Ecosystem review 2026-09-17: interests lapse, wikilink ratchet, seed split

.github/workflows/validate.yml

@@ -22,4 +22,14 @@ jobs:
           python-version: "3.12"
       - run: pip install pytest
       - run: make validate
+      # The wikilink leg of `make validate` is a ratchet: it fails only on an
+      # unresolved target that is not in scripts/wikilink_baseline.txt, because
+      # a link with no page yet is legal (wiki/CLAUDE.md, "Cross-references").
+      # This step re-runs it with --list so the log always carries the current
+      # wanted-page backlog — and, when the ratchet trips, shows how often the
+      # offending slug is referenced. `if: always()` so a failure above still
+      # gets the listing.
+      - name: Wikilink report
+        if: always()
+        run: python scripts/validate_wikilinks.py --list
       - run: make test

AGENTS.md

@@ -31,8 +31,7 @@ The repo has exactly **two content homes**, enforced by
   [`wiki/CLAUDE.md`](wiki/CLAUDE.md). New sub-wikis are added beside the
   existing ones, never at the repo root.
 - **`bibliography/`** — the *only* place BibTeX lives. One canonical file
-  (`pyautomemory.bib`) plus `bibkey_aliases.yaml`; never add loose `.bib`
-  files anywhere else.
+  (`pyautomemory.bib`); never add loose `.bib` files anywhere else.
 
 **Source PDFs live off-repo.** Never commit a paper (PDF/HTML, with or
 without a file extension) — read it, stub it in the right

Makefile

@@ -1,13 +1,16 @@
-.PHONY: validate validate-literature-citations validate-structure test board
+.PHONY: validate validate-literature-citations validate-structure validate-wikilinks test board
 
-validate: validate-literature-citations validate-structure
+validate: validate-literature-citations validate-structure validate-wikilinks
 
 validate-literature-citations:
 	python scripts/validate_literature_citations.py
 
 validate-structure:
 	python scripts/validate_structure.py
 
+validate-wikilinks:
+	python scripts/validate_wikilinks.py
+
 test:
 	python -m pytest tests/ -q
 

arxiv-interests.md

@@ -16,18 +16,28 @@ Format:
   into — these papers span domains, so unlike the strong-lensing inbox there is
   no single destination. A topic naming no real section falls back to
   `## Interests` (`FALLBACK_SECTION` in `scripts/interests_actions.py`).
-- **A backlog, not a timer.** Nothing here lapses. The Dashboard shows the
-  **oldest un-cleared batch only**, and its 🧹 *clear* button drops that whole
-  day and reveals the next — so a fortnight away is a fortnight of batches to
-  cycle through, one tap each, not a fortnight of lost recommendations. This is
-  the one way this list differs from `arxiv-inbox.md`, which lapses after seven
-  days.
+- **A backlog on the inbox's timer.** The Dashboard shows the **oldest
+  un-cleared batch only**, and its 🧹 *clear* button drops that whole day and
+  reveals the next — but a batch nobody clears **lapses after 7 days**
+  (`INBOX_WINDOW_DAYS` in `scripts/inbox_actions.py`, reused by
+  `interests_actions.py sweep`) and is swept by the nightly job. The batch is
+  the unit: the whole day goes, never part of one.
+
+  This list was built the other way — "a backlog, not a timer; nothing here
+  lapses" — so that a fortnight away would be a fortnight of batches to cycle
+  through rather than a fortnight of lost recommendations. The measurement
+  said otherwise: of 123 papers appended, **2 reached the reading queue and 1
+  batch was cleared** — a 1.6 % clearance rate, leaving 114 entries in 11
+  day-batches unread. A backlog nobody walks forward through is not a backlog,
+  and it makes the oldest batch — the one the Dashboard shows — the least
+  relevant thing on the board. Seven days is how long a suggestion is worth
+  looking at; the same rule as `arxiv-inbox.md`, applied a day at a time.
 - Each paper carries the same one-tap actions as the inbox: 📄 straight to the
   PDF, ➕ add to the reading queue, 📥 intake into memory, 📑 make citeable,
   ✖️ dismiss.
-- A cleared batch is not lost: git history holds it. Same reasoning as the
-  inbox sweep — the never-delete rule in `reading-queue.md` protects *reading
-  history*, and an un-acted suggestion is not history.
+- A cleared or swept batch is not lost: git history holds it. Same reasoning as
+  the inbox sweep — the never-delete rule in `reading-queue.md` protects
+  *reading history*, and an un-acted suggestion is not history.
 - One `last digest: <YYYY-MM-DD>` line, directly below the `---`, records the
   last run of the nightly digest — **papers or none** — so an empty list says
   *which* kind of empty it is: dated today it is a genuinely quiet day, four

index.md

@@ -41,12 +41,15 @@ the schema in [`wiki/CLAUDE.md`](wiki/CLAUDE.md).
   that is **not** strong lensing: black holes, dark matter, galaxy formation,
   statistics. One line per paper as `<YYYY-MM-DD> — [<Topic>] <title>[ — <ref>]`,
   where the topic names the reading-queue section ➕ files it into. It differs
-  from the inbox in one way: it is a **day-batched backlog, not a timer**. Each
-  run appends that day's ten, the dashboard shows the **oldest un-cleared batch
-  only**, and one 🧹 *clear* button drops that whole day and reveals the next —
-  so a fortnight away is a fortnight of batches to cycle through, not a
-  fortnight of lost recommendations. Same five per-paper actions as the inbox.
-  Format and transitions live in `scripts/interests_actions.py`.
+  from the inbox in one way: it is **day-batched**. Each run appends that day's
+  ten, the dashboard shows the **oldest un-cleared batch only**, and one 🧹
+  *clear* button drops that whole day and reveals the next. The seven-day
+  window is the inbox's: a batch nobody clears lapses whole (`sweep` in
+  `scripts/interests_actions.py`, on `INBOX_WINDOW_DAYS`), because at a
+  measured 1.6 % clearance rate an unlapsing backlog just makes the oldest
+  batch — the one on the board — the least relevant thing on it. Same five
+  per-paper actions as the inbox. Format and transitions live in
+  `scripts/interests_actions.py`.
 - [`reading-queue.md`](reading-queue.md) — the paper queue AND the reading
   history: `## ` section headers group papers by domain, one title per line,
   and a read paper is never deleted — its line gains a `DONE <date> — `
@@ -68,7 +71,7 @@ reviewed`), and source-page layout are uniform across sub-wikis.
 ## Citation metadata
 
 The sub-wikis explain which claims papers support. Canonical BibTeX metadata,
-key aliases, downstream-project resolution rules, and validation live in
+downstream-project resolution rules, and validation live in
 [`bibliography/`](bibliography/README.md).
 
 ## Provenance
@@ -77,6 +80,9 @@ The wikis were seeded in 2026-05 from a ~615-paper personal PDF library
 (the PDFs themselves were deleted from the repo the same day, backed up
 externally). Most legacy per-paper entries remain filename-inferred stubs:
 upgrade them to compact, claim-oriented `drafted` entries only after
-verifying the paper, and log the change in the relevant sub-wiki. The live
+verifying the paper, and log the change in the relevant sub-wiki. The eight
+pages on which *every* entry is still such a stub were split out of `sources/`
+into `wiki/<domain>/seed/` — same shape, nothing citeable, promoted a paper at
+a time as each is verified. The live
 counts — pages, maturity, unresolved citation keys, reading queue — are on
 the [dashboard](https://pyautolabs.github.io/PyAutoMemory/).

scripts/board.py

@@ -28,9 +28,11 @@
 the same overnight machinery pointed at everything that is *not* strong
 lensing — black holes, dark matter, galaxy formation, statistics — as one
 day-batch of ten. It renders the same five actions per paper, plus one 🧹
-*clear* button on the batch itself: the list is a backlog rather than a timer,
-so the board shows the OLDEST un-cleared day only and clearing it reveals the
-next. ``scripts/interests_actions.py`` owns that format and those transitions.
+*clear* button on the batch itself: the board shows the OLDEST un-cleared day
+only and clearing it reveals the next. That batch carries a days-left count of
+its own, because it lapses on the inbox's window if nobody clears it — whole,
+a day at a time. ``scripts/interests_actions.py`` owns that format and those
+transitions.
 
 Both suggestion tiers also carry a **freshness line**: the date of the last digest run,
 so an empty inbox says *which* kind of empty it is — arXiv was quiet, or the
@@ -217,6 +219,7 @@ def collect(root: Path | None = None) -> dict:
         # `interests_backlog` are how much is queued behind it.
         "interests": [],
         "interests_date": None,
+        "interests_days_left": None,
         "interests_batches": 0,
         "interests_backlog": 0,
         "interests_last_digest": None,
@@ -233,7 +236,10 @@ def collect(root: Path | None = None) -> dict:
         pages = sorted(wdir.rglob("*.md"))
         stems |= {p.stem for p in pages}
         statuses: dict[str, int] = {}
-        kinds = {"concepts": 0, "entities": 0, "sources": 0}
+        # `seed/` is the 2026-05 import's unverified half, split out of
+        # `sources/`; counted separately so the sub-wiki row does not
+        # read as if those pages were verified claim support.
+        kinds = {"concepts": 0, "entities": 0, "sources": 0, "seed": 0}
         sections = 0
         todo = 0
         resolved: set[str] = set()
@@ -288,9 +294,10 @@ def _flush():
                 papers.append(_parse_paper(line.strip()))
         _flush()
 
+    today = datetime.datetime.now(datetime.timezone.utc).date()
+
     inbox = root / inbox_actions.INBOX_FILE
     if inbox.exists():
-        today = datetime.datetime.now(datetime.timezone.utc).date()
         inbox_text = inbox.read_text(errors="replace")
         snapshot["inbox_last_digest"] = inbox_actions.last_digest(inbox_text)
         for entry in inbox_actions.papers(inbox_text):
@@ -314,12 +321,19 @@ def _flush():
         current = interests_actions.current_batch(text)
         if current:
             snapshot["interests_date"] = current[0]
+            # The batch is what lapses, so the days-left is the batch's, not
+            # each paper's — every line in it carries the same date. Kept on
+            # the papers too so an interests paper renders like an inbox one.
+            snapshot["interests_days_left"] = interests_actions.days_left(
+                current[0], today)
             for entry in current[1]:
                 snapshot["interests"].append({
                     "title": entry["title"],
                     "ref": entry["ref"],
                     "line": entry["line"],
                     "added": entry["added"],
+                    "days_left": interests_actions.days_left(entry["added"],
+                                                             today),
                     "topic": entry["topic"],
                     "section": interests_actions.section_for(entry),
                     "done": False,
@@ -763,8 +777,13 @@ def _render_md(snapshot: dict) -> str:
                    if left else " · nothing behind it")
         clear_url = _interests_clear_url(snapshot, date, len(interests))
         clear = f" · [clear this day \U0001f9f9]({clear_url})" if clear_url else ""
+        # The batch's days-left, beside the inbox's per-line one: a batch
+        # nobody clears is swept whole on the same window.
+        dleft = snapshot.get("interests_days_left")
+        left_bit = f" · {dleft}d left" if dleft is not None else ""
         lines += [f"_{date} · {len(interests)} "
-                  f"paper{'s' if len(interests) != 1 else ''}{backlog}{clear}_",
+                  f"paper{'s' if len(interests) != 1 else ''}{left_bit}"
+                  f"{backlog}{clear}_",
                   ""]
         for p_ in interests:
             label = p_["title"].replace("[", "\\[").replace("]", "\\]")
@@ -1264,11 +1283,16 @@ def _fresh_span(fresh: dict, now_text: str, stale_text: str,
         # backlog depth matters most, and dropping it from the stale variant
         # hid it twice over: on a render that is already stale, and on a fresh
         # render the moment _EXTRA_JS swaps in `data-stale-text`.
+        # The batch's days-left rides on both texts too, for the same reason
+        # `behind` does: a late digest is exactly when it matters that the day
+        # on screen is about to be swept.
+        idleft = snapshot.get("interests_days_left")
+        ileft = f" · {idleft}d left" if idleft is not None else ""
         imeta = _fresh_span(
             ifresh,
             f"{len(interest_items)} paper"
-            f"{'s' if len(interest_items) != 1 else ''}{behind}",
-            f"{len(interest_items)} paper(s){behind} · {istale_tmpl}")
+            f"{'s' if len(interest_items) != 1 else ''}{ileft}{behind}",
+            f"{len(interest_items)} paper(s){ileft}{behind} · {istale_tmpl}")
         # The batch's date IS its name — this is a backlog of days, and which
         # day you are looking at is the first thing to know.
         interests_block = (
@@ -1321,7 +1345,8 @@ def _fresh_span(fresh: dict, now_text: str, stale_text: str,
         wiki_rows.append(
             f"<tr><td class='name'>{link}</td>"
             f"<td>{w['pages']} pages · {w.get('concepts', 0)}c/"
-            f"{w.get('entities', 0)}e/{w.get('sources', 0)}s</td>"
+            f"{w.get('entities', 0)}e/{w.get('sources', 0)}s"
+            f"{('/' + str(w['seed']) + ' seed') if w.get('seed') else ''}</td>"
             f"<td>{_bar(s)} <span class='meta'>{s.get('stub', 0)} stub · "
             f"{s.get('drafted', 0)} drafted</span> "
             f"{_copy_btn(_stub_prompt(snapshot, w['name']), 'copy: upgrade a stub')} "
@@ -1392,7 +1417,8 @@ def _fresh_span(fresh: dict, now_text: str, stale_text: str,
 {inbox_block}
 <h2>arXiv interests <span class="muted">(everything that is not strong
  lensing — one day's ten at a time; \U0001f9f9 clears the day and shows the
- next{interests_backlog_note})</span></h2>
+ next; un-cleared batches lapse after {inbox_actions.INBOX_WINDOW_DAYS} days
+{interests_backlog_note})</span></h2>
 {interests_block}
 <h2>Reading queue <span class="muted">({trend_bits})</span>{spark}</h2>
 {filter_box}