v1 crashes at ~227ms with $0 cost — ruled out auth, action version, own CLI, model, config

100% failure rate on a private repo, ~20 runs. Claude Code initializes, then returns immediately:

{"type":"result","subtype":"success","is_error":true,"duration_ms":227,"num_turns":1,"total_cost_usd":0,"permission_denials_count":0,"modelUsage":{}}

No API call is made, nothing is billed. The action still posts its comment, then exits 1. Same signature as #776, #804, #853 and #914 (all closed or stale), so filing fresh against the current version.

Each of the following was tested independently, merged to the default branch before retesting:

Auth: workload identity federation and a plain API key. Console authentication events show Success for the exact failing run. Identical failure on both.
Action version: @v1 (resolving to v1.0.210) and pinned @v1.0.202.
Own CLI: npm install -g @anthropic-ai/claude-code@latest supplied via path_to_claude_code_executable. No change.
Model: action default and --model claude-sonnet-5.
Turn caps: 80, 40, 15.
Conversation size: a 160-comment issue and a 3-comment PR.
Task type: a one-sentence question and a create-a-file task.
Budget: credit balance and monthly spend limit both had headroom throughout.
Repo config: no .claude/, no CLAUDE.md, no .mcp.json anywhere in the repository.

Two independent credentials, two action versions, an externally installed CLI, two conversations and two task shapes all produce the same 227ms crash. That points at the bundled SDK's startup rather than any consumer-side configuration.