feat(seats): mimo-9b-agent becomes amd-gcn's bound agent seat

[dmmdea]

Summary

• amd-gcn (Ryzen 5 5625U / Vega 7 iGPU, Vulkan) now sets include_mimo_9b: true and binds config_seed.agent_model to mimo-9b-agent, keeping include_qwen35_4b: true as the rollback seat. MEASURED 2026-09-24 on the amd-gcn reference box (llama.cpp b11153 Vulkan): mimo-9b-agent ties qwen3.5-4b-agent on quality (shape B 18/18 vs 17/18, shape C 4/5 vs 5/5) at roughly half the wall (333s vs 614s).
• servingtmpl.Params.validate() no longer refuses IncludeMimo9B && IncludeQ354B. A new __Q354B_AGENT_ALIAS__ token (mirroring the existing __Q359B_AGENT_ALIAS__) drops qwen3.5-4b-agent's own claim on agent-seat when mimo-9b-agent also renders, so the 4B entry stays as an un-aliased rollback instead of being refused. The pre-existing include_qwen35_4b/include_qwen35_9b mutual exclusion is untouched.
• llama-swap.linux-vulkan.yaml's mimo-9b-agent block now serves the tier's own window/KV/flash-attn/cache-ram (__CTX__/__KV_K__/__KV_V__/__FLASH_ATTN__/__CACHE_RAM__) instead of the CUDA 8GB tiers' literal 65536/q8_0 pin — a Vulkan UMA box has no fixed 8GB-class fit to pin to. win-cuda.yaml and linux-cuda.yaml's mimo blocks are unchanged (still the literal CUDA pin).
• Not originally scoped, required by two existing gates: llama-swap.win-vulkan.yaml gained a mimo-9b-agent entry (same tier-token shape) it never had. TestEveryBoundAliasIsServed and TestVulkanAndCPUTiersRenderOnLinux both assert every vulkan-backend tier renders on BOTH operating systems ("a tier is a hardware class; the OS it boots must not decide whether it exists"). Setting include_mimo_9b on amd-gcn without this addition made the Windows render of the tier refuse outright (win-vulkan.yaml had no mimo-9b-agent entry), which is exactly the OS-decides-capability failure those two gates exist to catch.
• REQUIRES llama.cpp >= b11102 (upstream llama.cpp #29319) — documented in the template comments, the tier notes, and docs/tiers/amd-gcn.md.

Tiers that render linux-vulkan.yaml / win-vulkan.yaml

Three tiers share backend: "vulkan": amd-gcn, amd-rdna3, amd-rdna3-dgpu.

• amd-gcn: the only one affected — gains the mimo-9b-agent seat as described above.
• amd-rdna3 / amd-rdna3-dgpu: functionally unchanged. Neither sets include_mimo_9b, so the new mimo block strips out of their renders on both OSes exactly as before (no new model, no alias, no matrix membership). Byte-diffed both tiers on both OSes before/after: the only differences are the provenance stamp (harness version bump) and inert prose comment lines preceding the stripped block — the same "comment survives a dropped model" characteristic every other gated seat in these templates already has (verified against the pre-existing mimo comment in win-cuda.yaml, which leaks identically for e.g. blackwell-16).

Rendered output (verification)

go run . install render --profile amd-gcn --os linux --home /opt/offload --llama-bin /opt/llama-vulkan --models /opt/models --threads 6 --listen 127.0.0.1:11436 --root .

mimo-9b-agent:

mimo-9b-agent:
  aliases: [mimo-9b, agent-seat]
  env: ["${ld}", "${vk}"]
  cmd: >-
    /opt/llama-vulkan/llama-server --model /opt/models/MiMo-V2.6-Distill-Qwen-9B-Q4_K_M.gguf
    --n-gpu-layers 999 --parallel 1 --ctx-size 32768 --flash-attn on
    --cache-type-k f16 --cache-type-v f16 --cache-ram 8192  --threads 6
    --reasoning off --jinja --port ${PORT} --host 127.0.0.1
  checkEndpoint: /health
  ttl: 300

qwen3.5-4b-agent (rollback, alias dropped):

qwen3.5-4b-agent:
  aliases: [qwen35-4b]
  env: ["${ld}", "${vk}"]
  cmd: >-
    /opt/llama-vulkan/llama-server --model /opt/models/Qwen3.5-4B-UD-Q4_K_XL.gguf
    --n-gpu-layers 999 --parallel 1 --ctx-size 32768 --flash-attn on
    --cache-type-k f16 --cache-type-v f16 --cache-ram 8192  --threads 6
    --jinja --port ${PORT} --host 127.0.0.1
  checkEndpoint: /health
  ttl: 300

(The double space before --threads is the empty slot-save token, [slot-save token output] in the task's expected shape — unchanged, unrelated to this PR.)

Tests

• go test ./...: all green (one unrelated pre-existing flake, TestDrainWaitsForARegisteredRunAcrossTheStepGap in gpu_drain_runs_test.go, a live GPU-lease timing test under full-suite load — passes 3/3 in isolation, file untouched by this PR).
• go vet ./...: clean.
• setup/render.tests.ps1: ALL PASS (new amd-gcn assertions for the mimo seat, the alias handoff, and the tier-driven ctx/KV).
• setup/tests/install-config-seed.test.ps1: ALL PASS (replaced the table-wide include_qwen35_4b/include_mimo_9b mutual-exclusion assertion — the exact rule this PR changes — with amd-gcn-specific assertions mirroring the existing blackwell-8/ampere-8 ones).

New/changed tests, each proven able to fail by reverting its production line:

• TestMimoAndQwen354BRenderTogetherQwenLosesTheAlias (replaces TestMimoAndQwen354BAreRefused) — red when validate()'s old refusal is restored.
• TestMimo9BSeatKeepsItsMeasuredInvariants (linux-vulkan/win-vulkan branch) — red when the mimo block is reverted to the CUDA literal pin.
• TestAmdGcnSeedsTheSeatItWasMeasuredOn (Go) and the new amd-gcn assertions in both .ps1 suites — red when agent_model/include_mimo_9b are reverted.
• TestAgentWindowMatchesWhatTheAgentSeatServes's backend now resolves a tier's agent-seat ctx from the template it actually renders (linux-vulkan for vulkan, win-cuda otherwise) instead of always reading win-cuda.yaml — a latent bug the vulkan+mimo combination exposed (win-cuda's mimo entry is a literal 65536 while amd-gcn's own entry is __CTX__/32768).

Not done / findings

• install.sh (the Linux install path) does no model downloads at all by design (documented as an operator prerequisite in its header) — there is no separate Linux download gate to wire up. install.ps1's include_mimo_9b gate is purely profile-id-keyed, so amd-gcn triggers it automatically with no code change.
• agent_seat_tok_s (10) stays as measured for the previous qwen3.5-4b-agent seat — not touched, since the task's config_seed instructions named only agent_model. Flagging in case the operator wants a mimo-specific rate re-measured for contract-wall sizing.
• Two Vulkan-only templates (win-vulkan.yaml, linux-vulkan.yaml) also carry a __Q354B_ALT__/win-vulkan.yaml gap in q354b_deadtoken_test.go's pre-existing tokenTemplates contract map (it only lists linux-cuda/win-cuda for that token, missing both vulkan templates) — pre-existing, not touched by this PR, noted for a future pass.

🤖 Generated with Claude Code