[jehrr]

Fixes from the 2026-10-14 audit of this repo (base 42ed314). The live earnings breakage is in a separate PR.

Problems and what happens now

Finding	Before	Now
Writes not atomic	an injected write error turned a 748-byte out.json into 13 bytes of invalid JSON	temp file in the same directory, fsync, os.replace; previous file intact, no .tmp left. New files 0o666 & ~umask, existing files keep their mode
No CSV formula guard	=HYPERLINK(...) written raw	CSV text cells starting = + - @ TAB CR LF get '; numbers (e.g. -1.5) and JSON unchanged; csv_cells_escaped in the sidecar, under the caller's extra
Captcha allowance	a failing solver was called 2 (no pool) to 4 (3-exit pool) times against SOLVES_PER_PAGE = 1	one page_flow.SolveBudget per page, spent before the solver is called, at both call sites in all three engines: 1 call in every case
pyppeteer, no proxy	AttributeError (exit 1) on a navigation timeout	load failure ends the page, as in Playwright
pyppeteer/Selenium classify without --mode	markets/movers/earnings pages labelled empty (readiness wait skipped)	content, same as Playwright
GOOGLE_FINANCE_URL from .env	--symbols MSFT:NASDAQ → "pass --symbols or --url, not both"	only a typed --url conflicts; .env.example ships the URL commented out
Text about other sites	block message told users to buy an exit in another site's country; issue templates, CONTRIBUTING, diff_runs --help (JPY), --headful help, scraper_api_client docstring, a "matches N, will serve M" log on every pyppeteer/Selenium market run	rewritten with what was measured on Google Finance, or removed
Private citations	79 references to private notes/section numbers in 13 files	removed or stated in place; a suite guard fails if one reappears (release notes v0.1.0/v0.2.0 checked: none)
CHANGELOG	several behaviour changes on main since v0.2.0 unrecorded	listed under [Unreleased], with a behaviour-change blockquote
Canary	exit 3 (blocked) and exit 6 skipped the data checks and ended green	exit 3 is an error; exit 6 warns and the row checks still run (quote status may be partial, one dropped symbol is a warning). GOOGL volume is no longer required: before the US open Google publishes 0, which the parser nulls, so a 07:00 UTC dispatch failed on it — now a notice

Evidence

• Offline suite green in a core-only venv and in one venv per engine (requirements.txt + requirements-<engine>.txt, no browser download): each run executed its own engine's group.
• New tests/stub_engines.py drives all three engines' _fetch_one_page and parse_args with stubbed drivers, run by the suite in a subprocess. On main it reports: pyppeteer CRASH AttributeError; 6× empty states; solver calls 2/2/2/4; --symbols with env URL exit 2 in all three engines. On this branch: no crash, content, 1/1/1/1, accepted.
• Controls (fault planted, cache cleared, suite red on the NAMED check, restored green): non-atomic write_json; flat 0o644 & ~umask; CSV escape disabled; extra merge order reversed; Playwright allowance spend removed; pyppeteer pool guard removed; Selenium classify without mode; typed_url reverted; a section citation added to page_flow; another site's exit string added to README. All 10 named, restored suite green.
• Canary assertion block extracted and run against the 2026-10-14 scheduled run's artefacts: the only failure is no event carries an estimate (the earnings payload change, fixed in the separate PR).

Not changed

• The earnings parser (live breakage): separate PR on fix/site-change-2026-10-14. Until it merges, the canary on this branch stays red on that one assertion.
• Shared fetch loop refactor, schema/sinks, version bump/tag/release: out of scope.
• Repo description/topics: no defect found (27-word description is informational only).
• x-debug JSON-field redaction (L2): whether the API ever echoes a key that way is not established.
• Inline sample-output step in tests.yml (L3) and the scraper_api_client --retries double meaning (L7): left as they are.

🤖 Generated with Claude Code