Keep unfetchable openai.com cross-posts out of Show Your Work's pool #245

Task

Stop Show Your Work from planning posts that neither its writers nor its validator can read. The OpenAI alignment index's cross-posts live on openai.com/index/..., which returns HTTP 403 to both WebFetch (the writer subagents) and urllib/curl (syw_write.fetch_post_text). After this change, no plan puts an unreadable URL in the feature slot or a brief slot, and a week never fails because of one.

Context

This was measured in the 2026-09-26 rehearsal for #238:

• 7 of 201 lead items were on openai.com.
• The plan made two of them briefs (/index/an-alien-mind, /index/research-acceleration-view-inside-openai). Both writers were refused on a 403, so 2 of 4 briefs were dropped.
• Both URLs returned 403 to a plain curl -A "Mozilla/5.0" as well.

The dropped briefs only shortened the episode. The worse case is a week whose top-scored feature is an openai.com cross-post: the writer is refused twice, and SKILL.md step 4 ends the run as FAILED feature refused twice even when readable candidates exist.

Reproduction

1. curl -s -o /dev/null -w '%{http_code}' -A "Mozilla/5.0" https://openai.com/index/an-alien-mind returns 403.
2. Gather and plan a week where an openai.com cross-post ranks into a slot.
3. Expected: that item is not planned, or is planned only if it can be read.
4. Actual: it is planned, then the writer refuses it (brief dropped) or the run fails (feature).

Pointers

• skills/show-your-work/syw_gather.py:423: parse_openai_alignment. This is where the ↗ cross-posts from the index are unioned in, and a candidate place to tag or skip them.
• skills/show-your-work/syw_script_plan.py:~170: the build_plan pool filter, the other candidate place. It already excludes undated items (Harden Show Your Work's unattended edges (8 parked review items) #236).
• skills/show-your-work/syw_write.py:393: fetch_post_text, the validator's own fetch.
• skills/show-your-work/prompts/write_feature.md:16 and write_brief.md:14: the writers read the post with WebFetch only.
• ~/.claude/projects/-Users-cory-clodcast/memory/project_clodcast_webfetch_blocked_feeds.md: the daily show has the same class of blocked hosts.

Decision (Cory)

1. Exclude openai.com cross-posts at gather. This is simplest. The cost is that an OpenAI post published only there never airs.
2. Probe fetchability at plan time. Try one fetch per slotted candidate and skip anything that fails. This is general across hosts, but it puts a network step in a layer that is pure today.
3. Keep them but never feature them. They stay brief-eligible and are dropped when refused.

Recommendation: option 1, recorded as a host list next to the adapter. It keeps build_plan pure and fixes the feature-failure case. Revisit if openai.com stops blocking.

Constraints

• build_plan stays pure (no network), unless option 2 is chosen deliberately.
• URL identity (normalize_url, seen.json keys) must not change; the host is already seeded.
• render.py, orchestrate.py and tests/data/ stay unmodified. Tests never touch ~/.config/show-your-work.
• Write the failing test first.

Acceptance criteria

• A test shows an openai.com/index/... lead item never reaches the feature slot or a brief (or, under option 2, that an unreadable one is skipped).
• alignment.openai.com items, including the misalignment casebook, are unaffected.
• SKILL.md and the spec §2.1 source table state the exclusion.
• The full suite, ruff check . and ruff format --check . are green, with counts in the PR.

Out of scope

• Any other host. File separately if another one is measured failing.
• The daily show's blocked feeds.