registry: add 6 more Arm software tools (with scenarios)

[kavya-chennoju]

Builds on the merged #1 (context-first registry). Seeds 6 more Arm software tools so the registry is a real menu across all six contexts — each with a when-to-use and a situation → action → outcome scenario.

Added

Tool	Category	Contexts
Arm Toolchain	sdks-tools	compiled-languages, embedded-development
Neural Graphics Model Gym	sdks-tools	ml-developer, mobile-games
Arm ML Inference Advisor (MLIA)	profiling-optimization	ml-developer, performance-optimization
Arm Accuracy Super Resolution (ASR)	sdks-tools	mobile-games
Metis	sdks-tools	cloud-development, compiled-languages
Arm ML SDK for Vulkan	sdks-tools	ml-developer, mobile-games

Multi-context tools are copied identically into each of their context files (drift guard green). Populates the previously-empty mobile-games and embedded-development contexts.

Registry is now 8 resources / 17 listings across all 6 contexts.

Notes

• All 6 are public repos in the arm org.
• Metis is general-purpose AI security code review, not Arm-hardware-specific — flagged as such in its entry, included as the org's most-starred tool. Drop it if the registry should stay strictly Arm-specific.
• scripts/validate.py green (schema + drift guard + counts + alphabetical order).

[atsyplikhin]

Validated all 6 when_to_use / example_use_case entries against the actual repos. Looks great overall -- one real fix, one minor reword, everything else checks out.

Fix needed: Arm Toolchain example_use_case (identical copy in compiled-languages.yaml and embedded-development.yaml): the scenario is a Cortex-M firmware build, but the text says the agent "sets up the Arm Toolchain (LLVM/Clang for AArch64)". Cortex-M is 32-bit (AArch32/Thumb), so "for AArch64" is wrong in this scenario. The repo ships two flavors: Arm Toolchain for Embedded (bare-metal, which is what a Cortex-M build uses) and Arm Toolchain for Linux (native AArch64). Suggest "sets up Arm Toolchain for Embedded (LLVM/Clang)" and drop the AArch64 parenthetical. The when_to_use itself is accurate.

Minor: Arm ML SDK for Vulkan when_to_use -- "so models execute on any Vulkan-compute-capable (mobile) GPU" slightly overstates it. Native execution needs a driver implementing the ML extensions for Vulkan; what gives you "any Vulkan GPU" is the SDK's Emulation Layer, meant for development. Something like "on Vulkan-capable GPUs, with an emulation layer for development on any Vulkan device" would be exact. The VGF-conversion and scenario-runner details in the example are correct.

Verified accurate:

• Metis -- matches the README nearly verbatim (open-source, AI-driven deep security code review; LLM reasoning vs. hardcoded rules); the validate-findings/re-run flow matches its issue-validation feature. On the open question: I'd keep it -- it's from Arm's Product Security Team and the entry already flags it as not Arm-hardware-specific.
• Neural Graphics Model Gym -- train/finetune/evaluate + quantize/export for the Vulkan ML extensions, NSS use case: all match.
• MLIA -- operator compatibility, performance estimates, pruning/clustering guidance for Ethos-U/Cortex targets: accurate.
• Arm ASR -- mobile-optimized temporal upscaling derived from AMD FSR2, render-low-and-upscale flow: accurate.
• Structure -- counts (3/3/3/4/1/3 = 17 listings, 8 resources), alphabetical order, and identical multi-context copies all hold.

Non-blocking nit: the diff rewraps the three pre-existing entries (whitespace-only reflow across 4 files), which adds review noise -- worth avoiding in future PRs.

[kavya-chennoju]

Thanks for the careful validation — all three addressed in 548542f.

• Arm Toolchain example (fix): you're right, Cortex-M is AArch32/Thumb — changed to "sets up Arm Toolchain for Embedded (LLVM/Clang)" and dropped the AArch64 parenthetical. Also named both flavors in when_to_use (Embedded for bare-metal, Linux for native AArch64).
• Arm ML SDK for Vulkan (reword): taken — now "through the ML extensions for Vulkan — natively on GPUs whose driver implements those extensions, with an emulation layer for development on any Vulkan device."
• Reflow nit: fixed, not just noted — I regenerated with the two pre-existing entries (MCP, Performix) emitted verbatim from main, so the diff is now purely additive (e.g. cloud-development shows only the Metis block). Going forward these land append-only.
• Metis: kept, per your call (Arm Product Security Team; entry already flags it as not Arm-hardware-specific).

Validator green, drift guard holds.

[atsyplikhin]

Re-reviewed at 548542f -- all three points addressed, LGTM.

• Arm Toolchain: example now says "sets up Arm Toolchain for Embedded (LLVM/Clang)" with the AArch64 parenthetical gone, and when_to_use naming both flavors is a nice touch. Correct on both counts.
• ML SDK for Vulkan: the new wording (native on drivers implementing the extensions, emulation layer for development on any Vulkan device) is exactly right.
• Diff hygiene: confirmed the resource entries are now purely additive against main. (Two files still carry a rewrapped header comment line -- ml-developer and performance-optimization -- but that is cosmetic and not worth another push.)
• Also spotted the new "(derived from AMD FSR2)" addition to the ASR entry -- verified against the README, accurate.

Counts (3/3/3/4/1/3), alphabetical order, and multi-context copies all hold, and the validate check is green. Good to merge from my side.