Show the normalizer failure in the triage outcome recap

[tarekziade]

Why

In transformers#47631 the recap explained the longcat_flash group as:

LLM returned unparseable output (finish_reason=stop, 62 LLM turns · 60 tool calls · 82.3s · 2014266 in / 16210 out tokens)

That is the last symptom. The run was actually killed by the repo normalizer failing on check_docstrings — a failure the patch could not have caused or fixed (the runner image's baked transformers predated a model that had landed on main 51 minutes earlier). A reviewer reading the issue had no way to tell those two situations apart without opening the Serge dashboard.

The resnet row had the opposite problem: its reason was the normalizer, but the cell ended at the colon (…before a clean patch was found:) because the recap keeps only the first line — the checker name was cut off.

What

Consumes the new normalizer_error field from Serge's /status (huggingface/serge#84):

• the failing checker's name is lifted into the Reason cell — … — normalizer: 1 failed: docstrings
• the output tail goes in a collapsible under the table, bounded to 3k chars (an issue body is capped at 65536 and may carry one block per group)

Falls back to Serge's Normalizer failed (exit N) preamble when there is no checker summary (a make style crash, a timeout), and is a no-op against a Serge build that does not send the field.

Rendered against the real longcat_flash payload:

| Group | Reason | LLM | Tokens (in / out) |
| --- | --- | --- | --- |
| `longcat_flash` | LLM returned unparseable output (finish_reason=stop, …) — normalizer: 1 failed: docstrings | `moonshotai/Kimi-K2.7-Code` | 2,014,266 / 16,210 |

<details><summary>`longcat_flash` — normalizer output (tail)</summary>
… ModuleNotFoundError: Could not import module 'GraniteMoeSWAConfig' … 1 failed: docstrings
</details>

Test

tests/test_integration_failure_triage.py: the error-path shape above, the exit-line fallback, block placement after the table, tail preservation under the bound, and a fence longer than any backtick run in the captured output. 88 passed.

Pre-existing failures in test_export.py / test_resource_plugin.py are a missing opentelemetry in the local venv, identical at HEAD.

🤖 Generated with Claude Code