feat(arxiv): a second daily digest for non-lensing papers

[Jammy2211]

arxiv_papers.yml is strong lensing only. This adds its sibling, arxiv_interests.yml: the day's ten most relevant papers that are not strong lensing — black holes, dark matter, galaxy formation, statistics — filed into PyAutoMemory's new arxiv-interests.md, where the knowledge board renders them under the inbox with the same per-paper actions and a 🧹 button that clears a whole day and reveals the next.

Pairs with PyAutoLabs/PyAutoMemory#66, which builds that surface. Independent: this side writes papers, that side renders them.

Why a separate workflow

The lensing digest works, it is the morning's product, and its Claude step is the single point of failure for the Slack post. A broad second query, a much larger prompt and a second cross-repo push do not belong on that critical path — a bad day here must cost the interests list and nothing else. There is no Slack leg at all; this exists to fill one surface.

Two stages

The house pattern — a deterministic stage that ranks, a Claude stage that judges. Neither could do the other's job: no keyword query answers "most interesting", and no prompt reads five hundred abstracts a day.

1. .github/scripts/arxiv_interests.py pages the whole announcement band across astro-ph.CO/GA/HE/IM and scores every paper against the reading interest profile (a term in the title counts triple, in the abstract once; the best-scoring topic becomes the suggested reading-queue section). Top 60 survive, with full abstracts.
2. Claude reads those, drops what the keywords oversold, and picks the ten.

The scoring is a shortlist, never a verdict, and generous on purpose: a term that pulls in an off-topic paper costs one line of a 60-paper list Claude then ignores; a term missing costs a paper that never appears at all. Recall first, exactly as the lensing query is.

Strong lensing is flagged, not excluded

A deterministic exclusion would open a gap: a dark matter paper that mentions a lensing constraint in passing matches the lensing net, gets dropped by the lensing digest as off-topic, and falls through both lists. So each candidate carries strong_lensing and the prompt makes that one judgement, in the only place that can. The append on the Memory side then dedupes against arxiv-inbox.md, so nothing lands twice even if the judgement goes the other way.

Timing

02:30 UTC, Mon–Fri, half an hour behind the lensing digest and on the same announcement band (arxiv_fetch.announcement_band, imported not restated — two digests on different bands leave a seam a paper can fall through). Running second is load-bearing in one small way: by then the lensing digest has filed its papers into arxiv-inbox.md, which is what the dedup reads. It also keeps the two cross-repo pushes off each other's toes.

Empty days still stamp last digest: and commit, so a quiet list is provably a quiet day rather than a broken filing — the same invariant the inbox carries (PyAutoMemory#58). There is deliberately no sweep: this list is a backlog the human clears a day at a time, and ageing batches out would silently delete the recommendations they asked to cycle through.

Changes

• .github/scripts/arxiv_interests.py (new) — the ranker, with a --selftest the workflow runs before every fetch.
• .github/workflows/arxiv_interests.yml (new) — fetch → Claude → file into PyAutoMemory, with the auth-vs-race split on the push retry that arxiv_papers.yml learned the hard way.
• .github/scripts/arxiv_fetch.py — fetch() gains start for paging. The lensing query never needs it (a band is ~1.5 papers); a whole day of astro-ph always does.
• scripts/spawn.py + docs/pyautobrain/spawn_spec.md — both new files DROP from the spawned templates: org secrets, a cron, a cross-repo push by name, and one reader's personal interest vocabulary, which is instance content by definition.

Testing

pytest tests/ 245 passed, including the extended spawn-template contract. arxiv_fetch.py --selftest and arxiv_interests.py --selftest both PASS; the latter's 11 checks cover topic routing for each of the four buckets, title-vs-abstract weighting, the lensing flag, and that ranking is deterministic and caps correctly.

Not verified live: arXiv is unreachable from this session's network policy, so the ranker's scoring is tested offline only. The real band volume — and therefore whether the MAX_PAGES = 12 paging guard is generous enough — is unmeasured. The first run logs announced=/scored=/shortlisted= counts to check against, and a truncated band emits a ::warning:: rather than silently dropping the band's oldest papers.

---

Generated by Claude Code