MangaRelease — Track 1 Data-Trust Audit

Status (2026-07-12): resolved & shipped. The ship-blockers below were fixed in the follow-up sections at the bottom of this doc — coverage recovered to ~74–79% at launch, with Square Enix / TOKYOPOP filling via the scheduled crons. Kept as an engineering record; the “bottom line” below describes the state before the fixes.

Date: 2026-07-11 · Scope: books.csv (published dataset) as a licensed-English manga release calendar. Method: static analysis of the CSVs as committed (no scrape run) + independent web ground truth for coverage/accuracy. Bottom line: NOT launch-ready. Measured coverage of real May/June 2026 releases is ~26–32%, against a ≥95% bar. The shortfall is ~90% explained by four major publishers that are entirely or almost-entirely absent from books.csv (Kodansha, VIZ Media, TOKYOPOP, Square Enix). What is present from the other publishers is largely accurate.

0. Data at a glance

books.csv publisher mix (present): Yen Press 8,644 · Seven Seas 5,088 · Dark Horse 970 · Ize Press 507 · J-Novel 497 · One Peace 483 · Titan 199 · Inklore 166 · Denpa 129 · WEBTOON Unscrolled 128 · Udon 96 · Ablaze 68 · Square Enix 38 · Digital Manga 19. Kodansha 0 · VIZ Media 0 · TOKYOPOP 0.

1. Coverage audit — the headline number

Ground truth built from Yatta-Tachi monthly release lists (independent 14-publisher aggregator), cross-checked against VIZ’s own calendar and publisher slates. Denominator restricted to publishers MangaRelease claims to track (Kana/SuBLime/Alphapolis/Wattpad excluded as out of scope). Matched on normalized title+volume against any books.csv row (so a wrong date counts as captured, not a miss).

(These are conservative lower bounds — a handful of Yen/Seven Seas “misses” are title-variant artifacts of my normalizer, so true coverage of the present publishers is a few points higher. The 0%-publisher story below is exact and unaffected.)

Misses are not scattered — they are four whole publishers leaking

books.csv actually over-covers the healthy publishers vs. this ground truth (May: 74 Yen + 50 Seven Seas rows present, more than Yatta-Tachi lists), so the ~9 in-scope primary publishers are effectively near-complete. The entire coverage gap is the four missing publishers.

Root cause (proven from code + data)

lnrelease/parse.py:24 selects rows for books.csv with:

if i.source not in SECONDARY or i.publisher not in PRIMARY:
    lst.append(i)          # → a row is DROPPED when source ∈ SECONDARY AND publisher ∈ PRIMARY

This is a dedup rule: “if a PRIMARY publisher’s book arrives via a SECONDARY aggregator (PRH/Crunchyroll/BookWalker/stores), drop the aggregator copy because the publisher’s own primary scraper will supply it.” But Kodansha, VIZ, TOKYOPOP, and Square Enix are listed as PRIMARY (utils.py:24) while their primary scrapers produced nothing this run. Verified: 100% of their info.csv rows come from SECONDARY sources —

publisher=Kodansha    → Penguin Random House 2010, Crunchyroll 137   (source=Kodansha: 0)
publisher=VIZ Media   → Crunchyroll 215                              (source=VIZ Media: 0)
publisher=TOKYOPOP    → Penguin Random House 717, Crunchyroll 153    (source=TOKYOPOP: 0)
publisher=Square Enix → Penguin Random House 1958, Crunchyroll 128, Square Enix 38

So every Kodansha/VIZ/TOKYOPOP row is filtered out, and Square Enix keeps only its 38 primary-scraped rows.

Two distinct failure classes (exact filter-drop replay on info.csv)

Replaying the drop condition on May/June info.csv rows (distinct releases, format-deduped):

Square Enix chapter caveat (important): the 602 May “releases” are dominated by individual digital chapters — e.g. Cast-off Magic Tool Researcher #001…#011, Beast Tamer #097, Assassin & Cinderella #022 — dumped in bulk on 2026‑05‑01/15/29 (158/152/150 rows). A naïve filter fix would flood the calendar with ~600 chapter rows/month. The fix must also drop chapter-level entries (#NNN, index=0, eBook, from the SE digital store) and keep only numbered print volumes.

VIZ caveat (worst case): VIZ is a top-3 English manga publisher (One Piece, Jujutsu Kaisen, Chainsaw Man, Kagurabachi, Blue Box…). It has zero rows in books.csv and only ~215 stale Crunchyroll rows in info.csv (Chainsaw Man last seen 2025‑05, Kagurabachi never). Fixing the filter recovers ~nothing here — VIZ needs its primary scraper repaired before coverage can approach 95%.

Reverse check — junk / false entries in books.csv

books.csv is not polluted with junk — the opposite problem. Because the filter drops the PRH aggregator dupes and the SE digital-chapter flood, books.csv stays clean. Candidate false/erroneous entries are low (dozens, not hundreds) and consist of the mistags (§3) and dedup artifacts (§4), not fabricated releases. Junk weight on the confidence number is small; the number is driven almost entirely by misses.

2. Accuracy spot-check

Sample seed 20260711, stratified across sources (guaranteeing Udon/Ablaze/Dark Horse). 29 rows drawn; drift suspects + mainstream anchors verified against retail/publisher listings.

Per-field error rates (by source, from full-column scans + spot checks)

Release date — accurate for primary-scraped publishers (Yen, Seven Seas, Dark Horse, Ize, Titan, J-Novel verified/consistent). Systematic drift isolated to:

Total 515 rows carry no real date (0001‑01‑01).

ISBN — 1,095 rows missing an ISBN entirely (all present ISBNs are well-formed 13-digit): J-Novel Club 486 (98% of its catalog), One Peace 235, Seven Seas 231, Denpa 69, Udon 55, Square Enix 19. One Peace ISBNs are also occasionally wrong edition (Higehiro above).

Volume — generally reliable, with one structural bug: Seven Seas subtitles containing a comma are split into the volume field. e.g. The Cursed Sword Master’s Harem Life: By the Sword, For the Sword stores volume = "For the Sword". Worth a targeted grep of Seven Seas rows with non-numeric volumes.

3. Mistag audit

Origin/category default (tag.py + parse.py): any series not caught by a marker / imprint rule / JP_PUBLISHERS set → JP / manga. Dark Horse, Ablaze, and Udon are in no rule set, so their non-Japanese titles silently default to JP manga.

(a) Western comics tagged manga. Quantified for the mixed publishers:

• Udon — ~21 series / 35 rows are American-made comics wrongly JP/manga: all Street Fighter titles, Darkstalkers, Mega Man / Little Mega Man / Mr. Mega Man, Final Fight, Dragon’s Crown, Team Phoenix, Veil, the webcomics Ménage à 3 / Sandra on the Rocks / Sticky Dilly Buns, Giga Town, Manga Biographies: Charles M. Schulz. (Udon also publishes real manga — Rose of Versailles, Higurashi — so it can’t be blanket-tagged.)
• Ablaze — mixed: e.g. WAKFU (French/Ankama IP) is Western; The Breaker: New Waves is Korean manhwa; Gannibal is JP. Its 31 manga rows include several non-JP.
• Dark Horse — catalog is mostly genuine licensed JP manga (Akira, Berserk, Blade of the Immortal, Cardcaptor Sakura), so the Western leak is a subset (e.g. Jeff Lemire’s 10,000 Ink Stains, Star Wars “manga” art books). Lower priority than Udon.

(b) Yen Press manhwa defaulting to JP. Confirmed for the known example: 13thboy → "13th Boy" is tagged origin=JP. Yen Press (unlike its Ize Press imprint) publishes both JP manga and KR manhwa, and only the literal (manhwa) title marker catches them — so any Yen manhwa without that marker defaults JP. This is a small, curation-shaped set.

Recommended fix path (decision-on-file: auto-rules preferred, origins.csv fallback):

A blanket publisher rule is wrong for all three (each publishes real JP manga too). Realistic path = a small franchise auto-rule for the obvious Western IP + origins.csv for the tail.

4. Dedup check

True cross-key ISBN duplicates: 11 ISBNs under 2 distinct series keys (same physical book counted twice):

• Lore Olympus (Inklore) — split into a combined key loreolympus (as vols 1–7) and seven singleton keys loreolympusvolumeone…seven (each as v1), sharing the same ISBNs. Biggest single dedup failure.
• Viral Hit (WEBTOON Unscrolled) — viralhit v1 vs viralhitvolumeone v1, same ISBN.
• The Dirty Pair: Dangerous Acquaintances — base vs …ltded (Ltd. Ed.) share an ISBN.

Series split across two keys (same normalized title+volume, different key): 26 — includes a real dedup miss: afrosamurai (Titan Comics) collides with sevenseasbookarchive (Seven Seas) — a catch-all “archive” key acting as a junk bucket that should be investigated. Others are legitimate distinct editions (hellsing vs hellsingsecondedition, cirquedufreakthemanga vs …edition, early Dark Horse ohmygoddess retitled sub-volumes) — borderline, low priority.

Net: ~11 confirmed duplicate volumes to merge + the sevenseasbookarchive bucket key to audit. Small and fixable.

5. Launch-readiness verdict

Against the ≥95%-coverage working bar: FAIL. Current measured coverage 26–32%; even a perfect filter fix reaches only ~mid‑40s% until the VIZ (and to a lesser degree Square Enix) scrapers are repaired, because that data isn’t in info.csv to recover.

Confidence in the present data is high (~95%+ on date/ISBN/volume for the ~9 primary-scraped publishers). The problem is completeness, not correctness.

Top-5 punch list (ranked by impact)

1. Restore Kodansha + TOKYOPOP + Square Enix via the parse.py filter and add a digital-chapter exclusion. — Recovers ~40–60 real releases/month. Highest impact-per-effort: data already in info.csv; fix is at parse.py:24 (don’t drop a SECONDARY-sourced PRIMARY-publisher row when that publisher has no primary-source rows), plus a #NNN/index=0/eBook chapter filter so SE doesn’t flood. Ship-blocker.
2. Fix the VIZ scraper (lnrelease/source/viz.py produced 0). — VIZ is the single largest missing bucket (~30–40 releases/month, the biggest-selling titles). No shortcut; without it 95% is unreachable. Ship-blocker.
3. Fix dates for Ablaze / Udon / One Peace / Denpa. — 515 placeholder dates (Udon 100%, One Peace 86%) + Ablaze 100% month-precision (and sometimes wrong month). Either scrape real on-sale dates or mark these month-precision explicitly; don’t publish 0001‑01‑01 to a calendar.
4. Backfill missing ISBNs (1,095 rows; J-Novel 98%, One Peace, Seven Seas) and fix the Seven Seas comma-in-subtitle → volume parse bug.
5. Mistags + dedup cleanup — Udon/Ablaze Western-comic origin rules (~35+ rows) via franchise auto-rule + origins.csv; Yen manhwa overrides (13th Boy…); merge the 11 Lore Olympus/Viral Hit/Dirty Pair ISBN duplicates; audit the sevenseasbookarchive bucket key.

Reproducibility

• Ground truth: Yatta-Tachi May 2026 / June 2026 lists; cross-checked vs VIZ calendar. Retail spot-checks via Amazon / Simon & Schuster / publisher pages.
• Match logic: normalize title = strip (manga)|(light novel)|(comic)|(omnibus)|N-in-N edition|hardcover, edition words, and Vol./Volume/Part N, then [^a-z0-9]→∅, lowercase; volume = trailing integer; match against any books.csv row. Coverage counts a row captured if the title+volume exists anywhere (date errors not penalized as misses).
• Filter-drop replay: exact — replicated source ∈ SECONDARY ∧ publisher ∈ PRIMARY on info.csv, format-deduped on (key,index). No title normalization, so publisher-level drop counts are exact.
• Accuracy/date/ISBN scans: full-column CSV parse (proper quoting) over books.csv; day-01 share and 0001‑01‑01 counts are exact.
• Sample seed: 20260711 (Python random.seed), stratified across publishers.
• Tooling: standalone Python (csv module for quote-safe parsing) over the committed CSVs — no scrape run; ground truth fetched read-only with polite request pacing.

Fixes applied (this session)

Pipeline + data changes made after the audit, then books.csv regenerated (tag.py → parse.py → pages.py); dataset grew 17,032 → 20,373 rows.

1. parse.py filter fix + digital-chapter guard

• Conditional bypass: a SECONDARY-sourced row for a PRIMARY publisher is only dropped as redundant when that publisher actually produced primary-scraper rows this run (scraped_pubs). Publishers whose scraper produced nothing (Kodansha/VIZ/TOKYOPOP) keep their aggregator rows instead of vanishing.
• Chapter guard: rows whose title matches #\d+ in a digital format are skipped, so Square Enix’s #001-style singles never enter the calendar (Physical # issues, e.g. a few Dark Horse, are kept).
• Fixed publisher/tokyopop.py assuming an info['Paperback'] bucket (crashed once TOKYOPOP rows flowed through) — now info.get('Paperback', []).

Result — recovered publishers: Kodansha 0 → 2,028, TOKYOPOP 0 → 1,104, VIZ 0 → 215 (all its available, if stale, rows). Square Enix stays 38 by design — it did scrape 38 rows, so it doesn’t meet the “produced 0” bar; it belongs with the deferred VIZ scraper-rebuild.

Re-run coverage vs the same ground truth:

Kodansha coverage went 0% → 100% (May) / 78% (June). It did not reach the 95% bar: VIZ (0–3%) and Square Enix (0%) dominate the remaining gap, and their data is not in info.csv to recover — confirming both need scraper rebuilds, exactly as flagged. TOKYOPOP’s recovered rows are older; its May/June titles were never scraped (partial source), so it’s also 0–20% and needs scrape completion.

2. Dedup

• Generic ISBN-level dedup post-pass in parse.py: same ISBN under two series keys collapses to the key carrying the most volumes. Resolves the 11 cross-key ISBN duplicates (Viral Hit, Dirty Pair Ltd Ed, Lore Olympus v1–7) durably.
• Data migration (info.csv + series.csv): Lore Olympus fully merged — the 11 loreolympusvolumeN singletons folded into loreolympus with correct volumes; it now reads as one series, volumes 1–11 (HC + PB). sevenseasbookarchive bucket split into 14 proper per-title keys; Afro Samurai now a single afrosamurai series carrying both the Seven Seas and Titan editions (collision gone).

3. Mistags

• tag.py Western-franchise auto-rule (Street Fighter / Darkstalkers / Final Fight / Dragon’s Crown / WAKFU), gated to Udon/Ablaze → origin=other, category=comic. Added comic to CATEGORIES. Verified it does not catch the genuinely-Japanese Mega Man Mastermix (Ariga), Persona, or Blue Archive.
• origins.csv tail for the residual one-offs (Team Phoenix, Veil, Ménage à 3, Sandra on the Rocks, Sticky Dilly Buns, Little/Mr Mega Man) and 13th Boy → KR/manhwa. 34 rows now correctly tagged comic.

Deferred (held for their own sessions, per plan): VIZ scraper rebuild (live scraping) and Square Enix scraper (same class); the Seven Seas comma-in-subtitle → volume parse bug and the Lore-Olympus-style word-number volume parsing (both are parse-logic changes to do with tests); date/ISBN backfills. Also noted: one cosmetic non-determinism — Honkai Impact 3rd v3 has two ISBNs for the same digital slot and the pre-existing parser picks one by set order (newly visible now that TOKYOPOP is included); set PYTHONHASHSEED=0 or fix the parser’s per-volume dedup to make it stable.

Durability note: the info.csv/series.csv migrations (Lore Olympus, archive split) are corrected in place but a full re-scrape could recreate the bad keys; the durable version is scraper-side key canonicalization (belongs with the deferred parse-logic work). The parse.py filter/chapter/ISBN-dedup changes and the tag.py/origins.csv mistag rules are durable — they re-apply on every build.

Caveats: Yatta-Tachi is not exhaustive (denominator slightly under-counts Yen/Seven Seas, which books.csv covers well → reported coverage is a conservative floor). Square Enix’s own scraper is known-throttled (30–600 s delays, no skip-cache), so its info.csv presence is partial and its numbers are noted separately, not folded into the primary-publisher accuracy claims.

Follow-up — 2026-07-11 (VIZ rebuild + Square Enix cache)

Continued from the fixes above. Committed the Part-1 fixes, then rebuilt the VIZ scraper and added a Square Enix skip-cache. Dataset now 20,613 rows.

VIZ scraper — diagnosed and rebuilt. VIZ produced zero rows not because of broken code — the search and product-page selectors parse fine against live viz.com — but because it never completed: paging the whole catalogue at 30–60 s/request times out and viz.csv was never populated (CI-runner IP blocking likely compounds it). Rebuilt to seed from the monthly release calendar (viz.com/calendar/YYYY/MM) for a recent+upcoming window first, so the important data lands even if the deep search crawl is cut short; the full crawl remains as backfill. Also fixed a latent store/viz.py crash on non-manga category URLs. Ran the calendar scrape: 441 VIZ rows, May–Nov 2026. A 5-row spot-check against viz.com product pages matched date/ISBN/volume 5/5; VIZ series tag JP/manga with zero review-queue churn.

Square Enix — skip-cache added (not run live). Gave source/square_enix.py the same incremental skip-cache VIZ uses (cache series+volume pages by date; skip settled past-dated pages, re-check 20%), so repeat runs skip the expensive per-volume fetches. Logic validated; not run live to avoid hammering SE’s rate-sensitive endpoint — the full populate rides with the weekly cron. Chapter guard intact (Chapters (Digital) → None); SE books remain volumes-only.

Coverage progression (same May/June 2026 ground truth, same matcher)

VIZ itself went 3% → 100% (May) and 0% → 97% (June); Kodansha 0% → 100% / 78%.

Verdict: the ≥95% bar is not yet met (73.7% / 79.2%). The single biggest blocker (VIZ, a top-3 publisher) is resolved. What remains, in order of impact:

1. Square Enix (0%) — scraper skip-cache is in place but not yet populated; the next weekly cron should fill ~8–10 releases/month.
2. TOKYOPOP (0% / 20%) — no working primary scraper; its rows come from PRH, which hasn’t listed the May/June titles. Known residual; the cron’s PRH refresh should fill it. Not chased here.
3. Seven Seas 77% / Yen 56%–95% residual — a mix of match-normalizer artifacts (title variants) and genuinely-not-yet-scraped recent 2026 volumes; closes as the daily cron refreshes.

Once Square Enix populates and TOKYOPOP/PRH catches up via the cron, coverage should clear ~90%+; the last few points are the long tail of just-announced volumes that any calendar trails by a refresh cycle.

Still deferred (unchanged, per plan): Seven Seas comma-in-subtitle → volume parse bug; Lore-Olympus-style word-number volume parsing; Ablaze/Udon/One Peace date backfills; J-Novel ISBN backfill. CI determinism is now handled — PYTHONHASHSEED=0 pinned in python.yml.

Infra caveat (from the merge reconciliation): the CI runs without Cloudflare, so its Seven Seas / Dark Horse scrapers return 0 primary rows; the local Cloudflare-scraped data for those was preserved by union, but a future daily CI run may wipe it. The Cloudflare/CI decision is separate from this work.

Follow-up — 2026-07-11 (wipe guard, webtoon taxonomy, guardrails)

Zero-row wipe guard (scrape.py). Closed the wipe risk flagged above. On save, a source that previously carried rows but now returns fewer than a per-source fraction (default 10%) is not replaced — the prior rows are kept and a loud WIPE GUARD warning is logged. General (not an SS/DH special case), configurable per source (Square Enix uses 5% since it carries few rows), and never fires for a new source (0 prior). Tested: a 0-row Seven Seas result against the live info.csv preserves all 4,964 rows and warns; a plausible partial run passes through.

Webtoon taxonomy. Adopted: origin = creator’s country (platform-independent); category = reader vocabulary — manga/manhwa/manhua for JP/KR/CN works including webtoon-native ones in print, webtoon only for non-Asian webtoon-native print, comic for Western print, artbook separate. Removed the blanket WEBTOON Unscrolled → other/webtoon rule in tag.py (a mixed imprint must not default to one origin — this was mis-tagging every Korean title). Classified all 38 category=webtoon series by creator nationality via origins.csv: 11 → KR/manhwa (Tower of God, Noblesse, God of High School, True Beauty, Your Throne, Viral Hit, Act Like You Love Me [Xuann, Korean], Operation True Love, The Mafia Nanny, Get Schooled, Terror Man), 26 → other/webtoon (Cursed Princess Club, Boyfriends, Lore Olympus, Chasing Red, Everything is Fine, …). Merged the leftover viralhitvolumeone key into viralhit, and generalized the parse-time dedup to key on (ISBN, format) so a merged key’s same-edition rows collapse while legitimate Digital/eBook pairs (236 of them) are preserved. Added a Taxonomy legend to the README (via write.py, so it survives regeneration). Acceptance met: Solo Leveling KR/manhwa (unchanged), Tower of God KR/manhwa, Viral Hit KR/manhwa single key, Lore Olympus other/webtoon (review flag cleared).

Two left for review (creator nationality not confidently derivable — kept at their existing other/webtoon rather than guessing): Daybreak, GremoryLand.

Goodwill guardrails. Confirmed the pipeline does not rehost cover images — no binary writes, no image field in the schema or data.json; metadata and store links only. Added a lightweight robots.txt check to the fetch helper (session.py), applied to every host, not per-scraper. Effect: viz.com disallows /search, so the VIZ deep-crawl backfill is now correctly skipped (the calendar path it relies on is allowed); Seven Seas / Yen / Square Enix disallow only admin paths, so nothing else changes. The VIZ search loop degrades gracefully when a page is disallowed.

Follow-up — 2026-07-11 (VIZ historical backfill via the release calendar)

Extended the VIZ scraper to walk the release calendar backwards for a decade of backlist, with no new data source — viz.com/calendar/YYYY/MM renders complete historical months (verified live back to 2010) and is robots-allowed, unlike /search.

Two-tier ingestion (source/viz.py). The calendar month table alone yields title, product URL, format and day-precision date, so those land immediately as calendar-tier rows (ISBN blank). The ISBN lives on the product page (30–60 s/request), so those are fetched incrementally through the existing viz.csv skip-cache — oldest backfill month first, within a per-run time budget — so each weekly cron chips away at the backlog and never re-fetches a cached product. Backfill walks back from the oldest cached month, stops at a Jan-2015 floor (config, not code, to go deeper) or after 3 consecutive empty months, and runs only after the recent+upcoming window so a time-bounded run always commits current data first. Politeness unchanged (existing delays, robots.txt check, calendar pages only).

Clean upgrade path. The calendar-tier Info key (link, format) and series key (volume-stripped title) match exactly what parse() produces, so a later product fetch upgrades the row in place rather than duplicating it. ingest_calendar never clobbers a row that already carries an ISBN.

Wiring (python.yml). VIZ_BACKFILL_MONTHS/VIZ_BACKFILL_SECONDS are set weekly only (6 months back, 1500 s ISBN budget) and default to 0 on the daily run, leaving the incremental path unchanged. The budget keeps the cron well under its 350-min cap.

Validation.

• Bounded live run (2 historical months, reduced delay): 70 calendar-tier rows landed with correct dates/formats (2020-04/05, Paperback/Hardcover/Digital); a 3-product spot-check returned valid ISBNs (Daytime Shooting Star Vol. 6 → 978-1-9747-0672-3, Demon Slayer Vol. 12 → 978-1-9747-1112-3, Dr. STONE Vol. 11 → 978-1-9747-1479-7).
• Upgrade path: after ingesting calendar-tier rows and then fetching their product pages, 61 unique series keys, 0 duplicate (link, format) rows — the merge is clean, no duplicate series created.
• Offline invariants (network-free): parse correctness, dedup/clobber-safety, calendar-title series-key == series-field series-key, and BACKFILL_MONTHS=0 ⇒ no-op all pass.
• Determinism (PYTHONHASHSEED=0): backfill order is deterministic (walk order + table order; output sorted by the driver). Wipe guard untouched — backfill only adds rows, so it can never false-trigger. Purely additive, so the coverage matcher cannot regress.

No calendar gaps found above the 2015 floor (2010/2014/2016 spot-fetches all rendered), so the API-enrichment fallback stays deferred as planned.

Deferred, unchanged: Seven Seas comma-in-subtitle volume parse; Lore-Olympus word-number volume parse; Ablaze/Udon/One Peace date backfills; J-Novel ISBN backfill; Square Enix live populate.

Follow-up — 2026-07-12 (repo map + pipeline wiring)

Repo navigation (ARCHITECTURE.md). Added a hand-maintained file map and data-flow doc — every root file classified as generated output / hand-maintained input (origins.csv) / per-source cache, plus why the flat root is load-bearing (hardcoded root-relative paths + Jekyll serving from root) and where to edit static README prose so it survives regeneration. write.py’s footer links to it so the generated README.md points there on every scrape.

pages.py wired into the pipeline (lnrelease.py). Found that lnrelease.py ran scrape→tag→parse→write but never pages.main(), so data.json and the physical/digital/html/audiobook + year/*.md pages refreshed only on manual runs — the interactive mangarelease.github.io data lagged a day behind the README.md calendar (verified: data.json last touched by a human commit, not the daily bot). Added pages.main() as the final stage (after write, reading the freshly written books.csv/series.csv) and staged year/ in the commit step so a brand-new year/<n>.md at a year boundary isn’t dropped by tracked-only git add -u. Regenerated the then-stale site outputs from committed data (data.json valid: 4,574 series / 17 publishers / 15,708 rows).

Empty source caches — investigated, no change. bookwalker.csv and one_peace.csv are 0-byte but are page-visit skip-caches for live daily sources (both in the daily source list; both modules run full scrape logic and pages.save()), not dormant leftovers. Table tolerates a missing file and save() recreates it, so deleting them would only trigger regeneration — left in place.