[jehrr]

Problem

Five defects from the manual test of v0.2.0, each reproduced before fixing:

report #	symptom	cause
8	--gl USA → exit 3 after fresh-browser retries and proxy advice	Play answers a country it does not serve with its own HTTP 400 page (1.7 KB, no store images), which was classified as blocked
6	screenshot_count = 1 on every app row	the app page wraps screenshots one level deeper than a grid tile; the wrapper was counted
7	stop_reason: completed when the last page added nothing	a post-loop line set completed whenever the page budget was met, overwriting no_new_products (also present in the reviews loop)
12	--out x.json → x.json.json, x.json.csv	--out is a prefix and was used literally
13	--dump-html into a missing directory → navigation_failed, exit 5	the write raised mid-run, after the page had been fetched

Behaviour after

• New page state bad_request (HTTP 400, or Play's 400 heading + no store images where Selenium has no status). The run stops with exit 2 naming --gl, no retries, no proxy advice. --gl is shape-checked (two letters) before any browser starts; no country list is kept, since ZZ is well-formed and Play does serve it. The Scraper API path returns 2 for the same page.
• screenshot_count unwraps single-element wrappers: 9 for WhatsApp, 8 for Minecraft; 0 of 223 grid tiles in the captures change.
• The post-loop override is removed from both loops. Every other stop_reason assignment sits before a break, so the line could only overwrite. Status and exit code are unchanged — only the reason was wrong.
• A trailing .json / .csv / .meta.json is dropped from --out.
• --out and --dump-html create missing directories (~ expanded once, so the directory created and the file written match); a path that cannot be created is exit 2 before a browser starts.

Also: the committable-captures check now uses git check-ignore --no-index. Without it, a tracked file is never reported, so the check only held until the fixture was committed. A control went silent once the control harness began committing its disposable copies, which is how this was found.

Evidence

• Live, Playwright / Selenium / pyppeteer: --gl xx → exit 2, 0 block retries, 0 hints; --gl USA → exit 2 before launch; --category TOOLS --pages 3 --out new/deeper/tools.json --dump-html dumps/x/p → tools.json, tools.csv, tools.meta.json + 3 dumps, complete / no_new_products; --mode app --app-id com.whatsapp → screenshot_count 9.
• Fake-driver runs through engine_core.main for each behaviour. Play's 400 page is committed verbatim as a fixture and is also a second control for the asset-count signal.
• Controls: 8 of 8, each planted in a disposable copy against a green baseline and caught by its named check. PR A's and PR B's controls were re-run on the same harness with a baseline: 6/6 and 5/5.
• Offline suite green with core requirements only and in each engine venv; credential scan clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VrpkXSoXyZMGV5nqF9JV6P