that payload offset carries NO alignment guarantee — UB everywhere, a real fault on `build-test-cpu-arm64` and Jetson/Orin sm_110. THIRD recurrence of one class after [#301](https://github.com/mudler/vllm.cpp/issues/301) (closed; it left the `vt::LoadUnaligned` seam) and [#627](https://github.com/mudler/vllm.cpp/issues/627) (`qwen3_5_weights.cpp`, still open). The coverage that caught it was ACCIDENTAL — the fixture's JSON header happens to land that tensor odd — so the fix owes a case that FORCES the odd offset and asserts the parity | bug |
 | [#664](https://github.com/mudler/vllm.cpp/issues/664) | — | **FIXED 2026-08-13, `row/FIX-WINDOWS-POSIX-VIDEO-ENGINE`.** `windows-msvc-cpu` / `windows-msvc-vulkan` were RED on EVERY open PR (9 sampled across 5 unrelated lanes) from ONE file: `video_engine.cpp` reached Windows with `<sys/stat.h>`, `::stat` and `S_ISDIR` (landed `cefacd2d0`), which `check-windows-portability.py:1675-1688` flags under `full_source_posix` — every scanned source, not only the platform-boundary set. `main` was never a denominator because the Windows jobs are PR-only and `skipped` on push (#584). Repaired at the SOURCE, not the checker: `IsDir`/`Exists` now take the `std::error_code` <filesystem> overloads through a file-local `NativePath`, which preserves `::stat`'s return-false-for-an-uninspectable-path behaviour that the THROWING overloads would have turned into a `filesystem_error` escaping a registry query | bug |
 | [#608](https://github.com/mudler/vllm.cpp/issues/608) | `TOOLS-PARSER-BREADTH` | **W0 (record backfill) landed 2026-08-13; row `INVENTORIED` → `PARTIAL`. W1 then shipped `inkling`, taking the registry to 42 names / 38 families.** Re-derived from the two REGISTRIES rather than from recipe usage: **five** `--tool-call-parser` names are upstream-only at the pin — `openai`, `inkling`, `minimax_m3` (W1, recipe demand) and `cohere_command3`, `cohere_command4` (W2, ZERO recipe uses, so usage-driven audits miss them). **Only `inkling` was PORTABLE from vLLM source, and it LANDED** (a `ParserEngineToolAdapter` over the already-ported Inkling engine; the gap was the registry face, not the grammar). Of the four left: `minimax_m3` is backed by the Rust crate; `openai` is a stub delegating to `vllm/parser/harmony.py`, which IS vLLM source but wraps the out-of-tree `openai_harmony` package (the SGLang secondary-oracle check was run and REFUSED — vLLM implements this path, so the rule does not admit a secondary); both Cohere names are shims over the out-of-tree `cohere_melody` package. So W1-remaining/W2 each owe a recorded decision before code rather than a text port. `nemotron_json`, `kimi_k3` and `ling3` are in NEITHER registry and arrive with the pin advance, not here. W3 ports upstream's shared `ToolParserTestConfig` harness. The earlier "six missing" framing was usage-derived: it listed `nemotron_json` as portable (it is not registered at the pin) and missed both Cohere entries | feature |