[jehrr]

Problem

In canary run 39479200544 (2026-10-20), the question run through QUORA_PROXY passed: 262 rows, complete. The topic run in the same job then exited 5. The log shows:

18:59:36,315 Fetching page 1/1: https://www.quora.com/topic/Machine-Learning
18:59:36,379 Timeout loading https://www.quora.com/topic/Machine-Learning (attempt 1/4) — retrying in 30.0s.

The first failure came 64 ms after the navigation started. The navigation timeout is 60 s, so this was not a timeout. The three retries also failed straight away; the gaps between them were the 30/60/120 s backoff. Playwright's goto raised a PWError that is not a proxy error the engine recognises. The engine logged it as "Timeout" and dropped the message. The sidecar said page_load_timeout and nothing more, so this run's failure cannot be diagnosed after the fact. Selenium and pyppeteer printed str(e)[:120] without masking credentials.

Change (all three engines)

• A failed navigation is logged as Timeout loading only when the driver raised its timeout type. Any other error is logged as Navigation failed loading. Both lines include the error's first line, with proxy credentials masked by the engine's existing _mask_credentials.
• The final Gave up loading ... after N attempt(s): <reason> line names the reason too.
• The reason is stored on PageOutcome.load_error. The sidecar gains load_error: the reason for the first page that never loaded, or null.
• Exit codes and stop_reason are unchanged.

Evidence

• New driven smoke checks, one set per engine. A stubbed session whose navigation always raises (1) a non-timeout driver error with a credential in its text and (2) the driver's timeout type. For each case the checks require load_failed, load_error equal to the masked first line, the right label, the reason in both the retry and the give-up line, no credential in the log or the outcome, and no solve bought. A source check covers the sidecar key in each engine.
• smoke_test.py: OK in a core-only venv (driving skipped). Also OK in separate playwright, selenium and pyppeteer venvs (pip only, no browser download), with each engine's driven checks shown as PASS.
• Controls, run in the playwright venv:
  • Forcing the label back to an unconditional "Timeout" turns "Error is logged as 'Navigation failed loading'" red.
  • Removing outcome.load_error = load_error turns both "... -> load_failed with the reason on the outcome" checks red.
  • With both restored, the suite is OK.
• .github/ci_checks.py --all: ok.

What this does not establish

This PR does not say why the topic navigation failed on 2026-10-20. That error text was discarded. The next canary run will record it in the log and in canary_topic.meta.json.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKsd7q25N19WzUZjrAdzx8