fix(platform): back off workflow model-capacity failures #4530

Codex's exact terminal model-capacity failure is now classified by the shared parser, but workflow agent steps still settle it as harness_error in backend/core/automations/agent_host.ts. agent_retry.ts then counts the failure, excludes the selected account and immediately re-kicks. This is existing behavior; #4529 deliberately repairs project task-agent runs only.

Independent review: #4529 (review) (N1).

Reproduce with the pinned Codex terminal turn.failed message Selected model is at capacity. Please try a different model. through the real workflow host and persisted retry path. The account is not demonstrated faulty by a model-wide capacity refusal.

Expected repair:

• Reuse the typed shared parser classification, including checkpoint restoration. Never invent HTTP429 or classify incidental assistant/tool text.
• Preserve the workflow's conversation and healthy account, while keeping existing counted retry limits, spend/time guards, effect-ledger and run protocol/lease fences.
• Durably delay the retry through the existing workflow queue/retry mechanism. Recheck authority, cancellation and the run's current claim before execution; do not add a second scheduler or bypass a legacy hold.
• Prove old-fails/new-passes for initial, resumed and checkpoint-restored failures, one delayed job, duplicate/stale-worker refusal, cancellation, and unchanged real account-failure exclusions.

Include actual PostgreSQL integration and current-head CI. No provider call, live task reset, grant expansion or model switch is needed for the regression proof. Coordinate with the pending legacy automation upgrade repair before publication.