fix(e2e-exe-dev): never score a failure notice as a ping reply

[glifocat]

What changes?

The headless installer counted any non-empty ping reply as a pass. With a bad API key, NanoClaw answers the ping with its failure notice, and the run ended in:

PING: ok
REPLY: The agent run failed. Check the logs for details.

Found during the nanocoai/nanoclaw#3901 r6 e2e on 2026-10-01. The cause there was an unfilled placeholder in the credential file.

Ping classification (e2e-install.sh, shared by e2e-exe-dev headless, e2e-proxmox and e2e-macos):

• The ping is agent_failure (exit 2) when any session gets a new messages_out row to cli/local (what chat.ts prints) marked failureNotice. The runner sets that marker on every failure notice since fix(agent-runner): never answer a failure notice with another nanoclaw#3908. The CLI socket forwards only the text, so the installer reads the session outbound.db files read-only. Rows that existed before the ping are ignored, so reused checkouts work.
• chat.ts exit 4 is also agent_failure: since fix(setup): score the agent's failure notice as a failed first chat nanoclaw#3980 it forwards the flag and exits 4 for a flagged reply.
• Text patterns can match a real reply, so they only apply when the runner may predate the marker or a DB cannot be read: the notice text, or a first line starting Error: (cores before fix: preserve provider cancellation, failure delivery, and skill files nanoclaw#3746). "May predate the marker" means the checkout lacks it or NANOCLAW_HARDENED_IMAGE=true pulls an image.
• e2e-wizard and e2e-windows were not affected. They require the exact answer to an arithmetic challenge, which a notice cannot give.

Credential check:

• exe-run.sh and proxmox-run.py refuse an empty or placeholder Anthropic credential file, or one without an sk-ant- value, before creating a VM or guest. The value is checked exactly as it is sent (only CR/LF removed) and is never printed.
• With --base or a remote OneCLI vault, the secret may already exist, so the host leaves the check to the installer. The installer checks the file only right before handing it to the gateway, so a reused vault with an unused placeholder file still passes.

Plugin 0.12.2; CHANGELOG, catalog and the e2e-exe-dev SKILL.md are updated.

The same gap in nanoclaw's own setup ping is fixed separately in nanocoai/nanoclaw#3980.

Validation

• Offline (at cd9580d; exit-4 commit 5772a59 re-ran test_e2e + contracts, 80 OK): python -m unittest discover -s tests (with skills/e2e-wizard/requirements.txt installed): 354 tests, one error. That error is test_payload_transport.test_alternate_owning_remote_transports_exact_selected_commit, which also fails on main.
• The new installer tests fail against main's e2e-install.sh: notice text, late notice, the pre-#3746 Error: reply, the marker with arbitrary text, and placeholder/foreign/embedded-space credentials. They pass with this change. Tests also cover: earlier and other-channel notices ignored, the marker trusted over quoted text, fallback when no DB is readable, hardened-image fallback (case-insensitive, .env and environment), a reused vault ignoring a placeholder file, --base skipping the host check, and no VM or guest allocation on a bad credential.
• tests/test_repository_contracts.py passes (version 0.12.2 in plugin.json, catalog and CHANGELOG).
• Codex adversarial review, 3 rounds: 9 findings, 8 fixed, 1 refuted. Round 3 asked to scope notices to the e2e agent's group; that contradicts round 1, and notices on cli/local are what chat.ts prints, so scoping is by channel.
• Live, exe.dev headless, onecli, nanoclaw main 962d527c, installer from this branch at 5772a59 (2026-10-01):
  • Placeholder credential (the real unfilled anthropic_api_key file): exit 66 before any VM was created; value not printed.
  • Good OAuth token: STATUS: pass, PING: ok, real agent reply, VM removed by --rm (removal verified).
  • Fake but well-formed API key: exit 2, ping: no_reply, not a pass. The SDK retried the 401s, so the runner's marked notice (cli/local, failureNotice: true) landed about 3 minutes after the ping, past chat.ts's 120 s stop. Running this branch's failure_notice_ids against that VM's real outbound.db returned the notice's id (exit 0). So the marker read works on a real install; the in-window path is covered by the offline tests. VM deleted after inspection.