feat: say when a step was stopped by the model provider

[bborn]

Why

Follow-up to #686. While recovering pipeline #5120, one step ran three times, committed nothing each time, and looked like a bad prompt or a broken worktree. Its activity log said only:

Waiting for user input

The actual cause existed only inside the session transcript:

API Error: Request rejected (429) · you (brunobornsztein) have reached your
session usage limit, upgrade for higher limits: https://ollama.com/upgrade

On the board, a rate-limited step is indistinguishable from one genuinely waiting on a human — and "Waiting for user input" actively misleads, since it reads as a question the agent never asked.

What

executor.LastAPIError reads the transcript Claude Code already hands the hook (transcript_path) and reports the provider failure a session ended on. It keys off the transcript entry's own fields — isApiErrorMessage, apiErrorStatus, apiErrorIsTransient — rather than sniffing message text, so it doesn't depend on error wording.

Errors the session recovered from are ignored: Claude Code retries transient failures, and reporting those would train the reader to skip these lines. When the provider itself says a failure is transient, the log line says a retry may work.

Wired into all three places a task parks:

• Notification hook — idle and permission prompts
• Stop hook — the workflow-step path and the generic end_turn path

The board now shows, as an error line:

Executor stopped on a provider error: API Error: Request rejected (429) ·
you (brunobornsztein) have reached your session usage limit, upgrade for
higher limits: https://ollama.com/upgrade (ref: d5ebadad-…)

Verification

Run against the real transcript from the incident:

status=429 transient=false
msg=API Error: Request rejected (429) · you (brunobornsztein) have reached your
    session usage limit, upgrade for higher limits: https://ollama.com/upgrade …

Unit tests cover: the failure a session ended on, a failure it recovered from (must stay silent), the transient hint, a clean session, and a missing/empty transcript path.

• go test ./... — all green
• go vet ./..., gofmt -l — clean
• golangci-lint run with the CI-pinned v2.8.0 — 0 issues

No visible surface changes beyond the log line quoted above, which is text on the existing activity feed.

🤖 Generated with Claude Code