fix(tiers): seed blackwell-8 and ampere-16 with 2026-09-23/24 measured values

[dmmdea]

Summary

Both reference boxes (OptiPlex 7060 / blackwell-8, Lenovo M720q A2 / ampere-16) got a harness
remediation pass on 2026-09-23/24. This carries the measurements they produced into the tier
seed (setup/templates/profiles.json) so a fresh install gets them instead of stale defaults,
and regenerates docs/tiers/*.md (go run ./cmd/gentiers).

blackwell-8 (OptiPlex 7060, RTX 5060 8GB), all under config_seed_ram_mid_high unless noted:

• imagegen_timeout_sec 900 → 2400 — the qwen-image-2.1 binding's own measured max wall is 678.1s
  at 2752×1536 under the power tune (900 left only ~1.3x margin, 2400 leaves ~3.5x).
• sdcpp_extra_args added for z-image-turbo: the measured-working graph-cut arm
  (--vae-tiling --offload-to-cpu --diffusion-fa --max-vram 6.5 --stream-layers), 56-69s/1024²
  image. Two failed arms are explicitly not seeded: bare --vae-tiling silently ran on the
  Intel UHD 630 iGPU (Vulkan device-pick default), and an all-VRAM placement OOM'd/spilled at
  565-608s/step.
• videogen_wan_virtual_vram_gb added at 12 (measured: 7 OOMs at lower values under the
  no-sysmem-fallback policy).
• videogen_timeout_sec 3600 → 4500 = ceil(2,970.3 × 1.5), from a measured cold native 81-frame
  render of 2,970.3s.
• audiogen_timeout_sec added at 1500 in config_seed (previously unbound, framework default
  720): sized from a measured cold ACE-Step render (421.9s) + the audio-QA dead-air retry
  (~420s) + loudnorm ≈ 870s measured, ×1.5 rounded up; also bounds the measured 108.7s cold
  voice (Chatterbox) render.
• imagegen_families adds two named, opt-in bindings beside the default hidream-o1-dev /
  z-image-turbo pair (ADR 0058, ram_mid_high only — both need ≥64GB host RAM for staging):
  qwen-image-2.1 (BINDING.md's canonical bf16 overlay, verbatim; A2 evaluation CLEARED 16/16)
  and qwen-image / Qwen-Image-2512 (Apache-2.0, commercial; measured 190.9s at 1328² with
  exact text rendering).
• Not seeded, on purpose: a qwen-image-2.1-int8 checkpoint (benched — 2/2 clean, ~1.6x faster,
  but fidelity split 1/2 over the LPIPS≤0.03 bar — and not operator-authorized as a named
  opt-in) and animate_character (no config key exists for the route on this box yet; its
  model files are unmeasured at this VRAM class).

ampere-16 (Lenovo M720q, NVIDIA A2 16GB):

• videogen_wan_virtual_vram_gb added at 9 — the tier declared Wan I2V keys but never this one,
  so a fresh install silently rendered the graph script's own default (7). A live smoke render
  measured a peak of 9.6 of 15.0 GiB across the full render at 9 GiB, 5.4 GiB headroom, zero
  OOM — a measured-safe floor, not yet a measured optimum (the smoke used a reduced frame
  count/resolution than the 81-frame production default).

Confirmed unchanged, already correct (checked, no edit needed): blackwell-3x16's
comfy_cuda_device: "2" (matches the 2026-09-24 measurement and displaycard_placement_test.go's
documented rule) and videogen_pool_compute: "cuda:0" (the ComfyUI-MultiGPU #220 exception).

Not touched, and why: ampere-8's comfy_extra_args temp/output-dir redirect is a measured,
working value, but a literal per-box drive path (D:/ComfyUI-temp etc.), not a tier-general
rule — confirmed by dedicated research against the source docs. ffmpeg_path on any tier: PR
#471 already makes a bare "ffmpeg" resolve via PATH, and no tier currently seeds a broken
value.

Also bumps VERSION 0.140.11 → 0.140.12 (and the other three version sources
TestVersionSourcesAgree checks: internal/buildinfo.Version, .printing-press.json,
CHANGELOG.md) with a changelog entry for this change.

Test plan

• go build ./... — clean
• go test ./... — all packages pass, including displaycard_placement_test.go,
  media_roster_test.go, docs_tiers_test.go, TestVersionSourcesAgree, and every
  internal/tierseed, internal/config, internal/pipeline, internal/tierdocs suite
• go run ./cmd/gentiers regenerated exactly docs/tiers/ampere-16.md and
  docs/tiers/blackwell-8.md (matches the profiles.json diff), and the staleness gate
  (TestTierDocsAreCurrent) passes
• setup/templates/profiles.json stays CRLF throughout (checked byte-for-byte before and
  after every edit)
• Rebased onto latest origin/main (post PR feat(seats): mimo-9b-agent becomes amd-gcn's bound agent seat #474 / 0.140.11) and rechecked VERSION
  immediately before push

🤖 Generated with Claude Code