GCP router: add 8x A100 (a2-ultragpu-8g) + 8x H100 (a3-highgpu-8g, FLEX_START) machine specs so the free GCP lane can run wide parallel sweeps

Overview / Motivation

The GCP lane can only reach 1× and 4× A100 machines today — INTENT_TO_MACHINE (src/explore_persona_space/backends/gcp.py) maps lora-7b/lora → a2-ultragpu-1g (1× A100-80), ft-7b → a2-ultragpu-4g (4× A100-80), eval/debug → g2-standard-4 (1× L4), plus 1×/2× H100 (lora-7b-h100 → a3-highgpu-1g, eval-h100 → a3-highgpu-2g). There is no 8-GPU GCP machine spec, so a large embarrassingly-parallel sweep that would finish ~2× faster on 8 GPUs cannot use the free GCP credit lane and falls to a paid RunPod 8× box instead.

Driving case: #697's 64-cell causal-patch sweep (cells are independent, shard across GPUs via the dispatcher's --n-gpus). On 4× A100 the remaining work is ~14h; 8 GPUs would ~halve it — but the router can't size an 8-GPU GCP VM, so it went to paid RunPod.

Goal

Add 8-GPU GCP machine specs/intents to the router so big parallel sweeps can use the free GCP credit lane at 8× width:

• 8× A100-80 → a2-ultragpu-8g
• 8× H100-80 → a3-highgpu-8g (FLEX_START / SPOT only — a3 cannot be created on-demand; render_create_argv already raises on H100 + STANDARD)

Verified quota + availability (us-central1, project eps-persona-gpu-jun2026, 2026-06-29)

• PREEMPTIBLE_NVIDIA_A100_80GB_GPUS = 16 (4 in use → 12 free) — 8× A100 fits.
• preemptible_nvidia_h100_gpus (us-central1) = 8; preemptible_nvidia_h100_mega_gpus = 8 — exactly one a3-highgpu-8g, NO concurrency headroom (two 8× H100 pods at once would exceed quota).
• On-demand H100 ≈ 0 (only preemptible/committed buckets exist) — consistent with a3 being flex-start/spot-only.
• Machine types offered: a2-ultragpu-8g in us-central1-a/c; a3-highgpu-8g in us-central1-a/b/c.

Scope / files (grep first, then edit)

• src/explore_persona_space/backends/gcp.py — add MachineSpec rows to INTENT_TO_MACHINE (line ~283) for the two 8-GPU machines under sensibly-named intents (planner's call — e.g. an 8× A100 sweep intent + an 8× H100 sweep intent). The H100 intent MUST require provisioning_model SPOT|FLEX_START (mirror the existing lora-7b-h100/eval-h100 rows + the #631 comment block).
• src/explore_persona_space/backends/router.py — decide whether auto should ever select an 8-GPU rung (probably NOT by default — reserve 8× for an explicit intent, since the H100 quota is exactly 8 / no headroom and 8× A100 is heavy). At minimum the new intents must be dispatchable via an explicit --intent. Document the choice.
• scripts/gpu_heuristics.py — if an 8× GCP intent should have a GpuSpec analogue, add it (the RunPod inf-70b = 8× H100 row at line ~44 is the sibling).
• scripts/dispatch_issue.py — the --gpus-vs-machine-type preflight (GCP sizes from --intent alone) must accept the new intents.
• Tests: tests/test_gcp_backend.py — INTENT_TO_MACHINE coverage + a render_create_argv test for a3-highgpu-8g FLEX_START (and that H100 + STANDARD still raises) + a2-ultragpu-8g. tests/test_router*.py if the ladder changes.
• CLAUDE.md — GCP mechanics line (~217) lists the intent→machine map; add the 8-GPU rows.

Constraints / invariants

• The H100 a3 intents are FLEX_START/SPOT only (never STANDARD) — keep the existing H100+STANDARD raise.
• 8× H100 quota is exactly 8 in us-central1 → AT MOST one a3-highgpu-8g at a time; do not let auto schedule two concurrently. 8× A100 has more headroom (12 free).
• Flex-start CAPACITY is separate from quota — an 8× create can queue; the existing flex-start ladder/wait handling applies.
• Workflow-surface only; ruff on touched files + uv run pytest tests/test_gcp_backend.py tests/test_router.py pass; CLAUDE.md stays consistent with the code.
• Coordinate with #741 (the max-run-duration default change) if it touches the same gcp.py regions — they may merge close together.