fix(syw): exclude writer-unreadable openai.com cross-posts (#245)

[schmug]

Closes #245. Implements option 1 (exclude at gather) with the two refinements from the decision comment.

Change

• syw_gather.WRITER_UNREADABLE_HOSTS = frozenset({"openai.com", "www.openai.com"}), with a comment explaining why: WebFetch, the writers' only read path, gets 403 host-wide there, and urllib's occasional 200 doesn't help. The comment says to revisit if a WebFetch probe of an openai.com/index/... post ever succeeds.
• parse_openai_alignment drops any merged item (index or RSS) whose urlsplit(url).hostname is in that set. The host must match exactly; a substring never matches, so alignment.openai.com is kept. Each drop writes a dropped.jsonl row via append_dropped (stage: gather, adapter: openai-alignment, url, reason: "host not readable by writers (#245)"). append_dropped already swallows OSError, so logging can't fail the adapter. The "Lead parsers — pure" section header and the docstring note this one side effect.
• normalize_url and the seen.json keys are unchanged. build_plan is untouched and stays pure. render.py, orchestrate.py and tests/data/ are unmodified.
• Docs: spec §2.1 OpenAI row has an Amended (Keep unfetchable openai.com cross-posts out of Show Your Work's pool #245) note, and Keep unfetchable openai.com cross-posts out of Show Your Work's pool #245 is added to the status line's list of amendments. SKILL.md Layout has one line on the exclusion.

Tests

New test test_openai_alignment_drops_writer_unreadable_openai_com_cross_posts checks four things: no returned URL has host openai.com/www.openai.com, /index/an-alien-mind is gone, /metagaming is kept, and 7 dropped.jsonl rows were written (stage, adapter and #245 reason).

It was red before the fix, failing for the right reason:

>       assert not {u for u in urls if urlsplit(u).hostname in {"openai.com", "www.openai.com"}}
E       AssertionError: assert not {'https://openai.com/index/advancing-independent-research-ai-alignment', 'https://openai.com/index/an-alien-mind', ...}
1 failed, 48 deselected in 0.16s

Counts updated to exactly what the fix implies. The fixture index holds 7 openai.com cross-posts, and nothing else moved:

• test_openai_alignment_unions_rss_with_the_index: 27 → 20. The cross-post assertion is removed, and the Metagaming assertion is kept.
• test_openai_alignment_keeps_an_rss_only_post: 28 → 21
• test_gather_marks_new_items_and_records_first_observation: lead 148 → 141 (the comment's "149 parsed" becomes 142)
• test_the_gather_cli_refuses_until_seeded: GATHER ok lead=148 → 141
• test_seed_marks_everything_seen: SEED ok marked=148 → 141

Gates

$ python3 -m pytest -q
1761 passed in 37.32s        # baseline on main 73d9403: 1760; +1 new test
$ ruff check .
All checks passed!
$ ruff format --check .
55 files already formatted

Not verified here

I did not run gather or seed against the live ~/.config/show-your-work. The show is live, and this change touches no seeded state: the excluded URLs stay in seen.json/observed.json as they are. The next show-your-work-weekly run (Sat 10:00) is the first live exercise.

🤖 Generated with Claude Code