Overview / Motivation

Auto-filed workflow-fix from the #825 turn-dynamics-5000 incident (2026-07-16): the MAIN phase requested an explicit 8-GPU sweep intent and starved for 2.5h+ on capacity, because explicit wide intents do NOT auto-degrade width on capacity failure.

Goal

On repeated 8-GPU (wide) capacity failure, degrade the GCP GPU width (8→4→2) for EXPLICIT sweep-8g intents too — not only for --gpus N declarations — so a wide-sweep phase falls back to abundant partial-node capacity instead of starving on empty 8-GPU DWS pools.

Workflow gap

• Bug observed: #825 round-11 MAIN phase (sweep-8g-a100, then sweep-8g-h100) repeatedly hit terminal_no_compute_available / FLEX_START queue-vanish (#1116) on BOTH us-central1 8-GPU DWS pools; it never fell back to 4×/2× (which have abundant capacity) and stayed stuck 2.5h+. An orchestrator free-text "narrow to 4×" directive was NOT honored (the width is baked into the explicit intent).
• Why it is a workflow gap: the width-aware auto-lane (#1121, WIDE_A100_80_BY_WIDTH / _requested_wide_widths / _gcp_ladder_specs in backends/router.py) degrades width only for --gpus N requests; an explicit sweep-8g-* intent is pinned 8-wide with no degradation path, so a capacity crunch on the single scarcest config (a whole 8-GPU node) blocks the job with no automatic route-around, even though 8-GPU width is a throughput choice (wide-parallel sweep), not a correctness requirement.
• Confidence (emitter): medium
• verified-at-filing: grep -rln -E 'sweep-8g|WIDE_A100|_requested_wide_widths|_gcp_ladder_specs' src/explore_persona_space/backends/*.py → router.py (WIDE_A100_80_BY_WIDTH:208, _requested_wide_widths:2470, _gcp_ladder_specs:2507, _flex_start_rung:2454) + gcp.py (INTENT_TO_MACHINE, sweep-8g machine maps); confirmed the width-degradation ladder keys on --gpus N (_requested_wide_widths reads spec.gpu_count) and explicit sweep intents bypass it (2026-07-16 UTC).

Proposed change (candidate diff sketch — refine in planning)

• When an explicit wide intent (sweep-8g-a100 / sweep-8g-h100, or any width-eligible intent resolving to a2-ultragpu-8g / a3-highgpu-8g) exhausts its wide rungs with capacity failures (queue-vanish #1116 / queue-timeout #783 / no-capacity), have _gcp_ladder_specs append degraded-width rungs (8→4→2) from WIDE_A100_80_BY_WIDTH before the terminal no_compute_available, matching the existing --gpus N degradation.
• Gate it so genuinely width-required jobs (a stated shared-nothing 8-way memory/parallelism need) can opt out; default = degrade (throughput-only wide sweeps are the common case).
• Emit an epm:progress note when width is auto-degraded so the run records that it dropped from 8→N.

Scope / surfaces

• Primary target: src/explore_persona_space/backends/router.py (+ src/explore_persona_space/backends/gcp.py for the intent→machine width maps).
• Grep the surface for sweep-8g / WIDE_A100_80_BY_WIDTH / _requested_wide_widths and update every hit; add a tests/test_router*.py case that an explicit sweep-8g intent degrades width on simulated wide-capacity exhaustion.

Constraints / invariants

• Workflow-surface only (backends/*.py is in-scope per .claude/rules/workflow-fix-on-bug.md).
• Preserve the RunPod-last-rung + no-cascade invariants (tests/test_router.py); width-degrade stays within the GCP ladder before any RunPod/terminal fallback.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 (recursion guard).

Provenance

• workflow_fix_target: src/explore_persona_space/backends/router.py
• fingerprint: (computed by filer)

target_file: src/explore_persona_space/backends/router.py, src/explore_persona_space/backends/gcp.py bug_observed: explicit 8-GPU sweep intents (sweep-8g-a100/h100) do not auto-degrade width on capacity failure, so a wide-sweep phase starves on empty 8-GPU DWS pools (#825 turn-dynamics, 2.5h+ stuck, #1116 queue-vanish on both A100 and H100 pools). why_workflow_gap: the width-aware ladder (#1121) degrades only --gpus N requests; explicit sweep intents are pinned 8-wide with no degradation path, and 8-GPU width is a throughput choice not a correctness need. proposed_change: on wide-capacity exhaustion for an explicit sweep-8g intent, append degraded-width rungs (8→4→2) from WIDE_A100_80_BY_WIDTH before the terminal no_compute_available, with an opt-out for genuinely width-required jobs. diff_sketch: |

router.py _gcp_ladder_specs(): after the wide rungs for an explicit sweep-8g

intent, if width-degradable, append 4x and 2x rungs from WIDE_A100_80_BY_WIDTH

(same as _requested_wide_widths does for --gpus N), then the narrow chain.

confidence: medium related_task: #825