tracker

All repositories: gitoria

Readme Code Pull requests Releases Tickets Settings

Commit2667da05

2667da05tracker#30 (mission 025): /my/ pages from slim cached title cards, episode rows and watch sets (after the jobs /my/series 1.8 s -> 0.06 s, /my/unwatched 4.1 -> 0.18 s); timeline page shows its name once; franchise widget under the poster/title; movies with TV leftovers (First Contact) go through the details repair; tools/count-tmdb-ids.hl; gates 327/0, 52/0, 32/0mre2026-10-02 16:07

2667da05/STATUS.md

145.9 KB

# STATUS — tracker.worldapi.org
## 2026-10-02 — mission 025: tracker#30 /my/ pages fast after the jobs + timeline name once + widget position + First Contact (NOT deployed)
**Cause of #30**: every /my/series load fetched each followed title's whole record (cast + crew) and one episode per watch (11,648
fetches); /my/unwatched + /my/schedule every season + episode of 128 titles. With a big heap after the jobs each allocated byte
feeds the byte-triggered GC. **Fix** (shows.hl, watches.hl, README "mission 025"): slim caches `cardOf` / `episodeRowsOf` /
`episodeKeyOf` + `watchedSetOf`, warmed at boot (~460 ms), refreshed in `catalog.hl refreshCatalogShow`; also used by the tiles.
/my/series keeps going by the watched EPISODE's own `show` (Reacher S03 watches sit on a migrated duplicate season outside
the title's seasons — the first cut via the seasons gave S02E08 instead of S03E08; caught by the before/after text diff).
**Real copy** (live storage 14:50 — taken while live ran its seed, so the copy's first start = forced delta (316–333 titles) +
the rest of the seed (1,802 movies), ~20 min; signed in as az5b2; 20 loads/page, `server / browser ms avg`):
| page | before (3909810), after jobs | after (this), after jobs | after, fresh |
|---|---|---|---|
| /my/series (126) | 1767 / 1927 | **63 / 123** | 53 / 113 |
| /my/shows (3) | 3539 / 3397 | 39 / 73 | 64 / 73 |
| /my/unwatched (1,288) | 4063 / 4540 | **176 / 329** | 172 / 336 |
| /my/movies | 114 / 155 | 8 / 38 | 6 / 36 |
| /my/schedule (56) | 4518 / 4543 | 14 / 52 | 13 / 58 |
| /series | 50 / 67 | 18 / 62 | 16 / 55 |
| Picard + Show all | 164 / 207 | 159 / 228 | 147 / 221 |
| RSS over the run | 4.7 → 21.3 GB (peak), 10.6 end | 1.36 → 1.68 GB | 1.83 → 2.13 GB |
- Before was benched while the after-process ran its jobs; after was benched with 3 GB of it swapped out (the before-process had
pushed Loreana into swap). Fresh before vs after (5 node fetches): /my/series 166–277 → 21–41 ms, /my/unwatched 500–730 → 87–125.
- **Correctness**: page text of /my/series, /my/shows, /my/unwatched, /my/movies, /my/schedule, /series, / identical between
before and after (fresh), and between the after-process that ran the jobs (caches refreshed by 333 synced titles, 6,361 new
episodes) and a fresh process on the same storage.
- **Timeline page** (components/franchise.hl): heads carry `hasName`; a timeline's own page shows no `a.timeline-name`.
- **Widget** (show.hl/styles.hl): `FranchiseWidget` inside `showMeta` before the genres; hidden when empty (`:has`).
- **First Contact**: tmdbId 199 is correct (TMDB movie 199); tmdbSummary was TV 199's ("Kabuhayang Swak na Swak…"), status
`Ended`, seasonsCount/episodesCount 0 — from the #9 `/tv/` bug (report 006). `details.hl hasTvLeftovers` puts such a movie into
the details repair; on the copy the next start logged `details repair: start, 1 incomplete titles` → `done: 1/1 complete=1`,
record afterwards: tmdbSummary = the movie's text, status Released, counts null. Only 1 such movie on real data.
- **Shared TMDB ids** (`tools/count-tmdb-ids.hl`): 0 records share a TMDB id with a different title. 68 type+id pairs are held by
2–6 records of the SAME title (mostly adult filmography titles, also Archive 81 ×6, Ally McBeal ×2, The Flight Attendant ×2;
several even share one slug) — duplicate imports, not fixed here.
- Gates: browser 327/0 (+2: TV-leftover repair, repaired once), franchises 52/0 (+2: widget position 390/1280, timeline name
once), kinds 32/0. Shots looked at: gate `/tmp/w074/{shots,fr,kinds}`, real `/tmp/w074/real/*-{390,1280}.png` (First Contact
full page, A Kid in… timeline, /my/*) on Loreana.
**Open**: duplicate title records (above) — which import path makes them (search import / filmography racing?) is a ticket of
its own; the page caches assume one process per storage (as watches.hl already does).
How to repeat (Loreana, this folder; ports 8728–8730):
```
W=$PWD/.scratch/w074; mkdir -p $W/pristine $W/real-after
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C $W/pristine
cp -a $W/pristine/storage $W/real-after/
tests/realdata-074.sh jobs $PWD $W/real-after 8730 # wait for "collections done" in real-after/server-jobs.log
R=$W/real-after; REAL_PORT=8730 REAL_CHROME=8729 REAL_SID_FILE=$R/sid REAL_PID=$(cat $R/server.pid) REAL_OUT=$R/bench-jobs.json node tests/realdata-074-bench.mjs
tests/realdata-074.sh stop x $R 8730 # "before": a tree snapshot (rsync without .scratch/storage/.env, absolute .env link) as <tree>
```
## 2026-10-02 — mission 024 (old 071): Hybriel master 190aa11d measured, NOT adopted (vendor stays 8efba065, NOT deployed)
**Candidate**: master `190aa11d` (#127 both shapes d98926c6 + 04df4428, GC correctness fc838894, closure scopes 038d84b3),
built ReleaseFast read-only (`.scratch/hybriel-071`), bin sha256 `feb597c4…dedc8`; kept ready in `.scratch/w071/vendor-190aa11d/`
(bin/ + plugins/; only `plugins/web/{WebFramework,compile}.hl` differ in git). Lambda audit: no semantics change since 8efba065,
`tools/lambda-audit.py .` lists the same 13. Gates on it: browser 325/0, kinds 32/0, franchises 50/0.
**Verdict: rule "every big page ≤ 1.3× the old live binary AND RSS flat over 20 loads" fails** → bin/plugins restored from
`.scratch/pre-071` (= 8efba065, 8081350's vendor); gates on the restored tree 325/0, 32/0, 50/0.
**Gate fix (committed)**: `tests/kinds.mjs` sets `TRACKER_COLLECTIONS=0` — the collection seed (on by default since #19) made a
TMDB request now and then before check 1 ("kinds off … no TMDB request"), 1 of 4 runs failed.
**Real copy** (live `storage/` 2026-10-02 10:50, real TMDB, all first-start jobs + a forced delta run; signed in as az5b2; 20
loads per page, each = node fetch + Chrome full load; `browser / server ms, RSS before→after (max)`; Loreana shared with other
workers, ±30 % noise between runs):
| page | old 8590df63 fresh | new 190aa11d fresh | old after jobs | new after jobs |
|---|---|---|---|---|
| /my/unwatched (1,288 rows) | 1089 / 974, 7.8→10.3 GB (15.3) | 910 / 793, 1.7→2.1 (2.2) | 3004 / 2688, 13.0→25.6 (26.4) | 2967 / 2879, 9.4→8.2 (12.2) |
| Picard + Show all (212) | 240 / 109 | 245 / 225 (repeat 222 / 135) | 270 / 124 | 333 / 322 |
| /my/series (126) | 292 / 194, 6.5→8.2 | 529 / 398 (repeat 380 / 277), 1.9→2.1 | 1827 / 1834, 17.9→30.1 | **4410 / 4401**, 7.4→10.8 (11.6) |
| /series | 49 / 22 | 82 / 49 (repeat 89 / 30) | 102 / 64 | 166 / 160 |
| /franchises | 52 / 20 | 70 / 31 (repeat 75 / 20) | 79 / 262 | 126 / 23 |
| a timeline | 47 / 44 | 55 / 9 | 50 / 10 | 85 / 18 |
- Boot RSS (fresh, 60 s after listening): old 7.96 GB, new 1.69–1.74 GB, 8efba065 2.39 GB, new+`HL_GC_BYTES=0` 3.3 GB.
- First-start jobs (delta 288 titles, kinds 1,789, seed 4,774 → 110 timelines): new 57 min (sync 448 s, kinds 1,188 s, seed
2,151 s), RSS at 60 s 4.7 GB, **peak 9.8 GB**, median 6.4; old 53 min (430/1,082/2,020 s), 7.7 GB, peak 13.6, median 10.3.
- Fails: after the jobs new `/my/series` 2.4× (4.4 s), `/series` 1.6×; fresh `/my/series` 1.3–1.8×, `/series` ~1.7× (tens of ms).
Not flat: after the jobs new swings 7.4–12.2 GB (old 13–31 GB!). Fresh new is flat (1.7–2.2 GB) and `/my/unwatched` faster.
- vs current main 8efba065 (fresh, same copy): /my/unwatched 4098 / 3933, /my/series 407 / 347, Picard 303 / 177 — 190aa11d is
better than what main has now (/my/unwatched 4.5× faster; 070's after-jobs 8.3 s and +25 MB/load are gone: 2.9 s, no climb).
- New + `HL_GC_BYTES=0`: fresh all ≤ 1.2× old (unwatched 577, /my/series 299, /series 59) at 3.3→3.9 GB; after its jobs (peak
16.1 GB, 54 min) speed ≈ old (/my/series 1999, /series 135 = 1.3×) but RSS 15.8→25.9 GB — the old binary's memory.
- So: GC by bytes with a big live heap still costs (the /my/series and /series renders); hybriel #127 fixed the growth, not the
per-collection cost. Decision for the creator/architect: adopt 190aa11d anyway (better than main in every after-jobs number but
/my/series), or wait for a GC fix. Shots: gate shots `/tmp/w071/{shots,kinds,fr}` on Loreana.
How to repeat (Loreana, this folder; the old tree = this tree with `.scratch/pre-070` bin/plugins + an absolute `.env` link):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/w071/pristine
for w in new old; do mkdir -p .scratch/w071/real-$w && cp -a .scratch/w071/pristine/storage .scratch/w071/real-$w/; done
tests/realdata-071.sh jobs new 8758 # all jobs + delta run; wait for "collections done" in real-new/server-jobs.log (~57 min)
R=$PWD/.scratch/w071/real-new; REAL_PORT=8758 REAL_CHROME=8759 REAL_SID_FILE=$R/sid REAL_PID=$(cat $R/server.pid) REAL_OUT=$R/bench-jobs.json node tests/realdata-071-bench.mjs
tests/realdata-071.sh stop new 8758; tests/realdata-071.sh fresh new 8758 # prints boot RSS; then the bench again (bench-fresh.json)
# same with "old"; HL_GC_BYTES=… tests/realdata-071.sh … passes through; tools/realdata-071-table.py label=.scratch/w071/real-new/bench-jobs.json … → the table
```
## 2026-10-02 — mission 023 (old 070): t18 + t20 + t19 merged into main, migrated data kept, Hybriel 8efba065 (NOT deployed)
**Commits on main** (not pushed): `20e09d8` merge t18 (#18), `daf49fe` merge t20 (#20/#21), `749019b` merge t19 (#19), `46b2138`
migrated data kept, `1ad19c8` re-vendor + `&` params, then the docs commit. Worktrees tracker-wt18/-wt19/-wt20 + branches removed.
**Merges** (conflicts README/STATUS/show.hl/project.hl/search.hl/components/search.hl/browser.mjs/faketmdb.mjs; both sides kept):
- show.hl: `pageShowOf.summary = summaryOf(s)` (#18 on 062's small object), `slugInfoOf` (#20) beside 062's slug helpers, the
franchise widget (#19) from `showRow.id`.
- project.hl clock: adult backfill → details repair → **kind backfill** → credits job; the collection seed (own clock) now waits
for ALL of them and the daily run (one TMDB pacing at a time) — first start = kinds, then the seed.
- details.hl `withDetailsFields` also stores `tmdbType` + `kind` (delta light step, #26 completion).
- #19 pages: typed headings "Franchises", "Franchise | …", "Timeline | …" (also the timeline editor); title links via `titlePath`.
- search import: `path` (#21) + `details` (#28), log line with the real path.
- Gates: browser.mjs URLs → `/series/` `/movies/`, typed-h1 selectors; kinds.mjs/franchises.mjs: repair/credits/kinds off,
fixture name = seed name, default ports 8700–8710 (deploy.sh runs all three now).
**Migrated data is never destroyed** (conductor): the one-time step MOVES a copied summary to `migratedSummary` (marker
`summaries-moved.txt`); a migrated record's (oldId not `tmdb-…`) title/genres/homepage/tagline → `migratedTitle`/`migratedGenres`/
`migratedHomepage`/`migratedTagline` before TMDB's first replaces them; seasons/episodes (oldId not `tmdb-`/`tvmaze-`)
`migratedTitle`/`migratedSummary`. Set once. Real copy after the jobs (`tools/count-migrated.hl`): summary 9,647, title 16
(e.g. "The Daily Show with Trevor Noah → The Daily Show", "Kaun Banega Crorepati(कौन बनेगा करोड़पति) → Kaun Banega Crorepati"),
genres 71, tagline 16, homepage 4, episodes title 635 / summary 409, seasons 3 / 3. Gate: 3 checks (move at first start, renamed
show/movie/episode keep the old value, a second rename keeps the first).
**Hybriel 8efba065** (README "Vendored Hybriel"). Lambda audit (#48, `tools/lambda-audit.py`): 13 lambdas change a parameter —
details.hl `withCredits` 236, `withDetailsFields` 265, `addSum` 359, `firstStep` 412 (stores the record itself); franchises.hl
`storeOrder` 373; people.hl `startFill` 222, `fillStep` 269 (p, jobIn); tmdbsync.hl `mergeTvmaze` 279 (seasonsIn, episodesIn),
`syncMovie` 442 (show, res), `addTotals` 826 — every caller uses the RETURNED value or the stored record (the maps of
`mergeTvmaze` are not read after the call) → nothing to fix. **But the copy per call was slow**: 319 read-only parameters got `&`
(`tools/ref-params.py --apply`): /my/schedule 2.35 → 0.46 s, /my/unwatched 13.7 → 5.7 s (real copy, fresh process). Micro test:
a 20k-row list passed 2,000× = 15.5 s copied, 0.015 s with `&`. /my/unwatched: one merge sort instead of n² inserts (5.7 → 4.8 s).
**Gates** (final tree, new binary):
```
TRACKER_GATE_SHOTS=/tmp/w070/new-shots TRACKER_GATE_PORT=8751 TRACKER_GATE_IDENT_PORT=8752 TRACKER_GATE_CHROME=8753-8756 TRACKER_GATE_TMDB_PORT=8757 node tests/browser.mjs
325 passed, 0 failed
TRACKER_GATE_SHOTS=/tmp/w070/new-kinds TRACKER_GATE_PORT=8750 TRACKER_GATE_IDENT_PORT=8751 TRACKER_GATE_CHROME=8752-8756 TRACKER_GATE_TMDB_PORT=8757 node tests/kinds.mjs
32 passed, 0 failed
TRACKER_GATE_SHOTS=/tmp/w070/new-fr TRACKER_GATE_PORT=8750 TRACKER_GATE_IDENT_PORT=8751 TRACKER_GATE_CHROME=8752-8756 TRACKER_GATE_TMDB_PORT=8757 node tests/franchises.mjs
50 passed, 0 failed
```
**Real data** (copy of Byrodin `storage/` 2026-10-02 ~04:40, real TMDB/TVmaze, all jobs on, the daily run forced as a delta
from yesterday). Jobs: delta run 261 titles (14 followed), 447 requests, 377–612 s; kind backfill 1,789 titles in 970–1,246 s;
credits/repair/adult nothing to do; seed 4,774 movies, 110 timelines / 272 titles, 1,920–2,228 s. RSS (sampled every 10 s):
| | after boot | jobs peak | 200 page loads after the jobs (fresh process) | /my/unwatched |
|---|---|---|---|---|
| old 8590df63 | 8.0 GB | 15.2 GB (in swap pressure, see below) | 8.0 → 12.2 GB | 1.15 s |
| new 8efba065, final tree | 2.1 GB | 11.8 GB, median 2.7 GB | 2.1 → 2.8 GB | 4.8 s |
| new, `HL_GC_BYTES=268435456` | 2.7 GB | 12.4 GB at 24 min (run aborted) | 2.7 → 3.5 GB | 2.2 s |
| new, `HL_GC_BYTES=0` | 6.2 GB | — | 6.2 → 8.7 GB | 1.15 s |
- The SAME new process right after its jobs: 200 loads 3.0 → 8.1 GB, then 8.1 → 13.6, 13.6 → 18.9 GB; /my/unwatched 8.3 s.
After a restart + one delta run: 4.3 → 5.8 → 7.5 GB per 200 loads. → restart after the first-start jobs (README "Deploy").
- The first parallel run (old + new at once) was killed by **systemd-oomd** at 05:38 (Loreana's memory + swap > 90 %, other
workers' tests too); its curves are in `.scratch/w070/real-{new,old}/rss.log`, the controlled numbers above ran one at a time.
- Pages (curl, 200 loads, final tree, fresh process): `/` 0.02 s, lists 0.01–0.05, `/franchises` 0.04, a series 0.25–0.45,
`/my/series` 0.4, `/my/shows` 0.33, `/my/schedule` 0.42, `/my/unwatched` 4.8, a person 0.4.
- Browser (`tests/realdata-070.mjs`, signed in as az5b2, 390/1280, after a delta run): lists 39–146 ms, `/franchises` 104–129,
a timeline ~40, Star Trek: First Contact with its widget (‹ Generations | TNG Collection | Insurrection ›) 136–346, Picard
(202 cast) 192–368, SNL 444–459, /my/series 825–1,064, /my/shows ~770, /my/unwatched 5.6–6.1 s, /my/schedule ~780, search
77–476; the first `/` after boot 15 s (first hydration), then 121 ms. No sideways scroll, 0 console problems. Shots looked at:
Loreana `/tmp/w070/real/*-{390,1280}.png`. No series is in a timeline on real data (the seed makes movie timelines only), so
"a series page with widget" was checked in the franchise gate (Gate Test Show), cast on Picard.
How to repeat (Loreana, from this folder; ONE tar of live storage):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/w070/pristine
mkdir -p .scratch/w070/real-final/sessions && cp -a .scratch/w070/pristine/storage .scratch/w070/real-final/
tests/realdata-070-run.sh final 8750 # all jobs, the delta run now, session for az5b2 → real-final/sid, RSS → real-final/rss.log
grep 'done' .scratch/w070/real-final/server.log # wait for "collections done" (~50 min)
tests/realdata-070-load200.sh final 8750 # 200 signed-in loads: per-path avg/max + RSS before/after
REAL_PORT=8750 REAL_CHROME=8752 REAL_OUT=/tmp/w070/real REAL_SID_FILE=.scratch/w070/real-final/sid node tests/realdata-070.mjs
tests/realdata-070-bench.sh new 8750 [HL_GC_BYTES=…] # a FRESH process on a done copy: boot RSS, 200 loads, stop
kill $(cat .scratch/w070/real-final/server.pid); rm -rf .scratch/w070/pristine .scratch/w070/real-*/storage
```
(`.scratch/w070/oldtree` = this tree with `.scratch/pre-070` bin/plugins and a `.env` link — removed again.)
**Open**: (1) deploy + restart once after `collections done`; a daily restart after the 04:00 run? (2) Hybriel (architect):
GC by bytes — with a 2+ GB live heap every 64 MiB of plugin data costs a full collection (/my/unwatched 1.2 → 4.8 s); a process
that fetched for a while keeps ~8–25 MB per page load (fresh: 3.5 MB) — repro: the bench above; the #48 copy of a big list per
call (micro test in this entry). (3) `HL_GC_BYTES` stays default: 256 MiB gives /my/unwatched 2.2 s, its full jobs peak was not
measured (aborted at 12.4 GB). (4) Video → Series? (#21, creator). (5) #18's 24 kept "own" summaries.
## 2026-10-01 — ticket #19 (mission 020 (old 066)): franchises and timelines (branch t19 from fa1f9df, NOT merged, NOT deployed)
**Built** (details [docs/franchises.md](docs/franchises.md)): tables `franchises.db`, `timelines.db`, `collectionchecks.db`
(`franchises.hl`); `/franchises`, `/franchises/<slug>` (each timeline: Timeline | Release toggle, a series by its LAST released
episode), `/timelines/<slug>`; the widget `‹ Prequel | Timeline | Sequel ›` with the franchise above on every show/movie page
(one per timeline); the creator's editor `/franchises/edit` + `/timelines/<slug>/edit` (`TRACKER_EDITORS`, default az5b2,
checked in every face); the TMDB collection seed in the app (own clock, one movie per step, 260 ms per request, waits for the
daily sync and the adult backfill, resumes).
**Decisions**: a collection becomes a timeline only with ≥ 2 of our non-adult movies (a 1-title widget says nothing); seeded
timelines get no franchise (franchises are hand-made); pages/widget skip non-public titles (adult != false), the editor shows
all; sort toggle client-side; styles in styles.hl (own section), the widget a composed component so show.hl changes by 4 lines.
**Merge notes** (existing files touched): `components/show.hl` (1 import, 1 static, 1 member, 1 View line before
`showSeasons`), `project.hl` (4 component imports, 1 franchises import, 5 routes after `/genre/:genre`, `franchisesReady` at
boot, the seed clock block before `on Error`), `styles.hl` (a section before "/ (tracker.worldapi.org#13)" + 2 lines in the
media block), `tests/browser.mjs` (`TRACKER_COLLECTIONS: '0'` in startServer), `tests/faketmdb.mjs` (`collections`,
`belongs_to_collection`, `at` on requests), `README.md`, `STATUS.md`. New: franchises.hl, components/{franchises,franchise,
franchisewidget,franchiseedit,timelineedit}.hl, tests/{franchises.mjs,seed-franchises.hl,realdata-066.mjs}, docs/franchises.md.
**Gates** (both must be green):
```
TRACKER_GATE_SHOTS=/tmp/w066/shots TRACKER_GATE_PORT=8732 TRACKER_GATE_IDENT_PORT=8733 TRACKER_GATE_CHROME=8734-8736 TRACKER_GATE_TMDB_PORT=8737 node tests/franchises.mjs
48 passed, 0 failed
TRACKER_GATE_SHOTS=/tmp/w066/main-shots TRACKER_GATE_PORT=8732 TRACKER_GATE_IDENT_PORT=8733 TRACKER_GATE_CHROME=8734-8736 TRACKER_GATE_TMDB_PORT=8737 node tests/browser.mjs
266 passed, 0 failed
```
Shots looked at: Loreana /tmp/w066/shots/{franchises,franchise,timeline,widget-*,franchise-edit,timeline-edit*}-{390,1280}.png.
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~22:47, :8738, real TMDB): seed **4,774 movies** checked in **2,459 s**
(~41 min), 4,957 TMDB requests, **110 timelines / 272 titles**, 235 collections with < 2 of our movies, 4,143 movies without a
collection, 124 TMDB 404, 0 retries, slowest step 860 ms. Pages during the seed (curl ×40): `/` avg 0.08 s (one 2.2 s outlier),
`/movies` 0.01, `/shows/Foundation` 0.18, `/franchises` 0.11. No Star Wars / Alien MOVIES in our data (only series) → checked
Star Trek: The Next Generation (4), Harry Potter (7), Lord of the Rings (3): order right, widget right on the middle title
(e.g. First Contact: ‹ Generations | TNG Collection | Insurrection ›). Full loads: `/franchises` 74–109 ms (83 KB, 110
timelines), timeline pages 49–213 ms, movie pages with widget 403–464 ms (curl: 0.38 s vs 0.34 s for a movie without a
timeline). Creator (az5b2): Edit link + editor forms present. 0 console problems. Shots /tmp/w066/real/*.png (looked at).
How to repeat:
```
tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage | ssh loreana 'tar xzf - -C <wt>/.scratch/realdata' # from Byrodin
ln -s ../tracker.worldapi.org/.env .env # the real TMDB token for the worktree (never read it); rm afterwards
TRACKER_PORT=8738 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_BACKFILL=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8738 TRACKER_OLD_SHORT_IDS=$PWD/data/old-short-ids.json HL_HOST=127.0.0.1 nohup ./bin/hybriel project.hl > /tmp/w066/real/server.log 2>&1 &
grep '^collections' /tmp/w066/real/server.log # wait for "collections done"
# session file for 0mujqspn0aqb (name sha256(id)) in .scratch/realdata-sessions/, id → .scratch/real-sid
REAL_MATCH='Star Trek|Harry Potter|Lord of the Rings' REAL_PORT=8738 REAL_CHROME=8739 REAL_OUT=/tmp/w066/real node tests/realdata-066.mjs
kill <pid>; rm -f .env; rm -rf .scratch/realdata .scratch/realdata-sessions .scratch/real-sid
```
**Open**: (1) deploy: the first start runs the seed ~40 min (TMDB pace), after the adult backfill. (2) Not built: rename /
delete franchises and timelines, edit a franchise summary, franchise `shortId` (field null). (3) New parts of a collection that
already has a timeline are not added by a later seed; titles imported later are not seeded unless their check row is
removed. (4) The widget sits under the whole header — on a phone after the cast (the creator may want it higher). (5) A
timeline page shows its name twice (page title + block head). (6) A show page costs ~40 ms more with a widget.
## 2026-10-02 — tickets #20 + #21 (mission 022 (old 068)): typed headings, Series / Shows split (branch t20, NOT merged, NOT deployed)
**Built** (details: [docs/kinds.md](docs/kinds.md)):
- #20: headings "<Type> | <name>" on the title page (Movie / Series / Show), genre, person. The type word is in its colour, the
bar muted, the name white. /movies, /series, /shows, /my/series, /my/shows show their type word; so do the homepage row
headings, the search labels and the filmography labels.
- Colours: Movie yellow, Series greenDark (the accent), Show purple, Genre blueDark, Person orange. Franchise blueLight and
Timeline greenLight are classes only (for t19).
- #21: TV titles get `kind` = series | show, rule in `shows.hl kindOfTitle`:
- show = TMDB type Reality / Talk Show / News / Video, or one of our genres Reality / Talk / News;
- everything else = series.
- Who stores `kind` + `tmdbType`: the daily sync, the adult backfill, the import, and a new kind backfill (`TRACKER_KINDS`,
after the adult backfill, 1 request per title without a show genre, resumes).
- Pages: /series and /shows (+ /page/N); /series/<slug>, /shows/<slug>, /movies/<slug>; /my/series + /my/shows; home has
5 tiles and 3 rows.
- Old /shows/<slug> of a series → 301 /series/<slug>; of a movie → 301 /movies/<slug>.
**Decisions**:
- **Movies moved to /movies/<slug>**: /shows/ now means TV shows; old links answer 301.
- **/shows/<slug> is a FUNCTION route** (`project.hl showsRoute`), because hl:web cannot redirect from a component route. For a
show it renders the Show component itself through hl:web internals (`server.router.match`, `server.page`, `hostOf`,
`requestHeaders`, `sessions`). Re-check when hl:web is re-vendored; the kinds gate covers it.
- Consequence: a link to a show is a full page load. Series and movie pages stay client-side navigations.
- Search "Fetch from web" rows say **TV** (TMDB's search has no TV type).
- `/my/series` signed out: "Sign in to see the series you follow."
**Gates**:
```
mkdir -p /tmp/w068/shots /tmp/w068/main-shots
TRACKER_GATE_SHOTS=/tmp/w068/shots TRACKER_GATE_PORT=8761 TRACKER_GATE_IDENT_PORT=8762 TRACKER_GATE_CHROME=8763-8765 TRACKER_GATE_TMDB_PORT=8766 node tests/kinds.mjs
32 passed, 0 failed
TRACKER_GATE_SHOTS=/tmp/w068/main-shots TRACKER_GATE_PORT=8761 TRACKER_GATE_IDENT_PORT=8762 TRACKER_GATE_CHROME=8763-8765 TRACKER_GATE_TMDB_PORT=8766 node tests/browser.mjs
266 passed, 0 failed (URLs, h1 selectors and the home checks updated; TRACKER_KINDS=0 in its startServer)
```
One run of browser.mjs failed once at the short-id resume check (`count-short-ids.hl`: mpackdb LockTimeout right after
stopServer, under load). The rerun passed; nothing of #21 is involved.
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~23:35, :8760, real TMDB through a `.env` symlink that was removed afterwards):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' | tar xzf - -C .scratch/realdata
ln -s ../tracker.worldapi.org/.env .env # token only, never read; rm .env afterwards
.scratch/w068/start.sh 8760 TRACKER_BACKFILL=0 TRACKER_KINDS=1 # = TRACKER_SYNC=0, TRACKER_STORAGE=copy, own sessions dir
# session file for 0mujqspn0aqb in .scratch/w068/sessions (as mission 016 (old 060)), id → .scratch/w068/real-sid
REAL_PORT=8760 REAL_CHROME=8767 REAL_OUT=/tmp/w068/real node tests/realdata-068.mjs
TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb ./bin/hybriel tools/count-kinds.hl # server stopped
```
- **Backfill**: 1,789 TV titles in **838 s** (1,613 requests).
- Kinds: series 1,577, shows 212 (176 by genre without a request, 36 by TMDB type).
- 11 TMDB 404s, 0 retries, no slow step.
- TMDB types: Scripted 1,337, Miniseries 179, Documentary 50, Reality 16, Talk Show 15, Video 4, News 1.
- **Public lists**: before (genres only) 1,601 series / 173 shows; after **1,565 series / 209 shows**; 4,650 movies.
- **Spot checks** (search label → link):
- Show: Saturday Night Live, The Tonight Show Starring Jimmy Fallon, Last Week Tonight, The Late Show with Stephen Colbert,
Hot Ones.
- Series: Reacher, Foundation, Doctor Who (both), Chernobyl (Miniseries), Severance (2022).
- All correct.
- **Pages**, curl every 5 s during the backfill (156 each, avg / max):
| Page | avg | max |
|---|---|---|
| `/` | 0.094 s | 3.5 s |
| `/series` | 0.055 s | 2.6 s |
| `/shows` | 0.024 s | 0.21 s |
| `/series/Reacher` | 0.187 s | 0.48 s |
- 4 answers over 1 s, about every 170 s. Idle afterwards: avg 0.029, max 1.0 (1 of 196).
- Cause not found; probably mpackdb rewriting shows.db.
- Before the backfill: /series/Reacher 0.17 s, /shows/Saturday-Night-Live (function route) 0.28 s.
- **Browser** (signed in as the creator, full load to hydrated, ms 390 | 1280):
| Page | 390 | 1280 |
|---|---|---|
| `/` | 74 | 98 |
| `/series` | 97 | 43 |
| `/shows` | 351 | 43 |
| `/movies` | 123 | 43 |
| `/series/Reacher` | 290 | 243 |
| `/shows/Saturday-Night-Live` | 395 | 380 |
| `/movies/Star-Trek-First-Contact` | 251 | 246 |
| `/genres/Drama` | 46 | 44 |
| `/person/Alan-Ritchson` | 94 | 35 |
| `/my/series` (126 rows) | 1410 | 384 |
| `/my/shows` (3) | 220 | 217 |
| `/search/reacher` | 37 | 90 |
- Every heading's text and colour is right, nothing scrolls sideways, 0 console problems.
- Shots `/tmp/w068/real/*-{390,1280}.png` on Loreana (looked at).
- Copy, sessions and the `.env` link were removed and the server stopped.
**Open**:
1. Merge with main (062) and t19 / t26. They also change `components/show.hl`, `project.hl`, `styles.hl`, `tests/browser.mjs`,
`tests/faketmdb.mjs`. t19's franchise pages need the typed h1 (`type-franchise` / `type-timeline`), and its links to
`/shows/<slug>` should become `titlePath`.
2. TMDB type **Video** = anime OVAs here (Halo Legends, .hack//Liminality, Carnival Phantasm, Halo: The Fall of Reach). They
became Shows by the architect's rule; maybe Video → Series (creator).
3. Deploy: the first start runs the kind backfill after the adult backfill (~14 min).
4. The 2–3.5 s answers during the backfill (above).
5. Titles added by the person fill have no stored kind until the next start (the genres decide meanwhile).
## 2026-10-02 — ticket #18 (mission 021 (old 067), branch t18 from t26's head 10bb3f9, worktree tracker-wt18): daily sync by change lists, show record refresh, summary priority (NOT merged, NOT deployed)
**Built**: `deltasync.hl` (new): the daily run first reads TMDB `/tv/changes` + `/movie/changes` (since the stored day, all pages)
and TVmaze `/updates/shows?since=day|week|month`, one request per clock step → our changed titles (search index lookups, no scan),
followed first. Followed → full `syncShow` (as before); unfollowed → `syncShowDelta` (light: details + ids + poster + record + only
new/grown/newest seasons, no TVmaze). Full walk on: no `sync-state.txt`, gap > 14 days, failed change list (day not advanced).
`details.hl withDetailsFields` (title, tmdbSummary, homepage, tagline, status, genres, adult; shared with #26's completion) +
`applySynced` (fields + credits) replace `applyCredits` in the sync paths. TVmaze request = `/shows/<id>?embed=episodes` (summary →
`tvmazeSummary`, `stripHtml`). `summary` never written (sync, search import, person fill, completion); page `summaryOf`: summary >
tmdbSummary > tvmazeSummary. One-time boot clear of copied summaries (marker `summaries-cleared.txt`). Search: `searchTitleRenamed`,
`titleIdByTvmaze`. Tool: `tools/sync-tmdb.hl --delta [--since D] [--plan]`. Details: README "What it does (ticket #18)".
**Decisions (mine)**: (1) unfollowed changed series get the LIGHT step — the first real run with full steps blocked the server 20.8 s
for Today (15,716 episodes) and EastEnders (7,349) each and grew RSS to 14 GB + 2.9 GB swap; (2) TVmaze stays a followed-series thing
(its list only brings followed series); (3) the day stored = the day the finished run STARTED (overlapping windows, no gap); (4) light
step's seasons = missing, TMDB count > ours (fewer = TVmaze extras, skipped), newest; (5) a renamed title keeps its slug;
(6) `syncShowPart` reads only its part's seasons' episodes unless it finishes with the TVmaze merge (also speeds up #26's parts).
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~21:45 UTC = pre-#26/#28 live data, real TMDB/TVmaze, :8750, sync-state = 2026-09-30):
- Change lists for one day: TMDB tv 3,088 + movie 9,559 ids (31 + 96 pages), TVmaze 312 → **336 of our titles (28 followed)**
(152 series / 197 movies in the first variant).
- Delta run (final code): `tmdb sync done: mode=delta changeRequests=128 shows=336 newSeasons=79 newEpisodes=4475 updatedEpisodes=1305
updatedSeasons=70 posters=1 ids=0 requests=607 tvmazeRequests=29 errors=0 … seconds=536.359`; slowest step 2.5 s; pages meanwhile
(`/`, `/movies`, `/shows` every 0.5 s, n=925): median 45 / p95 226 / p99 693 / max 2216 ms. Same day again (steady state):
`shows=336 newSeasons=0 newEpisodes=0 updatedEpisodes=0 … requests=605 tvmazeRequests=29 errors=0 seconds=511.205`, slowest 2.06 s,
RSS 2.18 → 5.7 GB.
- OLD full walk on the same copy (no state file): `mode=full shows=128 … requests=259 tvmazeRequests=128 errors=1 seconds=312.241`
(the error: one TLS handshake failure, transient), RSS 2.1 → 4.9 GB.
- First variant (full step for every title): 349 titles, 659 + 155 TVmaze requests, 755 s, 46 slow steps up to 20.8 s, RSS 14 GB.
- Summaries: `summaries: cleared 9647 copies of tmdbSummary, kept 24 own texts, errors 0` (one time; ~+2 GB RSS at that boot). The 24
kept are all FOLLOWED shows whose tmdbSummary the #9 sync changed since the migration → older TMDB texts, likely no rewrites:
The-Flash, Magnum-P.I., Alex-Rider, Star-Trek-Strange-New-Worlds, Star-Trek-First-Contact (its tmdbSummary is ANOTHER title's —
"Kabuhayang Swak na Swak": wrong tmdbId on that record), Night-Sky, Good-Omens, Quantum-Leap-2, His-Dark-Materials, Not-Dead-Yet,
Star-Wars-Andor, The-Rookie, Will-Trent, Unstable-2, Reacher, Wednesday, Citadel, Miracle-Workers, FUBAR, The-Night-Agent,
The-Wheel-of-Time, Ahsoka, Percy-Jackson-and-the-Olympians, The-Daily-Show-with-Trevor-Noah.
- Record refresh, first day: 24 renamed (Star Wars: Andor → Andor, The Daily Show with Trevor Noah → The Daily Show, WWE Raw → Raw,
편의점 샛별이 → Backstreet Rookie, "Kaun Banega Crorepati(कौन बनेगा करोड़पति)" → "Kaun Banega Crorepati" — the old tracker's
"Name(original)" form goes), 79 tmdbSummary changed, 173 taglines.
- Screenshots (looked at): Loreana `/tmp/w067/real/{Star-Wars-Andor,Breaking-Bad,Kaun-Banega-Crorepati}-{390,1280}.png` (Andor: new
title; Breaking Bad: TMDB's current text after the clear), full load 134–649 ms, 0 console problems; gate `/tmp/w067-shots/`
(delta-show-1280, delta-movie-1280).
How to repeat (ports 8750–8759; ONLY `storage/`; the worktree has no `.env` → `ln -s ../tracker.worldapi.org/.env .env`, remove after):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
echo <yesterday> > .scratch/realdata/storage/mpackdb/sync-state.txt # none = the full walk
TRACKER_PORT=8750 TRACKER_SYNC=1 TRACKER_SYNC_HOUR=$(date -u +%-H) TRACKER_WATCH=0 TRACKER_BACKFILL=0 TRACKER_REPAIR=0 TRACKER_CREDITS=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8750 HL_HOST=127.0.0.1 ./bin/hybriel project.hl > .scratch/delta.log 2>&1 &
REAL_PORT=8750 REAL_LOG=.scratch/delta.log node tests/realdata-018-poll.mjs # latency until "tmdb sync done"
grep '^tmdb sync\|^summaries' .scratch/delta.log; RSS: grep VmRSS /proc/<pid>/status
REAL_TITLES=Star-Wars-Andor,Breaking-Bad REAL_OUT=/tmp/w067/real node tests/realdata-018.mjs
# afterwards: kill <pid>; rm -rf .scratch/realdata .scratch/realdata-sessions; rm .env
```
**Gate** (249 → 261; 12 new #18 checks; the TVmaze episode paths are `/tvmaze/shows/<id>?embed=episodes` now; the #26 completion
check now expects TMDB's overview — `summary` is no longer filled by the fill/completion; the seed movie's summary is a copy of its
tmdbSummary; `tests/peek-shows.hl` reads records with the app stopped):
```
TRACKER_GATE_SHOTS=/tmp/w067-shots TRACKER_GATE_PORT=8751 TRACKER_GATE_IDENT_PORT=8752 TRACKER_GATE_TMDB_PORT=8753 TRACKER_GATE_CHROME=8754-8756 node tests/browser.mjs
261 passed, 0 failed
```
**Open (for the architect)**:
- NOT "a handful of requests": movies change a lot on TMDB (9,559 ids/day → 96 pages) — a day = 128 change-list requests + ~480 title
requests + 29 TVmaze ≈ 634 (old full walk: 387) but every title stays current, followed ones only when they changed (28 vs 128).
Dropping unfollowed MOVIES would save ~96 pages + ~200 requests — creator's call.
- MEMORY: ~10 MB per synced title are never freed (Hybriel, see #28) → +3.5 GB RSS per daily run until a restart (old run +2.8 GB);
the live container already used 11.4 GiB. A daily restart or the runtime fix is needed either way.
- Titles: TMDB's en-US name replaces the migrated title (24 on day one, incl. the "Name(original)" form). If the creator wants to keep
some, a creator-title field like `summary` would be needed.
- The 24 kept "own" summaries look like older TMDB copies (list above) — clear them too? (architect's rule was "identical only").
- Merge: t18 is on t26's 10bb3f9; t19 (franchises) and the 062 merge also touch people.hl/search.hl/README/STATUS — expect conflicts
in README/STATUS only (code: people.hl 1 line, search.hl small additions).
## 2026-10-01 — mission 018 (old 062): t26 (#26 + #28) merged into main, short ids for new people, lazy guests, lean watch clicks (NOT deployed)
**Merge**: `git merge t26` into main (fa1f9df). Conflicts were in README/STATUS, `components/show.hl` (header comment;
skeleton poster + short id) and `tests/browser.mjs` (startServer env, import checks, #27 + #26/#28 blocks). Both sides kept.
**Fixes**:
1. **Short ids**: people.hl `castPersonId` calls `claimShortId('person', …)`. It is the only place that creates people:
- #26 completion, credits job, daily sync, `tools/sync-tmdb.hl`, search import;
- the guest route.
- shortids.hl also finds `../data/old-short-ids.json` from tools/.
2. **Guests are no person records**: details.hl `mergedCast` keeps a guest on the title as
`{ person: null|existing, tmdbId, actor, character, profile, episodes, guest }`.
- Link: `/person/tmdb/<id>?show=<show id>` → project.hl `guestRoute` → people.hl `guestPersonId`.
- It creates the person with a short id (from that title's entry; without `?show=`, from TMDB `/person/<id>`), then 302.
3. **Watch/follow clicks**:
- `showRow` is a small `pageShowOf` object. The whole record (~300 cast + crew) was shipped with the page and re-sent with
every session sync.
- cast/crew derive from the slug (`castOfSlug`/`crewOfSlug`; adult titles: none).
- `watchesOfUser` is cached per user until `setWatched`.
- The faces return rows only when the page opened/closed a season (`collapsedChanged`). Otherwise the sync's rows answer.
**Gate** (266 main + 35 t26 → 311; mission 018 (old 062): 10 new checks, the #26 new-person check extended, sync-tool expectation +27 → +25 people):
- count-short-ids runs after the sync tool and after the credits job;
- import and #26 people have short ids at once;
- guest links, guest click → person page with short id + TMDB fill, again → 302, TMDB fallback, 404s;
- a watch click's frames carry no cast/crew;
- a client-side navigation + watch click gets rows by the sync only.
```
mkdir -p /tmp/w062-shots
TRACKER_GATE_SHOTS=/tmp/w062-shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
311 passed, 0 failed
```
Shots looked at: `/tmp/w062-shots/guest-person-{390,1280}.png`, `credits-series-all-390.png`.
**Real data** (copy of Byrodin `storage/` 2026-10-01 22:05, real TMDB, :8730; the copy is deleted):
- **Jobs**:
- Details repair: 557/557 in 836 s, 0 failed.
- Credits job: 5,867 titles.
- Run 1 (1,961 titles) ended at ~22:41 with no message; the cause is not known (no OOM line visible).
- The restart resumed: `credits job done: 3906/3906 … people=47242 … seconds=2028`. The next start: nothing to do.
- `short ids: 112298 records, 112298 with a short id, 0 without` (every person created had one at once).
- **RSS** (`ps` every 60 s, `/tmp/w062/rss.log`):
- During the jobs: 1.2–6.3 GB, swinging; it falls back by GBs at times.
- 2 slow steps of 4.9–5.6 s, where RSS dropped too, so probably the runtime freeing memory. Pages wait meanwhile.
- **People / boot**:
| | people | search index | RSS ~30–60 s after boot |
|---|---|---|---|
| main fa1f9df, same copy (no token, no jobs) | 15,157 | 0.66 s | 4.3–4.5 GB |
| merged, after the jobs (jobs off) | 102,424 | 3.0 s | 6.4–7.0 GB |
| merged, after the jobs (jobs on, nothing to do) | 102,287 | 2.7 s | 8.1–8.8 GB |
- w061 (all guests as people): 160,084 people.
- Movie casts (all "main") are most of the 87k new people.
- **Pages** (`REAL_OUT=/tmp/w062/real node tests/realdata-062.mjs`, signed in as az5b2):
- Star Trek: Picard: 202 cast (94 guests linked lazily) + 45 crew, HTML 143 kB. Full load 201–342 ms; the first after boot
took 2.3 s.
- Superman: 90 cast, 8 crew, 117–173 ms.
- Guest click (Stephanie Czajkowski) → person created (short id 401md) + filled from TMDB (18 films) in 0.74–1.3 s. Again: 29–66 ms.
- **Watch click** (page clock, click → answer frame):
| | before | after |
|---|---|---|
| Picard | 184–200 ms (w061) | 159–174 ms, 4.9 kB |
| The Simpsons | 276–731 ms (w061) | 156–214 ms, 15 kB (was 29 kB) |
| Follow | | 119–221 ms |
About 1 in 6 clicks took 550–970 ms (same pauses as above).
- Server timing (temporary logs, removed): the face is ~40 ms, the page's re-mount ~25 ms of our code; the rest (~120 ms) is
hl:web's session-face path.
- Screenshots (looked at): `/tmp/w062/real/{Star-Trek-Picard,Superman}{,-credits}-{390,1280}.png`, `guest-person-{390,1280}.png`.
How to repeat (ports 8730/8731; ONLY `storage/`):
```
mkdir -p .scratch/realdata && ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8730 TRACKER_URL=http://127.0.0.1:8730 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_WATCH=0 TRACKER_SYNC=0 TRACKER_CREDITS_LOG_EVERY=250 ./bin/hybriel project.hl
# (the runtime reads .env → real TMDB: repair ~14 min, then credits ~50 min; RSS: ps -o rss= -p <pid> every minute)
# session for 0mujqspn0aqb as in mission 016 (old 060) → .scratch/real-sid; then:
REAL_OUT=/tmp/w062/real REAL_CHROME=8731 node tests/realdata-062.mjs
# afterwards: kill the server; rm -rf .scratch/realdata .scratch/realdata-sessions .scratch/real-sid
```
**Deploy (what to watch)**:
- After the start: the details repair (~14 min), then the credits job (~50 min).
- `docker stats`: expect RSS to roughly double after boot (4.3 → ~8 GB) and swing during the jobs.
- Log: `details repair done`, `credits job done`, `slow step`, `guest opened`.
**Open**:
1. Deploy.
2. 102k people is still ~6.8× the old count. Most are movie cast. If memory is too high, apply the same lazy rule to a
movie's cast after the first N (not built: the mission kept "main cast").
3. The Hybriel runtime: RSS swings, multi-second pauses, a silent process end in run 1. For the architect.
4. Watch clicks: about 1 in 6 is slow; the base cost is hl:web's session-face path (~120 ms).
5. Adult titles show no cast/crew now (they never get credits; at most a migrated cast was hidden).
## 2026-10-01 — ticket #29 (mission 019 (old 063)): unwatched checks looked "active" (NOT deployed)
**Cause**: NOT a code regression of #23 — on real data (as az5b2) all 1,287 `/my/unwatched` checks are `checkicon regular`
(fill transparent), fresh load and after a client-side navigation; live CSS = HEAD. The look itself was the problem: #11's
"inverted check" (outline circle + outline check) was drawn in the ACCENT colour, so a page full of them read as ticked
(creator: "in /my/unwatched all checkboxes are active"). **Fix** (`styles.hl`): the unwatched outline + check are muted grey
(`colorTextMuted`, rgb(145,150,155)), accent only on hover; watched unchanged (solid accent disc, dark check). Keeps #11's
shape (not a bare empty circle). Applies everywhere the check is used: /my/unwatched, show seasons + episodes, movie page.
/my/schedule has no checks; /my/movies only marks watched movies (#24).
**Gate** (259 → 266): `checkLook(sel)` in tests/browser.mjs measures a REAL check on the page (no hidden copy any more):
computed circle fill/stroke + check stroke, topmost element, and the SCREENSHOT pixel inside the disc (svg 12,6) →
`lookEmpty` / `lookSolid`. On /my/unwatched (both rows), show page (unwatched episode + season, watched episode, #23 mixed
state), movie page (both states), /my/movies (watched mark). Against the old styles.hl it fails 4 checks (stroke accent).
```
mkdir -p /tmp/w063/shots # the gate needs the shots dir to exist
TRACKER_GATE_SHOTS=/tmp/w063/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
266 passed, 0 failed
```
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~19:30; procedure as mission 016 (old 060) — the inner ssh from Loreana to
Byrodin was refused once, so the tar was piped from Byrodin into Loreana instead; server :8730 also needs
`TRACKER_OLD_SHORT_IDS=$PWD/data/old-short-ids.json`): `REAL_OUT=/tmp/w063/real node tests/realdata-063.mjs` → /my/unwatched
1,287 checks all regular/muted/unfilled (390: 1913 ms, 1280: 1050 ms full load), Foundation 11 muted + 2 solid, Star Trek:
First Contact movie check muted, /my/movies none (creator has no watched movie); one click 1287 → 1286 rows in 1.5 s, gone
after a reload; 0 console problems. Shots /tmp/w063/real/{my-unwatched,show,my-movies}-{390,1280}.png (looked at).
**Open**: deploy; creator to confirm the muted look (if they want a bare empty circle that reverses #11).
## 2026-10-01 — ticket #27 (mission 016 (old 060)): short ids for titles and persons (NOT deployed)
**Built** (details README "What it does (mission 016 (old 060))"): `shortId` (5 × `[a-z0-9]`) on shows (movies + series) and persons, unique
across both through one in-memory map (`shortids.hl`). The 702 old ids come back from `data/old-short-ids.json` (made once by
`tools/old-short-ids.hl` from the old export); they are reserved for new ids. New records get one at creation (search import, the
person fill's titles). A background backfill (own 0.1 s clock, 40 ms batches, resumes) gives everyone else one. `#shortid` under
the poster / photo (no photo: under the facts), small, muted, monospace. `/<shortId>` → 301 to the page, else 404.
**Decisions**: backfill in the background, not at start (all 25k at once ≥ 16 s blocked; batches measured ≤ 62 ms). Uniqueness via
the in-memory map, not an mpackdb `!shortId` index (per table only; adding an index to the live tables is untested, hybriel#21).
Old-id file in the code (`data/`, deployed), not in `storage/`. Upper case in `/<id>` is read as lower case. A record shows no id
until the backfill reaches it (first ~5 min after the deploy).
**Gate** (238 → 259):
```
TRACKER_GATE_SHOTS=/tmp/w060/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
259 passed, 0 failed
```
The gate gets its own old-id file (`TRACKER_OLD_SHORT_IDS`: sh1 → 100jh). Shots looked at: Loreana /tmp/w060/shots/shortid-*.
**Real data** (copy of Byrodin `storage/` taken 2026-10-01 ~14:09 Loreana time: 10,010 shows, 15,157 persons; :8730):
- Backfill: 25,167 records in **297 s** (another run 307 s), 1,859 batches, slowest batch 62 ms, slowest persist 31 ms; boot map 0.12–0.22 s.
- `tools/count-short-ids.hl` afterwards: all 10,010 + 15,157 have one, old kept **702/702**, lost 0, old id on another record 0,
**duplicates 0**; `tools/count-duplicate-ids.hl`: 0 duplicate @ids in all tables.
- Pages by curl, back to back, during the backfill (avg / max / answers > 1 s of ~520 each): `/shows/Foundation` 0.234 / 1.56 / 9,
`/person/Jared-Harris` 0.170 / 1.37 / 9, `/` 0.070 / 1.34 / 2, `/movies` 0.073 / 1.18 / 2. Without a backfill (120 s, 436 each):
0.173 / 1.34 / 5, 0.046 / 1.10 / 5, 0.022 / 0.99 / 0, 0.015 / 1.01 / 1 — the > 1 s answers happen without it too (Loreana was
busy with other workers); the backfill adds ~0.1 s on average.
- Browser (`node tests/realdata-060.mjs`, signed in as the creator, full loads ms 390 | 1280): Raised by Wolves (via `/100jh`)
1372/297/310 | 335/266/301, Foundation (`100sp`, an old id too) 302/306/250 | 1246/205/195, Star Trek: First Contact (`aafif`)
275/287/275 | 211/235/212, Jared Harris (`yg7lk`) 164/138/122 | 104/119/121; each id under its image; every `/<id>` 301 in 1–27
ms and the browser lands on the page; `/zzzzz` 404; 0 console problems. Shots /tmp/w060/real/{old-id,series,movie,person}-{390,1280}.png (looked at).
How to repeat (ONLY `storage/`; inner ssh with `< /dev/null`):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_OLD_SHORT_IDS=$PWD/data/old-short-ids.json ./bin/hybriel tools/count-short-ids.hl
OUT=/tmp/w060 tests/realdata-060-backfill.sh 1 450 "" # starts :8730 on the copy, curl-times 4 pages until "short ids done"
# (3rd arg `full` = keep timing until MAX — the baseline once nothing is left). Then, server still up: session file for user
# 0mujqspn0aqb (name sha256(<32-hex id>), JSON {"created":<now ms>,"data":{"tag":"t"},"id":"<id>","seen":<now ms>,"user":{"id":"0mujqspn0aqb"}}
# in .scratch/realdata-sessions/), id → .scratch/real-sid
REAL_OUT=/tmp/w060/real node tests/realdata-060.mjs # REAL_SERIES, REAL_MOVIE, REAL_PERSON, REAL_OLD, REAL_REPEAT
kill $(cat /tmp/w060/real.pid); TRACKER_STORAGE=… tools/count-short-ids.hl again; rm -rf .scratch/realdata .scratch/realdata-sessions .scratch/real-sid
```
**Open**: (1) deploy — the first start runs the backfill (~5 min, pages ~0.1 s slower meanwhile; ids appear as it goes).
(2) Inside the app a record costs ~3 ms vs ~0.6 ms in a plain script (cause not found; a big `static` list read through a field
was copied on every access — fixed by keeping the queue a plain `static` list). (3) Nothing creates persons in the app today; if a
later step does, it must call `claimShortId('person', …)`. (4) `/<shortId>` of an adult title redirects; its page says "Not found"
to non-followers as before.
## 2026-10-01 — tickets #22–#25 (mission 014 (old 058)): air dates, season check, movie watched check, genre pages (NOT deployed)
**Built** (details README "What it does (mission 014 (old 058))"): #22 every episode row shows its air date (server rows + opened
seasons). #23 a season's check is DERIVED = all its episodes watched (live with the face's answer; a season click on a partly
watched season watches the rest). #24 a movie's page has `#moviewatch` (watch `targetType 'movie'`, face `showToggleMovie`;
signed out → modal), `/my/movies` says "N movies · M watched" + a check per watched row. #25 `/genres/<genre>` (+ `/page/<n>`,
`components/genre.hl`, catalog.hl `index.genres`): movies AND series, newest first, 24 per page, adult hidden; pills link there,
`/genre/<x>` → 301.
**Decisions**: the season record is still written by a season click but no longer read for the check. A series' place in a
genre list = its newest released episode date (else its release), a movie's = its release (one comparable date). Genre by
`urlSegment` (case-insensitive fallback). TMDB's TV and movie genres stay separate (see Open).
**Gate** (214 → 238):
```
TRACKER_GATE_SHOTS=/tmp/w058/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
238 passed, 0 failed
```
One old check changed: "an unwatched check is the inverted icon" now styles a detached copy classed `regular` (with #23 every
check on that page is solid at that point). Fixture: genre Action (`g0000000000000g2`) on Pager Movie 01–30, Pager Future Movie,
Pager Older Series + the adult `Hidden Action Film`. Shots looked at (Loreana /tmp/w058/shots/: show-open-*, movie-watched-*,
genre-action-*, my-movies-*).
**Real data** (copy of Byrodin `storage/` 2026-10-01 12:03 UTC, :8730, creator 0mujqspn0aqb, `node tests/realdata-058.mjs`, full
load goto → hydrated, 3 repeats, ms 390 | 1280): `/shows/Foundation` 549/277/221 | 224/228/225 (10 open rows all dated, top season
check == all its episodes); `/shows/Star-Trek-First-Contact` (movie) 234/250/250 | 237/238/239; `/genres/Action` 118/991/42 |
92/39/43 (643 titles, 27 pages; server 6–8 ms by curl — the spikes are poster loads); page 2 41/103/749 | 93/48/110; `/my/movies`
66/63/66 | 68/63/63. Movie check on the copy: click 361 ms → reload solid → click back → reload regular. Genre kinds page 1:
Action 24 movies/0 series, Drama 5/19, Action-Adventure 0/24. 0 console problems. Screenshots /tmp/w058/real/ on Loreana (looked at).
How to repeat (ONLY `storage/`; inner ssh with `< /dev/null`):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8730 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_BACKFILL=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8730 HL_HOST=127.0.0.1 ./bin/hybriel project.hl
# session file for 0mujqspn0aqb (name sha256(id), created/seen = now ms), id → .scratch/real-sid
REAL_OUT=/tmp/w058/real node tests/realdata-058.mjs # REAL_SHOW, REAL_MOVIE, REAL_GENRE, REAL_REPEAT
# afterwards rm -rf .scratch/realdata .scratch/realdata-sessions .scratch/real-sid
```
**Open**: (1) deploy. (2) TMDB's TV genres ("Action & Adventure", "Sci-Fi & Fantasy", "War & Politics") are separate genres from
the movie ones, so `/genres/Action` has no series on real data — merge them? (creator). (3) A series tile shows its start year
while it is placed by its newest episode (same as /shows). (4) The season's own watch record can now disagree with its episodes
(not read anywhere for the check).
## 2026-10-01 — ticket #28 (mission 017 (old 061), branch t26 on top of #26, worktree tracker-wt26): full cast + crew (NOT merged, NOT deployed)
**Found** (live copy 2026-10-01 ~14:40): the page showed `show.cast` — the old tracker's migrated list for 1,820 of 10,011 titles
(Star Trek: Picard **4**: Stewart, Ryan, Hurd, Speleers; Dune (1984) 39), nothing for the rest (Superman 1978: **0**); #26 only
added TMDB's first 20 to filmography titles; no crew anywhere; the daily sync and the search import never stored cast.
**Built**: details.hl `castOf`/`crewOf`/`mainMinOf`/`withCredits`/`applyCredits` + the credits job (`creditsQueue`/`creditsStep`,
project.hl `creditsTick`); tmdbsync.hl asks `external_ids,aggregate_credits|credits` in the details request it makes anyway and
returns it (`res.details`); search.hl `importTitle` asks once (ids + credits) and hands that answer to `syncShowWith` (one request
less), components/search.hl stores the credits; tools/sync-tmdb.hl too; show page: crew block (job labels) + cast `Name (Character)`
collapsed after 20 with a client-only "Show all (N)" (crew: 6 per job, "Show all crew (N)"), "Guest stars" label; shows.hl
`showBySlug` by a slug → id map. Details: README "What it does (ticket #28)".
**Decisions**: cast ≤ 300 (`TRACKER_CAST_MAX`; SNL 2,771 / Grey's 3,243 one-episode guests would each be a person), crew ≤ 30 per
job; guest = fewer than half the series' episodes (≤ 10, ≥ 1) — Picard (30 ep.): ≥ 10; composer = Original Music Composer /
Composer / Music; Writer, Screenplay, Story separate groups (Teleplay not shown). The credits job runs after the backfill and the
repair, only public titles (`adult == false`), one request per title + 260 ms, pauses above `TRACKER_CREDITS_MAX_MB` (16000).
**Hybriel finding (not filed; for the architect)**: the native runtime never gives back what a `fetch` body or a decoded record
took — repro (10 lines): a node server answering a 440 KB JSON; a .hl loop `fetch(url)` + `r.json()` 300× → RSS +2.7 GB (fetch
without `.json()`: +0.43 GB), sampled with `ps -o rss=`; nothing is given back while the process lives. The same
for `showsTable.find(null, null)`: the old `showBySlug` decoded every title per call → run 1 below grew to 20 GB in 20 min. The LIVE
container used **11.4 GiB** today (docker stats, before any of this). Fix in this branch: the slug map (no per-page scan) + the job's
RSS limit; the leak itself is Hybriel's.
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~14:40, real TMDB/TVmaze, :8750; copy deleted):
- Before → after: **Star Trek: Picard 4 → 202 cast** (17 main ≥ 10 episodes incl. season-3 regulars Gates McFadden, Todd Stashwick, then
guests from John de Lancie (9)) + 45 crew (Created by Kurtzman, Beyer, Goldsman, Chabon; Directed by Frakes …; Written by; Story;
Music Jeff Russo …). **Superman (1978) 0 → 90 cast** + Directed by Richard Donner, Written by Tom Mankiewicz, Screenplay Newman /
Benton / Puzo / Newman, Story Puzo, Music John Williams. Dune (1984) 39 → 41. The Simpsons 6 → 300 (cap) + 67 crew.
- Run 1 (`TRACKER_REPAIR=1 TRACKER_CREDITS=1`): `details repair done: 557/557 … cast=30286 … seconds=800`, then the credits job
`start, 5867 titles`; stopped after 714 (RSS 20 GB — pages scanned every show per call, see the Hybriel finding). Run 2 (slug map,
light persist, `TRACKER_CREDITS_MAX_MB=16000`): resumed `start, 5153 titles` → `credits job done: 5153/5153 titles, done=5153
failed=0 cast=262634 crew=39624 people=104614 requests=5153 retries=0 slowestStepMs=1316 seconds=2681` (44.7 min; whole job ≈ 51
min); RSS 3.8 → 6.3 GB (the pause did not trigger; it is gate-tested). Next start: `credits job: start, 0 titles` / `nothing to do`.
- Page latency meanwhile (`/movies`, Picard, Foundation, `/` in turn every 0.5 s, n=4836): median 28 ms, p95 194, p99 436, max 1125 ms.
(Run 1 with the old `showBySlug`: Picard median 318 ms, max 3.8 s; during the repair median 179 ms, max 4.1 s.)
- After (no job running): show page full load 105–235 ms (Picard 206 ms, first after boot 2.1 s = the slug map), Picard HTML 166 kB,
Simpsons 364 kB; "Show all" 2–18 ms, 0 faces sent; no sideways scroll at 390. Signed-in watch click (session face re-derives the cast
rows too): Picard 184–200 ms (first 1.3 s), The Simpsons 276–731 ms. Boot: search index 10,011 titles / **160,084 people** (was
15,157) in **4.3 s (was 0.66 s)**, RSS after boot **4.9 GB (was 1.6 GB)**.
- Screenshots (looked at) on Loreana `/tmp/w061/real/{Star-Trek-Picard,Superman,Dune,The-Simpsons}-{collapsed,all}-{390,1280}.png`;
gate `/tmp/w061/shots/credits-{movie-1280,series-collapsed-390,series-all-390,series-1280}.png`. 0 console problems.
How to repeat (ports 8750–8753; ONLY `storage/`; the worktree has no `.env` → `ln -s ../tracker.worldapi.org/.env .env`, remove after):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8750 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_BACKFILL=1 TRACKER_REPAIR=1 TRACKER_CREDITS=1 TRACKER_CREDITS_LOG_EVERY=250 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8750 HL_HOST=127.0.0.1 ./bin/hybriel project.hl
# latency meanwhile: a node loop fetching /movies, /shows/Star-Trek-Picard, /shows/Foundation, / every 0.5 s until 'credits job done'
# (mission 017 (old 061)'s: .scratch/w061/poll.mjs); grep '^credits job\|^details repair' in the log; RSS: ps -o rss= -p <pid>
REAL_TITLES=Star-Trek-Picard,Superman,Dune,The-Simpsons REAL_OUT=/tmp/w061/real node tests/realdata-028.mjs
# afterwards: stop the server, rm -rf .scratch/realdata .scratch/realdata-sessions; rm .env
```
**Gate** (231 → 249; 18 new #28 checks; fake TMDB fixtures: show 1001 aggregate_credits + created_by, movie 1003 credits, import
1004, Backfill Movies' credits; the sync/import request paths now carry the credits; `TRACKER_CREDITS=0` by default):
```
TRACKER_GATE_SHOTS=/tmp/w061/shots TRACKER_GATE_PORT=8754 TRACKER_GATE_IDENT_PORT=8755 TRACKER_GATE_CHROME=8756-8758 TRACKER_GATE_TMDB_PORT=8759 node tests/browser.mjs
249 passed, 0 failed
```
**Merge notes** (main has missions 014 (old 058)/016 (old 060) meanwhile): `components/show.hl` — header comment, statics after `genreRowsOf`
(`castShown`…`crewCountOf`, replacing `castRowsOf`), `completedPageOf`'s answer, members after `hasCast`, the View's crew + cast
blocks (replacing the old `showCast` loop), `completeTick` assignments, handlers `toggleCastAll`/`toggleCrewAll` before
`toggleEpisode`. `styles.hl` — the cast/crew block replaces the two `.artist-link` lines. `shows.hl` — `showBySlug`. `tmdbsync.hl` —
details paths + `res.details`. `search.hl` — `importTitle` (one request, `syncShowWith`). `project.hl` — `readFile` import, sync tick
`applyCredits`, the credits job block before the clock, the clock's tick. Gate expectations that name request paths changed (sync tool,
import, failed-import message). **Deploy**: repair (~557 titles, ~13 min) then credits job (~5,900 titles, ~51 min); watch RSS
(`docker stats`) — live already used 11.4 GiB before this.
**Open**: (1) merge + deploy. (2) The Hybriel leak (fetch bodies / decoded records never freed) — architect to file; the live app
grows with every sync/backfill too. (3) 160k people: boot +3.6 s, +3.3 GB RSS — if too much, store cast without creating people for
guests (link resolved on click) — not built. (4) Session faces (watch/follow clicks) re-derive the cast rows (`castRows` derives from
`showRow`, which reads the session): Simpsons click up to 0.7 s; deriving them from the slug would avoid it (adult pages need care).
(5) A followed title's credits refresh with the daily sync; others are never refreshed. (6) Cast persons' slugs as #26 (`-2` rule).
## 2026-10-01 — ticket #26 (mission 015 (old 059), branch t26, worktree tracker-wt26): filmography titles complete (NOT merged, NOT deployed)
**Found**: the person fill (#16, people.hl `addMinimalTitle`) stores only title, year/release, overview, type, tmdbId, adult, genres,
language and that one person as cast — no poster, seasons/episodes, imdb/tvdb/tvmaze ids, homepage/tagline/status, `tmdbSync`; the
show page just showed that, forever. A search import (`importTitle` + `syncShow`) has all of it except the cast. Live copy: 321 such
titles (6 people filled so far).
**Built**: `details.hl` (`completeStep`), show page loading state (skeleton poster / cast line / 3 season rows, probe → face
`showComplete` until done, no reload), `project.hl repairTick` (the details repair, after the adult backfill), tmdbsync.hl
`syncShowWith`/`syncShowPart` (sync with a given details answer, a series in parts), people.hl `castPersonId` + `minimal = true`,
search.hl `slugBaseOf`/`indexPerson`. Details: README "What it does (ticket #26)".
**Decisions**: complete = what a full import has + the cast (TMDB's first 20, billing order; series: aggregate_credits; adult
people left out; new people get a record + slug + search entry, their page fills on first visit). Marker `detailsAt`; incomplete =
`minimal` or `imported` without `tmdbSync`. Adult titles are never completed (not on open — the public face refuses them — not by
the repair). TMDB 404 → `detailsCheck`, not asked again. A series in steps (≤ 300 episodes per step) — the first real run did the
Late Show in ONE 10.7 s step, pages waited up to 10.3 s. Titles a fill adds while the app runs: completed on open or at the next
start (the queue is built at start).
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~13:55, real TMDB/TVmaze, :8750, `TRACKER_BACKFILL=0`; copy deleted):
- Click from a person page (fill first), `node tests/realdata-026.mjs` — ms from the click: skeleton shown / complete / poster:
| title (person, width) | shown | complete | poster | server (log) |
|---|---|---|---|---|
| Violent Night 2, movie (Jared Harris, 390) | 186 | 712 | 765 | 1 step, 255 ms, cast 14 |
| Royal Kill List, series (Jared Harris, 390) | 227 | 1361 | 1363 | 3 steps, 592 ms, 1 season / 3 ep |
| Teen Wolf: The Movie (Tyler Hoechlin, 1280) | 216 | 662 | 662 | 1 step, 189 ms, cast 20 |
| Wolf Watch, series (Tyler Hoechlin, 1280) | 258 | 1316 | 1318 | 3 steps, 537 ms, 3 seasons / 28 ep |
| Avengers: Doomsday (Patrick Stewart, 1280) | 178 | 642 | 694 | 1 step, 247 ms |
| Bait, series (Patrick Stewart, 1280) | 181 | 1262 | 1263 | 3 steps, 594 ms |
/movies asked every 100 ms meanwhile: max 283–586 ms. Reload afterwards: complete at once, 157–205 ms full load, same parts.
Direct open (1280): Late Show with David Letterman 11.4 s (23 steps, 4,252 episodes), Watch What Happens Live 11.1 s (18 steps,
3,220 ep); /movies meanwhile max 1.3 / 1.5 s (before the steps: one 10.7 s step).
- **Repair** (restart with `TRACKER_REPAIR=1` on a fresh copy): `details repair done: 321/321 titles, complete=321 failed=0 adult=0
newSeasons=638 newEpisodes=16528 posters=299 cast=5196 requests=787 tvmazeRequests=218 retries=0 seconds=507.248`; no slow step
(> 1.5 s); /movies + /shows/Foundation every 0.5 s meanwhile: n=1298, median 0.13 s, p95 0.25 s, p99 1.03 s, max 2.35 s. Next
start: `start, 0 incomplete titles` / `nothing to do`. (First try, one step per title: 401 titles in 628 s, max wait 10.3 s.)
- Screenshots (looked at) on Loreana `/tmp/w059/real/`: `{movie,series}-{skeleton,complete,reload}-{390,1280}.png`,
`direct-*-1280.png`, `repaired-{Licorice-Pizza,Late-Show-with-David-Letterman}-{390,1280}.png`. 0 console problems.
How to repeat (ports 8750–8753; ONLY `storage/`; the worktree has no `.env` → `ln -s ../tracker.worldapi.org/.env .env` for the
real token, remove afterwards):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8750 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_BACKFILL=0 TRACKER_REPAIR=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8750 HL_HOST=127.0.0.1 ./bin/hybriel project.hl
REAL_PEOPLE=Jared-Harris:390,Tyler-Hoechlin:1280,Patrick-Stewart:1280 REAL_OUT=/tmp/w059/real node tests/realdata-026.mjs
REAL_PEOPLE= REAL_TITLES=Late-Show-with-David-Letterman node tests/realdata-026.mjs
# repair: restart with TRACKER_REPAIR=1, poll /movies, grep '^details repair' in the log (~8.5 min)
# count incomplete: the log line 'details repair: start, N incomplete titles' (N = 321 on 2026-10-01)
```
**Gate** (214 → 231; 17 new #26 checks, the offline home check now expects 6 series):
```
TRACKER_GATE_SHOTS=/tmp/w059/shots TRACKER_GATE_PORT=8754 TRACKER_GATE_IDENT_PORT=8755 TRACKER_GATE_CHROME=8756-8758 TRACKER_GATE_TMDB_PORT=8759 node tests/browser.mjs
231 passed, 0 failed
```
Gate env: `TRACKER_REPAIR=0` by default, `TRACKER_DETAILS_PART_EPISODES=1` (a season per step), the repair block sets
`TRACKER_REPAIR_FAST=1`. Shots `title-skeleton-1280.png`, `series-skeleton-1280.png` (looked at).
**Merge notes** (main has mission 014 (old 058) meanwhile — #22–#25 touch `components/show.hl`): my show.hl changes are additive — an
import, 2 statics before "reactive members", 4 members after `adultShow`, the poster/cast/loading blocks in the View, the
`completeTick` handler before `toggleEpisode`, the face `showComplete` first among the faces. If 058 changes `buildRows`' signature,
`completedPageOf` calls `buildRows(show, null, cl)`. `tmdbsync.hl`: `syncMovie(show, res, pre)` gained a parameter, `syncShow` is
now `syncShowPart(showId, null, null)`. `project.hl`: 1 import, the repair block before the clock, the clock's tick. Gate: fixtures
(movies 2001/2004/2005/2101–2178, tv 2002, TVmaze 602, person 3010), startServer defaults, the #26 block inside #16, the repair
block before #10, the offline home check. **The live deploy**: the repair waits for the adult backfill (~80 min), then ~9 min.
**Open**: (1) merge + deploy. (2) A search import still has no cast (could use `completeStep`). (3) Client-side navigation keeps
the person page's scroll position on the title page (hl:web; seen on the phone shots, not new). (4) Titles added while the app runs
are not queued for the repair (open or next start). (5) Cast persons of a title are never refreshed; people's slugs from cast use
the same `-2` rule as the migration.
## 2026-10-01 — ticket #17 (mission 013 (old 057)): season caret up/down, skeleton rows, faster opening (NOT deployed)
**Built** (`components/show.hl`, `styles.hl`, gate): caret DOWN closed / UP open. Opening a season: caret up + min(episodes, 8)
shimmering skeleton rows at once (client-side), then the sessionless face `showSeasonEpisodes(seasonId)` → `filledRows`
replaces them. Closing: client-only, no request. Watch state from member `watchedEpisodes` (per season `{ all, except }`).
Adult title pages (followers only) keep the session face `showRows(slug, collapsed)` (signature changed: no toggle inside).
Details: README "What it does (ticket #17)".
**Why it was slow** (measured): server work per open was small (probe on the copy: showBySlug 60–78 ms linear scan,
watchesOfUser 20 ms, watchedSet 6, episodesOfSeason 6) — but the face took `session`, so hl:web RE-MOUNTED the page on the
server and shipped every session-derived member (`rows` with the initial seasons!) as `sync`, which the client repainted,
then the handler repainted again. TRAP for future faces: a face with `session` = a full page re-render server-side + sync.
**Real data** (copy of Byrodin `storage/` 2026-10-01 12:25, :8730, creator az5b2, `node tests/realdata-057.mjs`, ms in the
page from click; 3 repeats; 390 and 1280 alike):
| | Daily Show S30 (142 ep) | Daily Show S28 (123) | LWT S12 (30) | LWT S10 (21) |
|---|---|---|---|---|
| before: episodes visible | 380–450 | 355–385 | 265–1093 | 235–285 |
| after: skeleton visible | 77–106 | 76–90 | 25–54 | 24–38 |
| after: episodes visible | 233–288 | 226–268 | 58–102 | 54–74 |
| close before → after | 420–1388 → 135–175 | 420–500 → 135–165 | 285–390 → 28–45 | 260–370 → 26–30 |
Skeleton rows 8 (cap) on all; faces sent: only showSeasonEpisodes, never showRows; client-filled watch state == the server's
rebuild (Daily Show 296 checks/121 solid, LWT 74/29). Page load unchanged (Daily Show ~310–360 ms warm both; HTML 157 → 167 kB
from `watchedEpisodes` — a plain id map was 274 kB, hence `{ all, except }`). Adult path checked once on a 2nd copy with LWT
set `adult = true` (:8728): signed out "Not found", creator opens seasons via showRows (skeleton 8, rows 300–400 ms, no dup).
The Daily Show's skeleton needs ~80 ms because hl:web re-renders the whole flat row list (~264 rows) on every `rows` write.
How to repeat (ONLY `storage/`; inner ssh with `< /dev/null`):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8730 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_BACKFILL=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8730 HL_HOST=127.0.0.1 ./bin/hybriel project.hl
# session file for user 0mujqspn0aqb (mission 012 (old 056) entry: name sha256(id), created/seen = now ms), id → .scratch/real-sid
REAL_OUT=/tmp/w057/real-after node tests/realdata-057.mjs # REAL_SHOWS, REAL_SEASONS (indexes, default 1,3), REAL_REPEAT
# afterwards rm -rf .scratch/realdata .scratch/realdata-sessions .scratch/real-sid
```
**Gate** (208 → 214):
```
TRACKER_GATE_SHOTS=/tmp/w057/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
214 passed, 0 failed
```
Skeleton state tested with a FAKE DELAY: an init script holds the websocket frame naming `showSeasonEpisodes` back 1.5 s
while `window.__delayFace` is set. Shots `show-skeleton/open/closed-{390,1280}.png` looked at (Loreana /tmp/w057/shots/).
**Open**: (1) deploy. (2) Episode/season check clicks still use session faces that rebuild all rows (+ the re-mount sync) —
not measured here; the same split (client patch + a lean face) could speed them up. (3) Big-show skeleton ~80 ms (whole-list re-render in hl:web).
(4) The adult fallback is not in the gate (no fixture series is adult; changing seed-adult shifts many counts).
## 2026-10-01 — mission 012 (old 056): #16 merged into main, 3 fixes, Hybriel master 8590df63 (NOT deployed)
**Commits** (main, not pushed): 2c53d5e merge t16-person (worktree tracker-wt16 + branch removed; its `.scratch/w055` moved to
`.scratch/w055-wt16`), 2b7fdd6 header + backfill posters, f40c250 Hybriel + adult show page, + this docs commit.
- **Merge**: all conflicts additive. Filmography (people.hl `filmographyOf`) lists only `isPublicTitle` (`adult == false`)
titles whose credit is not adult — unknown hidden like the public lists.
- **Header** (signed out, phone): ident's "choose ident" + "Log in with ident" wrapped (73 px). Now nowrap, the phone says
"Log in" (`span.wide` " with ident" from 40rem): 46 px at 320/390, 54 px at 1280. TRAP: a style key twice in one block
REPLACES the first (a 2nd `'ident-selector::part(button)'` dropped its padding/border) — merge into one line.
- **Backfill**: an adult title's poster is not downloaded (flag + ids stored). Real data: posters 5,433 (was 8,740), dir
269 MB (was 459 MB) → ~190 MB saved.
- **Hybriel** master 8590df63 (≥ 73267707: #122 18cf2fb9, #121 426944c7), sha256 `ca830ef8…`; old in `.scratch/pre-056/`.
Duplicate @ids on live copy: 0 in all 8 tables (episodes 232,384, shows 9,453, watches 11,663, …). Workarounds: none removed
(static helpers kept). SVG namespace check already in the gate (047).
- **Adult show page** (components/show.hl `visibleShow`/`adultHiddenFor`): `adult == true` → "Not found" for non-followers,
signed out too; no member carries the row (`showRow` itself is null — every member is shipped to the browser). A follower
sees it; Unfollow on the page hides it at once (session sync, #122; old build: only after reload). Faces showRows/
showToggle* refuse it. TRAP (why 054's try failed): `collapsed` derived from `found`/`showRow` (session-derived) gets
re-derived on the server with MOUNT state after each session face → seasons snap back, 2nd click dead. `collapsed`
now from the slug (`initialCollapsedOf`). Unknown (`null`) pages are not hidden.
- **Gate fixes**: the backfill "start" line comes with the clock's first tick, after the server answers → wait for it
(was a deterministic fail after the merge); every `stopServer()` first sends the tab to about:blank (websocket reconnect
errors failed the console check intermittently).
**Gate** (203 after merge → 208):
```
TRACKER_GATE_SHOTS=/tmp/w056/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
208 passed, 0 failed
```
Proven failing without the fix: poster check (forced download → FAIL), Unfollow-hides check on the old Hybriel (FAIL, page
stayed until reload). Shots `header-signed-out-{320,390,1280}.png`, `person*.png` looked at.
**Real data** (copy of Byrodin `storage/` 2026-10-01 09:47, :8730, real TMDB, backfill run on the copy 78 min, copy deleted):
```
adult backfill done: 9453/9455 titles, adult=3450 notAdult=5866 unknown=137 posters=5433 ids=6459 requests=9455 errors=139 retries=2 seconds=4701.056
```
`node tests/realdata-056.mjs` (full load ms 390/1280): `/` 426/91, `/movies` 92/90, `/movies/page/2` 93/37, `/shows` 287/41,
`/shows/page/2` 92/42, `/search/the` 336/126, `/search/sex` 442/91, `/person/Jared-Harris` 242/102 (first visit filled 97
credits, 80 added, 487 ms), `/shows/Foundation` 189/145; signed in (az5b2): `/my/shows` 650/219, `/my/movies` 41/38,
`/my/unwatched` 1478/1655 (repeats 1.5–2.4 s, server 1.33 s, 1,595 rows — the same with the old Hybriel), `/my/schedule`
457/322, show 170/267. Header one row on every page (46/54–56 px). Adult page `/shows/This-Aint-Star-Trek-XXX`: "Not found"
signed out (slug not in the HTML) and for az5b2 (follows no adult title). Every link of all 251 list pages, `/`, 6 searches
(the, sex, porn, xxx, girl, love) and the person: 5,914 slugs → `tools/check-public-slugs.hl`: notAdult=5914, adult=0,
unknown=0. 0 console problems. Screenshots `/tmp/w056/real/{out,in}-*-{390,1280}.png` on Loreana (looked at).
How to repeat (ports 8730/8731; ONLY `storage/`; inner ssh with `< /dev/null`):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8730 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_BACKFILL=1 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8730 ./bin/hybriel project.hl
# wait for 'adult backfill done' (~80 min); session file for user 0mujqspn0aqb as in ticket #8 (created/seen = now ms), id → .scratch/real-sid
REAL_PERSON=Jared-Harris REAL_SHOW=/shows/Foundation REAL_ADULT=This-Aint-Star-Trek-XXX node tests/realdata-056.mjs
# stop the server, then: TRACKER_STORAGE=… SLUGS=/tmp/w056/real/slugs.txt USER_ID=0mujqspn0aqb ./bin/hybriel tools/check-public-slugs.hl
# afterwards rm -rf .scratch/realdata .scratch/realdata-sessions .scratch/real-sid
```
**Open**: (1) deploy: the live backfill (~80 min, now ~270 MB posters). (2) `/my/unwatched` 1.5–2.4 s full load at the 2 s
target (1,595 rows). (3) hl:fetch blocks the server during a slow TMDB answer (max step 5 s in this run). (4) Unfollowing an
adult title on its page turns the page into "Not found" at once (wanted: hidden for non-followers). (5) 137 TMDB-404 titles
stay hidden; adult people are not hidden; people are never refreshed.
## 2026-10-01 — ticket #16 (mission 011 (old 055), branch t16-person, worktree tracker-wt16): person pages
**Built**: `/person/<slug>` (`components/person.hl`, `people.hl`, route + `/profiles/:name` in `project.hl`, styles block, search.hl
`titleIdByTmdb`): photo, facts, short bio, filmography (poster, title, year, type, role; newest first; adult hidden). Incomplete
person → "Loading filmography…" and a step-by-step TMDB fill without reload (1 TMDB request + photo, then 40 titles per face
call), missing titles added as minimal records with TMDB's `adult`, person marked `filmographyAt`. Details: README "What it does
(ticket #16)".
**Decisions**: the fill is driven by the page (face `personFill` per step; first step from an invisible CSS-animation tick — hl:web
has no mount hook), so it needs no server timer and other requests are served between steps. No sign-in needed. Credits stored in
the migrated `person.shows` (+ `posterPath`, `adult`); migrated credits TMDB no longer lists are kept. Rows without our poster file
show TMDB's w92 thumb (like search's web rows) — posters of added titles otherwise come later (sync/backfill). Crew credits count
(role = job). New titles get this person in `cast` (other existing titles' `cast` is not touched). Bio clamped to 4 lines, click =
all. Hidden: `show.adult == true` OR the credit's TMDB `adult`.
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~10:00, deleted afterwards; dev server :8777, signed out, `TRACKER_SYNC=0`,
real TMDB; `node tests/realdata-person.mjs`): cast click on the show page → person page shown / filled (ms), titles added, server log:
| person (show) | shown | filled | credits | added | rows | steps | TMDB request | other requests waited max |
|---|---|---|---|---|---|---|---|---|
| Tyler Hoechlin (Superman & Lois, 33 migrated credits) | 131 | 643 | 36 | 7 | 38 | 2 | 249 ms | 678 ms |
| Grant Gustin (The Flash) | 68 | 593 | 33 | 21 | 33 | 2 | 242 ms | 203 ms |
| Jared Harris (Foundation) | 85 | 869 | 97 | 80 | 97 | 4 | 287 ms | 220 ms |
| Niamh Algar (Raised by Wolves) | 76 | 584 | 39 | 20 | 40 | 2 | 250 ms | 209 ms |
| Eric Roberts (stress, direct load) | — | 3.8 s fill | 758 | 712 | 758 | 20 | 646 ms | 612 ms |
Second visit (complete, full load in Chrome): 41–120 ms at 390/1280, no loading, same rows; row → show page client-side 122–149 ms;
0 console problems, no sideways scroll at 390. Screenshots `/tmp/w055/real/*.png` on Loreana (looked at).
How to repeat (ports 8777–8779; the fill WRITES — use a copy):
```
mkdir -p .scratch/realdata && ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' | tar xzf - -C .scratch/realdata
TRACKER_PORT=8777 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8777 ./bin/hybriel project.hl
node tests/realdata-person.mjs # REAL_CASES='/shows/X:Person-Slug,…' for others; grep 'person fill' in the server log
# afterwards: stop the server, rm -rf .scratch/realdata .scratch/realdata-sessions
```
**Gate** (167 → 186 checks):
```
TRACKER_GATE_SHOTS=/tmp/w055/shots TRACKER_GATE_PORT=8770 TRACKER_GATE_IDENT_PORT=8771 TRACKER_GATE_CHROME=8772-8775 TRACKER_GATE_TMDB_PORT=8776 node tests/browser.mjs
186 passed, 0 failed
```
Screenshots `/tmp/w055/shots/person{,-complete}-{390,1280}.png` (looked at).
**Merge notes**: shared files touched additively — `project.hl` (1 import, `profilesDir`/`profileRoute` before the "old addresses"
block, 2 routes after `/search/:q?`), `search.hl` (`titleIdByTmdb` before "the query as it comes in the address"), `styles.hl` (one
block before `@media` + 2 lines at the end of it), `tests/faketmdb.mjs` (`persons`, `personRequests`, `personDelayMs`), `tests/seed-show.hl`
(Test Actor gets `tmdbId = 3002`, genre Drama `tmdbId = 18`, two persons), `tests/browser.mjs` (`gatePersons` fixture, a block
before "#10 THE INSTALLABLE APP"), README/STATUS. Mission 010 (old 054) (`adult`): people.hl already reads `show.adult == true` and writes `adult`
on added titles; if 054 also filters the public lists, the gate's "Hidden Adult Film" (2012) leaves /movies — no gate check here counts it.
**Open**: the TMDB step blocks the server for its request (~0.25–0.65 s; hl:fetch is blocking). People are never refreshed after
complete (the ticket's "daily sync may refresh followed people later" — not built). Adult PEOPLE are not hidden (only titles). Added
titles have no poster file / seasons until synced. Signed-out header at 390 wraps "choose ident" / "Log in with ident" to two lines
(not from #16 — seen in the real-data shots, every page).
## 2026-10-01 — mission 010 (old 054): adult titles hidden from the public lists + the adult-flag backfill (NOT deployed)
**Built**: `shows.adult` true/false/null(unknown). Public lists (catalog.hl: homepage rows, `/movies`, `/shows`) and the
search (search.hl `titleOk`; "Fetch from web": `include_adult=false` + only results with `adult: false`) show ONLY
`adult == false` (shows.hl `isPublicTitle`) — unknown is hidden, so the first deploy is safe. `/my/*` and the show page
are not filtered (a followed adult title stays reachable). The daily sync, the search import and the backfill store
TMDB's flag. **Backfill** in the app (project.hl `backfillTick` → tmdbsync.hl `backfillTitle`): one title per step on a
0.1 s clock shared with the daily sync (which wins), `GET /movie|tv/<id>?append_to_response=external_ids` + the poster
when missing, 260 ms pause per API request; 404 → `adultCheck='TMDB 404'` (hidden, never asked again); no answer/429/5xx
→ retried (≤ 2 per start); resumes after a restart (queue = what is still unknown); log every 500 + done line. Env
`TRACKER_BACKFILL=0` off, `TRACKER_BACKFILL_LOG_EVERY`. Details: README "What it does (mission 010 (old 054))".
**Decision / Hybriel finding**: hiding the SHOW PAGE from non-followers was built and dropped. `found = showRow != null &&
!adultHiddenFor(showRow, me)` (a static reaching mpackdb, reading `me`) → after any session face on the page
(`showToggleEpisode`) the browser re-derives `found` and then `rows`, whose `buildRows` reaches mpackdb → console
`'watchesOfUser' is a static of 'Watches' whose initializer reaches 'watchesTable', a name the realm 'client' … cannot bind`,
the click does nothing (gate: "ep3 solid" timeout). A plain copy in between (`hideAdult = hiddenAdult`) did not stop it.
So the show page of an adult title opens for anyone who has the link (it is in no list and no search).
**Gate** (167 → 184 checks):
```
TRACKER_GATE_SHOTS=/tmp/w054/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
(backfill second start: 8 titles left, 6.614 s; 285 pages asked during it, slowest 634 ms; adult backfill done: 8/8 titles, adult=0 notAdult=7 unknown=1 posters=7 ids=2 requests=8 errors=1 retries=0 seconds=6.714)
184 passed, 0 failed
```
New block (end of the gate, before the PWA part): `tests/seed-adult.hl` (app stopped) = 1 adult movie alice follows +
14 unknown (fake TMDB: 1 adult, 1 404). Backfill off: none on `/`, `/movies` (2 pages), `/shows`, search; the adult one on
alice's `/my/movies` and its page; web search drops TMDB's adult result. Backfill on with 300 ms per fake answer: stopped
after 6 → restart asks only the other 8 (resume; mpackdb `update` survives SIGTERM without `persist`), pages < 1.2 s
meanwhile (slowest 634 ms), lists + search updated at once, poster served, 3rd start "nothing to do". All other fixtures
are `adult = false` (seed-show.hl); `startServer()` runs with `TRACKER_BACKFILL=0` unless a block switches it on; on a
throw the gate now prints the browser console + server log tail. Screenshot `adult-home-1280.png` (looked at).
**Real data** (copy of Byrodin `storage/` 2026-10-01 07:37, real TMDB, :8730, `TRACKER_SYNC=0`, deleted afterwards):
```
adult backfill: start, 9453 titles without an adult flag (65 ms to find them)
adult backfill done: 9453/9454 titles, adult=3450 notAdult=5866 unknown=137 posters=8740 ids=6459 requests=9454 errors=140 retries=1 seconds=5935.021
```
- 98.9 min, ~0.63 s per title. Adult 3,450 (3,448 movies, 2 series), not adult 5,866, unknown 137 (all TMDB 404 →
`adultCheck`), errors 140 = 137 × 404 + 2 poster 404 + 1 TLS "no answer" (retried OK). Posters: 126 → 8,866 files,
459 MB (adult titles' posters included). A restart afterwards: `start, 0 titles` / `nothing to do`.
- Lists after: `/movies` 4,242 movies (177 pages), `/shows` 1,624 series (68 pages). EVERY slug on `/`, all 245 list
pages and 10 searches (the, sex, porn, slut, milf, girl, xxx, anal, teen, hot): 6,008 links → all `adult == false`
(0 others). TMDB-non-adult titles like "Porn King: The Rise & Fall of Ron Jeremy" (2022 doc) and "Slut." (2004) stay.
- Page latency DURING the run (curl every ~5 s, 1,155 samples each): `/` avg 0.156 s max 4.87 s, `/movies` avg 0.059
max 1.67, `/shows` avg 0.045 max 1.54, `/search/the` avg 0.075 max 1.33; 14 of 4,620 > 1 s. Cause: 13 steps > 1.5 s
(one 5.8 s) — all a slow TMDB/image answer (movies, no catalog work); hl:fetch blocks the one request loop.
- After: full load (Chrome, signed out) 390/1280: `/` 0.57 (cold)/0.09 s, `/movies` 0.07/0.04, `/shows` 0.04/0.04,
`/search/the` 0.46/0.08 s, 0 console problems. Screenshots `/tmp/w054/real/{home,movies,shows,search-the}{,-bottom}-{390,1280}.png`
on Loreana — looked at: no adult title or poster. Logs `/tmp/w054/real-backfill.log`, `real-latency.log`, `real-restart.log`.
How to repeat (ports 8730/8731; ONLY `storage/`; note the inner ssh needs `< /dev/null` inside `ssh loreana bash -s`):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' < /dev/null | tar xzf - -C .scratch/realdata
TRACKER_PORT=8730 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realsess/ TRACKER_URL=http://127.0.0.1:8730 ./bin/hybriel project.hl
# progress: grep 'adult backfill' in its log; afterwards stop it, rm -rf the copy + sessions
```
**Open**: (1) deploy: the live app starts the backfill at once (~100 min, ~450 MB of posters into `storage/` → the
nightly backup); until it reaches a title the lists are emptier (all unknown at the first start). (2) A slow TMDB answer
blocks pages for its duration (max seen 4.9 s) — needs a non-blocking fetch in Hybriel or shorter timeouts (20 s now).
(3) Posters of adult titles are downloaded too (served only by their file name; could be skipped to save ~170 MB).
(4) The show page of an adult title is not hidden (Hybriel limit above). (5) 137 titles TMDB no longer has stay hidden.
(6) Seen in passing: signed out at 390 px the header buttons wrap to two lines (pre-existing).
## 2026-10-01 — mission 009 (old 053): #15 (TVmaze) and #14 (search) merged into main (with #13)
**Merges** (on main, never pushed): `git merge t15-tvmaze` → `03edc78`, then `git merge t14-search` → `dcc5eec`.
Branches t14-search / t15-tvmaze and worktrees tracker-wt14 / tracker-wt15 removed afterwards.
**Conflicts** (both sides kept): README (intro steps 13/14/15, "What it does" sections, Test block, Files rows merged),
STATUS (all entries, newest on top), `tests/realdata-check.mjs` (#13's page list + #15's env REAL_PORT/REAL_CHROME/
REAL_OUT/REAL_SHOW; REAL_SHOTS still works), `project.hl` (imports of both), `styles.hl` (both blocks),
`tests/browser.mjs` (fixtures: movie 1005 + TVmaze episode lists + thetvdb 4004→504; screenshot list = all pages).
**Behaviour fixes from combining**: (1) #13's 32 extra fixture titles were named "Gate …" and flooded #14's "gate"
search checks → renamed "Pager Movie NN / Pager Future Movie / Pager Undated Movie / Pager Older Series"
(`tests/seed-show.hl`). (2) `search.hl` `importTitle` now calls `catalog.hl` `refreshCatalogShow` → an imported title is
on /movies, /shows and the home rows at once (+1 gate check; the offline check's home Shows row is 4 tiles now).
**Gate**:
```
TRACKER_GATE_SHOTS=/tmp/w053/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
167 passed, 0 failed
```
**Real data** (copy of Byrodin `storage/` 2026-10-01, :8730, `az5b2`, `TRACKER_SYNC=0`, deleted afterwards; procedure =
ticket #8; session file `created`/`seen` = now ms): search index 0.58 s at boot. Server (curl): `/` 6 ms, `/movies` 11,
`/shows` 8, `/search/doctor%20who` 14, `/search/the` 56, `/my/shows` 178, `/my/unwatched` 1204, `/my/schedule` 252,
`/my/movies` 6 ms. Full load in Chrome (`REAL_OUT=/tmp/w053/real node tests/realdata-check.mjs` + a search timing),
390 / 1280: `/` 1.32 (cold) / 0.08 s, `/movies` 0.11 / 0.04, `/shows` 0.10 / 0.09, `/search/doctor%20who` 1.04 (cold) /
0.06, `/search/the` 0.12 / 0.66, `/my/shows` 1.27 / 0.26, `/my/unwatched` 1.39–2.0 / 1.37–1.39, `/my/schedule` 0.33 /
0.27, show page 0.11 / 0.11 s; 0 console problems, no sideways scroll. Screenshots `/tmp/w053/real/` + gate shots
`/tmp/w053/shots/` on Loreana (looked at).
**Open**: `/my/unwatched` (1594 rows) is still the slowest page (server 1.2 s); not deployed; open points of #13/#14/#15
below stay (adult titles, imported titles not followed → no daily sync, …).
## 2026-10-01 — ticket #14 (mission 008 (old 052), branch t14-search, worktree tracker-wt14): search
**Built**: header magnifier → `/search` / `/search/<text>` (`components/search.hl`), `search.hl` (in-memory word-prefix
index over 9453 titles + 15157 people, built at boot; TMDB `search/multi`; `importTitle` = new show record + the daily
sync's `syncShow`). Signed out: search + web work, "Add" opens the sign-in modal. Details: README "What it does (ticket #14)".
**Decisions**: address `/search/<text>` instead of `/search?q=` (hl:web gives a page no query string; `/search?q=x` opens
the empty page). Min 2 letters. ≤ 30 titles + ≤ 10 people. One TMDB request per "Fetch from web" (button, not per
keystroke). Web list leaves out what we have by TMDB id + kind. Import is not auto-follow (so the daily sync won't
refresh it later — architect's call). Genres matched to ours by name; cast not imported. People link to `/person/<slug>`
(no page yet, like the show page's cast).
**Measured on Hybriel**: a module `static x = f()` is evaluated LAZILY (never, if nothing reads `x`) — the index is built
by `ensureIndex()` called from a `project.hl` member. Static hashes/lists in a module can be mutated in place
(`buckets[k].push(i)` works, 200k pushes 174 ms). The server percent-decodes the path BEFORE routing (`%2F` → 404, invalid
UTF-8 passes through raw); a client-side navigation hands the param still encoded. A record literal entry `id = …`
shadows a local `id` for later entries (bit me: `tmdbId = id` took the new 16-hex id).
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~07:00, deleted afterwards; dev server :8770, az5b2, real TMDB):
index build 0.60 s at boot (log `search index: 9453 titles, 15157 people, 4180 keys, 612 ms`); `dbSearch` 0–1 ms for
"doctor who"/"star trek"/"sopranos", 38–44 ms for "the"/"th" (2000+ candidates); server render `/search/doctor%20who`
9 ms, `/search/the` 56 ms; browser: per keystroke → results + address 56–76 ms (incl. CDP), `/search` cold 0.85 s / warm
0.04 s; Fetch from web "pluribus" 0.12–0.23 s (real TMDB); Add "Pluribus (2025)" tv 225171 → `/shows/Pluribus` in
0.77–1.1 s: Season 1 9 episodes, poster, TMDB/IMDb/TVDB/TVmaze, genres Drama + Sci-Fi & Fantasy; afterwards found in our
database and gone from the web list; 0 console problems. Screenshots `/tmp/w052-t14/real/*.png` on Loreana (looked at).
How to repeat (ports 8770–8772; ONLY `storage/` is copied):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' | tar xzf - -C .scratch/realdata
cp -r .scratch/realdata/storage .scratch/probe-store # the import WRITES — work on a second copy
TRACKER_PORT=8770 TRACKER_SYNC=0 TRACKER_WATCH=0 TRACKER_STORAGE=$PWD/.scratch/probe-store/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ TRACKER_URL=http://127.0.0.1:8770 ./bin/hybriel project.hl
# AFTER the server is up (it sweeps old session files at boot): session file name = sha256(<32-hex id>), JSON
# {"created":<now ms>,"data":{"tag":"t"},"id":"<id>","seen":<now ms>,"user":{"id":"0mujqspn0aqb"}}; the id into .scratch/real-sid
TRACKER_REAL_WEB=pluribus node tests/realdata-search.mjs # TRACKER_REAL_IMPORT=0: no import
# afterwards: stop the server, rm -rf .scratch/realdata .scratch/probe-store .scratch/realdata-sessions .scratch/real-sid
```
**Gate** (111 → 136 checks):
```
TRACKER_GATE_SHOTS=/tmp/w052-t14/shots TRACKER_GATE_PORT=8773 TRACKER_GATE_IDENT_PORT=8774 TRACKER_GATE_CHROME=8776-8778 TRACKER_GATE_TMDB_PORT=8779 node tests/browser.mjs
(in-app run 5.413 s; 12 pages asked during it, slowest 1301 ms)
136 passed, 0 failed
```
Fake TMDB: `/3/search/multi` (→ `fake.searchRequests`, w92 thumbs → `fake.thumbRequests`, both kept out of
`fake.requests` so the sync checks are untouched); tv 1004 + movie 1005 importable, tv 1006 in the search but 404 (the
failed import = the hybriel#121/#122 check), TVmaze thetvdb 4004 → 504. Screenshots `/tmp/w052-t14/shots/search*.png`
(+ the existing ones; header still one row at 390 with the magnifier) — looked at.
**Merge notes** (#13 likely touches the same): `project.hl` (imports, one route, `searchIndexReady` before `server`),
`components/main.hl` (the magnifier inside `userBox`), `styles.hl` (one block before the `@media` block),
`tests/browser.mjs` (fixtures at the top, a signed-out block before the "#3 borders" checks, `search` in the 390/1280 page
list, a signed-in block before "#10 THE INSTALLABLE APP"), `tests/faketmdb.mjs`, README (intro, a section, Test, Files).
**Open**: no person page (`/person/<slug>` 404s, like the cast links). Imported titles aren't followed → not re-synced.
Web search is open to signed-out visitors (one TMDB request per click; no rate limit of ours). The field's autofocus is proven
on a full load only (real-data run: `focused: searchinput`); after the client-side click on the magnifier it is not checked,
and not on a real phone.
## 2026-10-01 — ticket #15 (mission 008 (old 052), worktree tracker-wt15, branch t15-tvmaze): TVmaze merge in the sync
**Built**: `tmdbsync.hl` `mergeTvmaze` (after the TMDB step, series with a tvmzId): `GET /shows/<tvmzId>/episodes`, by
S/E number: ADD missing episodes/seasons, FILL empty/placeholder titles + empty air dates; never overwrites a TMDB
value; TMDB's own placeholder ("Episode 4") no longer overwrites a real title; TMDB later adopts TVmaze rows by
number. `numberingAgrees` skips shows TVmaze numbers differently (≥ 80 % of shared S/E within 1 day or same title).
Details: README "What it does (step 15)". Tool line + totals: `tvmazeEpisodes tvmazeSeasons tvmazeFilled tvmazeSkipped`.
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~07:00, deleted afterwards; procedure = ticket #8 below, server
on :8739, `REAL_PORT=8739 REAL_CHROME=8738 REAL_OUT=/tmp/w052-tvmaze/real REAL_SHOW=/shows/The-Lord-of-the-Rings-The-Rings-of-Power node tests/realdata-check.mjs`):
```
run 1: shows=126 newEpisodes=0 updatedEpisodes=0 requests=255 tvmazeRequests=126 errors=0 tvmazeEpisodes=0 tvmazeFilled=21 tvmazeSkipped=7 seconds=178.9
run 2: updatedEpisodes=8 (Greek Salad "Épisode n" → "Episode n", after "Épisode" became a placeholder) tvmazeFilled=0 tvmazeSkipped=6 seconds=174.9
run 3: all 0, tvmazeSkipped=6, seconds=169.9 (idempotent)
```
Filled: Rings of Power S3E1–8 real titles (TMDB "Episode n", airs 2026-11), Last Week Tonight 8x26–30. No episode
was missing on TMDB that day (tvmazeEpisodes=0) — the add path is proven by the gate only. Skipped (checked by hand,
correct): Daily Show (TVmaze seasons by year), Star Trek: Prodigy + Blood & Treasure (shifted numbering), Alex Rider,
Intergalactic, The Rising (other premiere dates + placeholder titles). Loads (390 / 1280): /my/shows 1.51 / 0.26 s,
/my/unwatched 1.46 / 1.41 s, /my/schedule 0.36 / 0.29 s, Rings of Power 0.13 / 0.13 s direct, 0 console problems.
Screenshots `/tmp/w052-tvmaze/real/*.png` on Loreana (looked at), logs `sync{1,2,3}.log`, `skipped*.txt`, `filled.txt`.
**Gate** (111 → 117 checks):
```
TRACKER_GATE_SHOTS=/tmp/w052-tvmaze/shots TRACKER_GATE_PORT=8732 TRACKER_GATE_IDENT_PORT=8733 TRACKER_GATE_CHROME=8734-8737 TRACKER_GATE_TMDB_PORT=8738 node tests/browser.mjs
117 passed, 0 failed
```
Screenshots `tvmaze-{schedule,show}-{390,1280}.png` looked at.
**Open**: adds a new TVmaze-only episode on real data not yet seen live; the daily run takes ~70 s longer (1 TVmaze
request/series). Shows skipped by the numbering check get no TVmaze help (by design).
## 2026-10-01 — ticket #13 (mission 007 (old 050)): homepage, /movies, /shows, /my/movies
**Built**: `/` (`components/home.hl`): signed in 4 icon tiles → `/my/unwatched` `/my/schedule` `/my/shows` `/my/movies`;
for everyone the text, "Movies" (→ `/movies`, 10 latest RELEASED movies) and "Shows" (→ `/shows`, 10 series by newest
released episode) as poster rows (phone: sideways scroll inside the row; ≥ 40rem grid of 5). `/movies`, `/shows`
(`components/movies.hl`, `allshows.hl` → shared `components/tilelist.hl`): all, 24 per page, pagination =
components.hybriel `pagination.hl` copied verbatim (4b4dabe). `/my/movies` (`components/mymovies.hl`); `/my/shows` is
series only now. `catalog.hl`: lists built once at start, `refreshCatalogShow` after each synced show (`project.hl`
`syncTick`), day roll on first read of a new day. `follows.hl` `insertByFollowDesc` (moved from myshows.hl),
`shows.hl` `allShows`/`allEpisodes`.
**Decisions**: page in the PATH (`/movies/page/2`), not `?page=2` as the mission said — a component can't read the
query (hybriel#11) and Back drops it (hl:web `popstate` uses `location.pathname`), so `?page` would mean full reloads +
Back to page 1. Pagination click → host: the copied component has no data out (hybriel#87 ruling: a passed function),
so a `pager-box { on click }` around it reads the button text — no edit of the copy. One page → no pagination.
/movies = ALL movies incl. future-dated (first) and undated (last); the homepage row only released ones. Series without
a released episode come after the others, by their own `release`. Home text says episodes come from TMDB (TVmaze is
only used for ids — the mission's "TMDB/TVmaze" would be wrong). Sign-in hint only when signed out.
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~06:45, deleted; `az5b2`, :8730, procedure = ticket #8 below; the
session file needs `created`/`seen` = now in ms, `1` is swept at boot): boot → first answer 1.47 s (catalog build).
Server time (curl) 5–7 ms for `/`, `/movies`, `/movies/page/200`, `/shows`, `/shows/page/30`, `/my/movies`. Full load
in a FRESH Chrome per page (document + modules + socket + every poster in view), signed in / out:
| page | 390 in / out | 1280 in / out |
|---|---|---|
| `/` | 1.11 / 1.34 s | 1.13 / 1.29 s |
| `/movies` | 0.30 / 0.91 s | 0.57 / 0.57 s |
| `/movies/page/2` | 0.53 / 0.63 s | 1.31 / 0.50 s |
| `/shows` (24 real posters) | 1.07 / 1.36 s | 1.37 / 1.11 s |
| `/my/movies` | 0.80 / 0.66 s | 0.83 / 0.58 s |
Same Chrome (`node tests/realdata-check.mjs`): `/movies`, `/shows`, `/my/movies` 0.04–0.45 s; a page change by the
pagination 60–75 ms (client-side); Back works; no sideways scroll; 0 console problems. `refreshCatalogShow` for all
126 followed shows: 0.31 s (SNL 23 ms). Screenshots `/tmp/w050/real/*-{390,1280}.png` (looked at). Not mine but seen:
`/my/unwatched` 1.47–1.64 s (1594 rows), `/my/shows` 1.25 s cold at 390.
**Gate** (111 → 135 checks):
```
TRACKER_GATE_SHOTS=/tmp/w050/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
(in-app run 5.415 s; 11 pages asked during it, slowest 1433 ms)
135 passed, 0 failed
```
**Session-sync double rows** (team lead: since hybriel 64527baa every face taking `session` also syncs the members derived
from session; a handler that ALSO appends the returned row shows it twice): checked — no tracker handler appends, every
one assigns the whole list (`rows = r.rows`) or a flag; the new #13 pages have no faces. 4 gate checks "row appears
once" (exact count + unique button values, 400 ms after the face for a late sync): /my/unwatched click, show page
expand, episode click, Follow toggles.
Changed existing checks: the three "page stays empty" (now home content), the in-app run expects `newEpisodes=2` (the
fake TMDB adds a released episode to show 2 too, to prove the in-process re-sort). Screenshots `/tmp/w050/shots/`
(home, movies, movies-page-2, shows-list, my-movies at 390/1280 — looked at).
**Open**: (1) **adult titles** — the migrated movie table has many porn titles (e.g. on /movies page 1 and 4); they
are on the public homepage and /movies. No `adult` flag stored; the creator/architect decides (filter via TMDB `adult`,
hide unfollowed movies, …). (2) Movie posters: only followed shows/movies have real posters → nearly every movie tile is
the placeholder until the sync covers more. (3) If `?page=N` links are wanted anyway, that needs hybriel#11's query
half. (4) Real-phone check after deploy.
## 2026-10-01 — ticket #12 (mission 006 (old 049)): link icons to TMDB, IMDb, TheTVDB, TVmaze
**Built**: show page header badges `TMDB` `IMDb` `TVDB` `TVmaze` (`components/show.hl` `linkRows`, `shows.hl`
`externalLinksOf`, `styles.hl` `showLinks`/`a.extlink`: CSS-only monochrome badges, accent on hover), only for ids the
show has, `target=_blank rel=noopener`; a movie: TMDB (`/movie/<id>`) + IMDb. Phone: between title row and poster;
desktop: under the title. Sync (`tmdbsync.hl`): details request now `?append_to_response=external_ids` (no extra
request) → fills a MISSING imdbId/tvdbId; still no tvmzId → TVmaze `lookup/shows?imdb=`, on 404 `?thetvdb=` (550 ms
pause each, `pauseMsAfter` in both runners); never overwrites. Totals: `ids=`, `tvmazeRequests=`. New
`tools/count-external-ids.hl` (copy only).
**Bug found + fixed**: movies (`type = 'movie'`, 7814 rows, 1 followed: Star Trek: First Contact, tmdbId 199) were
synced as `/tv/199` — another title. Now `/movie/<id>?append_to_response=external_ids`: imdbId + poster only.
**Live data still carries the damage**: that record has `status = 'Ended'` and `tmdbSummary` = the overview of TV
show 199 ("Kabuhayang Swak na Swak…") — not visible (the page shows `summary`), the new sync leaves both alone; a
one-off fix (set `tmdbSummary = null`, `status` back) is the architect's call.
**Decisions**: TheTVDB link = `https://thetvdb.com/dereferrer/series/<id>` (it and `?tab=series&id=` both 301 to the
slug page, checked 2026-10-01); IMDb `/title/<tt>/`; TVmaze `www.tvmaze.com/shows/<id>`. Badges with text, not
drawn pictograms (no logos copied; legible at 0.75rem). An imdb id is only taken as `tt`+digits, a tvdb/TVmaze id
only as a number > 0. A show TVmaze doesn't know is asked again each daily run (2 lookups/day on real data).
**Real data** (copy of Byrodin `storage/` 2026-10-01 ~06:20, deleted afterwards; `tools/count-external-ids.hl`):
| | shows | tmdbId | imdbId | tvdbId | tvmzId |
|---|---|---|---|---|---|
| followed series, before → after 1 sync | 125 | 125 | 125 | 125 | 123 → **124** |
| followed movie | 1 | 1 | 1 | – | – |
| all series (only followed are synced) | 1639 | 1639 | 907 | 910 | 845 → 846 |
| all movies | 7814 | 7814 | 379 | – | – |
Sync: `shows=126 newSeasons=0 newEpisodes=0 updatedEpisodes=0 updatedSeasons=0 posters=1 ids=1 requests=256
tvmazeRequests=4 errors=0 seconds=105.6` (poster = the movie's real one; id = Greek Salad tvmz 60613; Mapping the
World: TVmaze knows neither tt0395050 nor tvdb 109331). TMDB's external ids match ALL 125 stored imdb/tvdb ids (0
differ). Show pages (az5b2, :8730): Librarians 1.08 s cold / 0.13 s, Greek Salad 0.18 s, Star Trek: First Contact
0.13 s (TMDB movie/199 + IMDb, real poster now); no sideways scroll, 0 console problems. Screenshots
`/tmp/w049/real/*-{390,1280}.png` (looked at). Procedure = ticket #8 below (sync tool with `.env` sourced).
**Gate** (101 → 111 checks):
```
TRACKER_GATE_SHOTS=/tmp/w049/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
(in-app run 5.416 s; 12 pages asked during it, slowest 1323 ms)
111 passed, 0 failed
```
Fixture: show 1 imdb `tt0000001` only (TMDB says tt9999999 + tvdb 4001), show 2 tvdb `4002` only (TMDB: tt0000002 +
4999), movie `gate-test-movie` (tmdb 1003, followed by a 2nd user so alice's lists are unchanged → the sync now has
3 shows). Fake TVmaze: imdb tt0000001 → 501, thetvdb 4002 → 502. Screenshots `/tmp/w049/shots/` (show-*, synced-show
{,-390}, synced-movie{,-390} looked at). The new-tab check closes the tab at once (it would load the real site).
**Open**: real-phone tap test after deploy (new tab). Star Trek: First Contact's wrong `tmdbSummary`/`status`
(above). Unfollowed shows never get ids filled (sync covers followed only, by design).
## 2026-10-01 — mission 005 (old 047): re-vendor Hybriel (invisible checks), mobile first, show page additions
**Bug** (live, the creator: "i dont see the check marks"): after a CLIENT-SIDE navigation (and after expand/collapse
or a check click) the check `<svg>` was built in the XHTML namespace (circle width 0). Fixed by **re-vendoring Hybriel
master ff51cf46** (bin sha256 `55a724a0…c5`, README "Vendored Hybriel"; old bin/plugins in `.scratch/pre-047/`).
Proof the new check fails on the old build: `/tmp/w047/old-build-gate.log` (`ns: http://www.w3.org/1999/xhtml,
circleWidth: 0`), passes on the new one.
**Built**: `styles.hl` mobile first (phone base rules, one `@media (min-width: 40rem)` block; header one row at
390 px; /my/unwatched + /my/schedule rows two-line on a phone). Show page: phone order title (+ Follow) → full-width
poster → genres/plot/cast (desktop unchanged); season caret (down/right); episodes not indented; artists orange
`#ce9178`; unwatched check = accent outline circle + accent check; **Follow toggle** (face `showToggleFollow`,
`follows.hl` `isFollowing`/`setFollowed`: new record `at = now`, unfollow deletes); signed out checks + Follow visible,
a click opens the sign-in modal (`components/modal.hl`, copied verbatim from components.hybriel becbd59), nothing sent.
Session-sync workaround in show/myshows/unwatched/schedule: `rows` reads `me` (a plain copy of `userId`), and
`following` goes through a LOCAL static `followingOf` — an imported mpackdb helper called directly in a member
initializer came back false after a client-side navigation (gate caught it; not reduced to a minimal repro yet).
**Duplicate @ids** (hybriel#113) on a copy of live storage (2026-10-01 04:13), `tools/count-duplicate-ids.hl` + the
`*.id.txt` indexes: **0 in every table** (episodes 232384, follows 128, genres 27, persons 15157, seasons 6450, shows
9453, users 1, watches 11663 rows).
**Real data** (copy, `az5b2`, :8730, `node tests/realdata-check.mjs`, full load incl. socket + posters):
| page | 390 px | 1280 px |
|---|---|---|
| `/my/shows` (128) | 1.29 s (cold posters) | 0.24 s |
| `/my/unwatched` (1594) | 1.51 s | 1.46 s |
| `/my/schedule` (54) | 0.35 s | 0.32 s |
| `/shows/The-Librarians-The-Next-Chapter` direct / client-side | 0.14 / 0.19 s | 0.13 / 0.18 s |
Client-side nav, expand, collapse: 14/26/14 checks, 0 wrong namespace, 0 undrawn; header 46 px at 390, no sideways
scroll. Screenshots `/tmp/w047/real/*.png` (looked at). Procedure = ticket #8 below; the session id goes in
`.scratch/real-sid` (the script reads it); copy + sessions deleted afterwards.
**Gate** (76 → 101 checks):
```
TRACKER_GATE_SHOTS=/tmp/w047/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
101 passed, 0 failed
```
Screenshots `/tmp/w047/shots/` ({home,show,my-*}-{390,1280}.png, signed-out-modal-1280.png) — looked at.
Gate note: the follow test re-follows show 1, so the sync's request order between the two shows may flip (the check
accepts both orders) and `/my/shows` after the sync has show 1 on top.
**Open**: the phone poster at full width is tall (2:3 → ~550 px on a 390 phone) — as asked, the creator may want a cap.
The unwatched (outline + check) and watched (filled) icons differ only by fill — as asked. No navigation links to
the `/my/` pages yet. Real-phone check after deploy.
## 2026-10-01 — ticket #10: installable app (PWA) + icons
**Built**: `icons/` (`icon.svg` source → `icon-192.png`, `icon-512.png`, `apple-touch-icon.png` 180; `favicon.svg`,
`favicon.ico` 16/32/48 — rsvg-convert + magick, commands in README "What it does (step 10)"); `project.hl`:
`appIcons` (192/512 × any/maskable, like calendar), `appTouchIcon`, `appFavicon`, `appThemeColor = darker.value`,
`appBackgroundColor = dark.value` (tokens), `offline = [ Home ]`, file routes for the icons, `/favicon.ico` now a
real file (was `direct = ""`); `components/main.hl` offline note + `netProbe` tick; `styles.hl` `offlineNote`,
`netProbe`, `@keyframes tracker-net-tick`; gate +12 checks.
**Decisions**: icon = a TV (antennas) with a check mark on the screen = "watched shows", green `#4ec9b0` on
rgb(25,30,35), all inside the maskable safe zone so one PNG serves any + maskable. Theme colour = `darker` (the header
bar, so the phone's status bar blends in), background = `dark`. Offline keeps only the shell/`/`; data pages get
hl:web's "Unavailable offline". The offline note is driven by `navigator.onLine`, read on a 1 s CSS-animation tick
(`animationiteration`) — hl:web has no connection-state member and no mount hook (reported as a Hybriel gap).
**Gate** (64 → 76 checks):
```
TRACKER_GATE_SHOTS=/tmp/w045/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
76 passed, 0 failed
```
Looked at: `/tmp/w045/icon-48.png`, `icon-192.png`, `icon-512.png`, `maskable-circle.png`, `small-on-white.png`
(favicon 16/32/48), `/tmp/w045/shots/pwa-phone-{online,offline}.png` (390 px) on Loreana.
Test gotcha: CDP `setOffline` only cuts the TAB — the service worker's own fetches still reach the server, so the
gate also stops the server before the offline reload.
**Open**: offline, the header's `<ident-selector>` (script from ident's origin, not cached) shows "choose ident"
next to "Log out". No note when the device is online but the server is down. Real-phone install test (ticket's
**Test**) after deploy: open https://tracker.worldapi.org → "Add to home screen".
## 2026-10-01 — ticket #9: TMDB sync (new episodes + posters)
**Built**: `tmdbsync.hl` (`syncShow`: details + ≤ 20 seasons per request via `append_to_response`, add/update
seasons+episodes, poster w342), `tools/sync-tmdb.hl` (`--limit N`, paced, prints counts), the daily run in
`project.hl` (`syncTick` on an hl:time `every(1)` clock: at `TRACKER_SYNC_HOUR` UTC (4), once a day, ONE show
per tick, next show after n × 260 ms), `follows.hl` `allFollowedShowIds`, fake TMDB `tests/faketmdb.mjs`,
fixture shows got tmdbIds 1001/1002. Details: README "What it does (step 9)".
**Decisions**: match episode by tmdbId, else by S/E when the local one has no tmdbId (161 migrated); season by
number. Empty TMDB values never overwrite. Season 0 only if the show already has one (else specials flood
/my/unwatched); empty TMDB seasons not created. Own 16-hex ids for new rows (hybriel#113). Posters in
`storage/mpackdb/posters/` (the folder the existing `/posters/:name` route reads — inside `storage/`, so in the
backup) rather than a new `storage/posters/`. Shows synced = all followed (also Ended ones — updates happen).
Sync is off without `TMDB_READ_TOKEN` or with `TRACKER_SYNC=0`.
**Blocking**: `fetch()` blocks the whole hl:web process (measured: requests wait for the running step) — so the
run is cut into one-show steps. Real data: ~0.7 s per step average, SNL (53 seasons, 4 requests) 1.7 s; persist
of the 3 tables 36 ms.
**Real-data proof** (copy of Byrodin's `storage/` in `.scratch/realdata/`, deleted afterwards; procedure as
ticket #8 below; session file needs `created`/`seen` = now in ms, else the boot sweep deletes it):
```
set -a; . ./.env; set +a
TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb ./bin/hybriel tools/sync-tmdb.hl --limit 128
tmdb sync done: shows=126 newSeasons=20 newEpisodes=800 updatedEpisodes=2651 updatedSeasons=174 posters=125 requests=380 errors=0 seconds=185.718
(again) shows=126 newSeasons=0 newEpisodes=0 updatedEpisodes=0 updatedSeasons=0 posters=0 requests=255 errors=0 seconds=89.888
```
Fresh-process recount: seasons 6430 → 6450, episodes 231584 → 232384 (exactly +20/+800, no row loss), 0
dangling season/episode refs, 0 duplicate season numbers; 12 duplicate tmdbIds + 1 wrong parent were already in
the migrated data (same before/after). 126 shows (128 follows), 125 posters (one show has none on TMDB).
In-app daily run on the copy (TRACKER_SYNC_HOUR = current hour, real TMDB): 126 shows, 255 requests, 136 s,
0 errors; `/my/shows` (signed in) polled every 0.5 s meanwhile: 181 answers, avg 0.25 s, max 4.0 s (one), then
1.8/1.7/1.4 s. Full page loads after the sync (headless Chrome, load + socket + all images):
| page | full load | rows |
|---|---|---|
| `/my/schedule` | 0.7–0.8 s | 55 (was 0) |
| `/my/shows` | 1.3–1.55 s (128 real posters now, 1 placeholder) | 128 |
| `/my/unwatched` | 1.5 s (server 1.3 s) | 1593 (was 831) |
| `/shows/The-Librarians-The-Next-Chapter` | 0.33 s | poster + new S2 (12 eps) |
Screenshots + logs: `/tmp/w044-real/` on Loreana (`after-*.png`, `real-sync{1,2}.log`, `real-lat.txt`).
**Gate** (49 → 64 checks):
```
TRACKER_GATE_SHOTS=/tmp/w044-shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs
(in-app run 4.405 s; 10 pages asked during it, slowest 1371 ms)
64 passed, 0 failed
```
Screenshots `/tmp/w044-shots/` (`synced-*.png` looked at: posters, S3, schedule rows fine).
**Deploy (architect)**: `TMDB_READ_TOKEN` must be in Byrodin's `.env` (env_file) — without it the log says
`tmdb sync: off (no TMDB_READ_TOKEN)`. First daily run 04:00 UTC after deploy; check `docker logs
tracker.worldapi.org | grep tmdb`. Or stop the container and run the tool once for an immediate catch-up.
**Open**: a step blocks requests up to ~2 s (a non-blocking `fetchStart` + `done()` state machine would fix it).
`/my/unwatched` render grew with the rows (1.3 s server, 1593 rows) — watch it against the 2 s target. A season
watch record stays solid when the sync adds a new episode to that season (existing season-watch semantics).
Dev servers get the real token from `.env` automatically → start them with `TRACKER_SYNC=0`. Shows nobody
follows are never synced (by design).
## 2026-09-30 — ticket #8: /my/ pages, S01E01, years, speed
**Built**: `/my/unwatched`, `/my/schedule` (routes in `project.hl`); old `/unwatched`, `/schedule` →
**301** (function routes `unwatchedMoved`/`scheduleMoved`). Rows as `S01E02 · <title>` (was
`Ep 2 · …`), list titles with the year `Doctor Who (2005)`, show page `1 episode` singular.
New shared helpers in `shows.hl`: `pad2`, `episodeCode`, `showYear`, `titleWithYear`,
`posterUrlOf`; in `watches.hl`: `watchedSet` (used by `/my/unwatched` and the show page's
`buildRows`; `isWatched` stays for the single-target toggles).
**Decisions**: year = `year`, else `release.slice(0,4)`, else none (197 shows lack `year`, some
have `release = ''`). Year only in LISTS (`/my/*`), not in the show page's `h1`. The 301 drops any
query (the pages take none). A row whose poster FILE is missing gets `/posters/none` (one cached
URL) instead of `/posters/<oldId>.jpg`.
**Why /my/shows was ~25 s live though its data is 0.2 s**: hl:web handles ONE request at a time.
Measured on the copy: `/my/shows` requested 1 s into a `/unwatched` render took 13.7 s (the
`/unwatched` render itself 14.5 s). Every request (pages, 128 poster requests) of every visitor
queued behind a slow `/unwatched`. Plus 128 distinct poster URLs per load, all the placeholder.
**Real-data check** (copy of Byrodin's `storage/` only, `az5b2` = user `0mujqspn0aqb`, dev
server :8730, headless Chrome, navigate → load → websocket open, 1280×900):
| page | before | after |
|---|---|---|
| `/my/shows` (128 rows) | 1.2 s (server 0.19 s) | 0.6–0.8 s, 1 poster request |
| `/unwatched` → `/my/unwatched` (831 rows) | 14.8 s | 0.7–1.2 s (server 0.55–0.65 s) |
| `/schedule` → `/my/schedule` (0 rows, data not synced since 2025-12) | 0.36 s | 0.3 s |
| `/shows/The-Librarians-The-Next-Chapter` | 0.34 s | 0.29 s |
How to repeat (Loreana; never touch Byrodin's folder beyond the tar read):
```
ssh [email protected] 'tar czf - -C /CONTAINERS/projects/tracker.worldapi.org storage' | tar xzf - -C .scratch/realdata # ONLY storage/
TRACKER_PORT=8730 TRACKER_WATCH=0 TRACKER_STORAGE=$PWD/.scratch/realdata/storage/mpackdb TRACKER_SESSIONS=$PWD/.scratch/realdata-sessions/ ./bin/hybriel project.hl
# session file: name = sha256(<32-hex id>), JSON {"created":1,"data":{"tag":"t"},"id":"<id>","seen":1,"user":{"id":"0mujqspn0aqb"}}
curl -s -o /dev/null -w '%{time_total}\n' -b trackersid=<id> http://127.0.0.1:8730/my/unwatched
# Chrome: tests/cdp.mjs launchBrowser, Network.setCookie trackersid, goto + wait __hl.socket open
# afterwards: stop the server, rm -rf .scratch/realdata .scratch/realdata-sessions
```
Screenshots: `/tmp/w041-real/{before,after}-{shows,unwatched,schedule,show}.png` (Loreana).
**Gate** (45 → 49 checks: two 301s, the browser landing on `/my/unwatched`, "1 episode"; the
existing ones now expect `/my/…`, `S01E02 · Second`, `S02E01 · Season Two Opener`, titles with
year, `/posters/sh1.png` + `/posters/none`). Fixture: show 1 `image = 'png'` (gate writes a 1×1
`posters/sh1.png`), show 2 `year = null`, `release = '2025-03-01'`.
```
TRACKER_GATE_SHOTS=/tmp/w041-shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 node tests/browser.mjs
49 passed, 0 failed
```
Screenshots `/tmp/w041-shots/` (looked at: all fine).
**Open**: `/my/unwatched` rebuilds all rows per check click (~0.6 s, fine). Still no navigation
to the `/my/` pages. The single-request server means any future slow page blocks everyone —
keep every page well under 1 s. `/my/schedule` stays empty until the TMDB sync step.
## 2026-09-30 — ticket #7: /my/shows
Built `/my/shows` (`components/myshows.hl`, route in `project.hl`, styles in `styles.hl`): every
followed show, newest follow (`follows.at`, epoch ms) first — small poster (2.5rem, `/posters/…` +
placeholder like the show page), title → `/shows/<slug>`, last watched episode `SxxEyy`.
**Decisions**: "last episode" = highest (season, episode) among the user's EPISODE watches of that
show (ticket); a season watch alone doesn't count (the season check watches every episode anyway).
No watch → no episode element. A show without a poster name → `/posters/none` (the route's own
placeholder), so every row has a poster. Read-only page, no face (like `/schedule`).
**Reuse**: new `follows.hl` `followsOfUser` (whole records; `followedShowIds` now builds on it) and
`shows.hl` `episodeById` (one line); `showById`, `posterName`, `watchesOfUser` unchanged.
**Performance** (against a COPY of `storage/mpackdb/` in `.scratch/perf-store`, removed after; live
files md5-verified unchanged): the page touches only the user's follows and episode watches
(one fetch per watch), never all episodes of a followed show — az5b2's 128 follows → 128 rows (123
with an episode, e.g. `S36E11`, `S30E91`) built in **160–255 ms**. For comparison **`/unwatched`
takes ~15 s** on the same copy (805 rows): it fetches every episode of every followed show and
checks each against the user's 11,692 watches with a linear `isWatched` — not in this ticket's
scope, flagged below.
**Gate** (`tests/browser.mjs`, 38 → 45 checks; `tests/seed-show.hl` adds a 2nd show "Unwatched Gate
Show", no seasons/poster name, followed AFTER the first but stored after it; the first follow is now
dated a day back): signed out → sign-in message; signed in → newest follow first, title hrefs,
posters loaded (`naturalWidth > 0`), 40px wide, `S01E02` after /unwatched's click, later `S02E01`
(watched BEFORE season 1's bulk watch → proves highest, not most recent), none for the other show.
```
TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 node tests/browser.mjs
45 passed, 0 failed
```
**Open**: `/unwatched` ~15 s on real data (above) — a watched-id map instead of `isWatched` per
episode would fix it (own ticket). No navigation links to `/my/shows` yet (homepage stays empty).
## 2026-09-27 — ticket #6: /schedule
Built `/schedule` (`components/schedule.hl`): every episode of a show the signed-in user follows
that is NOT yet released (`e.release > today`), soonest first. Creator: "/schedule has a similar
list without check icons for the upcoming episodes of shows followed" — same row shape as
`/unwatched` (`components/unwatched.hl`) minus the check icon.
**Reuse, not copy**: imports the exact same `followedShowIds` (`follows.hl`), `showById` /
`seasonsOfShow` / `episodesOfSeason` (`shows.hl`) that `/unwatched` already uses — nothing new
added to either file. `watches.hl` is not imported at all: a row here is never watched/unwatched,
there is no check icon and no click, so the page needed no per-user watch state and no `on server`
face — the whole component is a single read-only server-side render (`rows` computed once from
`session`, same as every other reactive member here).
**Order**: soonest first, ascending by `release` (a plain `'YYYY-MM-DD'` string, ordering
operators confirmed working against this app's own vendored `bin/hybriel`, same as `/unwatched`'s
STATUS entry) — the ticket named this the architect's own reading ("the creator named the order
only for /unwatched"), so `insertByReleaseAsc` mirrors `/unwatched`'s `insertByReleaseDesc` with
the comparison flipped.
**`decided`**: "upcoming" = `release` strictly after today (not `>=`) — the natural complement of
`/unwatched`'s own `release <= today`, so every released-or-today episode is unambiguously on one
list, never both, never neither.
**Gate** (`tests/browser.mjs`, 35 → 38 checks): reused the same fixture (`tests/seed-show.hl`,
unchanged) — episode 3 already carried a far-future release (`2099-01-01`) for `/unwatched`'s own
"excludes the future one" check; `/schedule` now proves the mirror image with the same fixture:
signed out, no rows, the sign-in message; signed in, only that one future episode shows (the two
already-released ones excluded), no `button.checkicon` anywhere on the page, unaffected by any of
the watch toggles the other checks perform on the same fixture (run in any order — `/schedule`
never reads watch state).
```
node tests/browser.mjs
38 passed, 0 failed
ps -eo pid,args | grep [h]l-browser-tier
(nothing)
```
Updated `README.md` ("What it does (step 6)", Test, Files) and this file.
**Open**: nothing shown on the homepage still points to `/schedule` (same as `/unwatched` before
it — the homepage stays empty per ticket #1's decision).
## 2026-09-27 — ticket #5: /unwatched
Built `/unwatched`: every episode of a show the signed-in user follows that is already released
(`release` ≤ today, a plain `'YYYY-MM-DD'` string — migrated as-is by `tools/migrate.hl`, never
run through `dateOf`) and not yet watched, newest release first (creator: "/unwatched lists all
unwatched episodes of shows followed in release date desc order").
**Reuse, not copy** (the ticket's own instruction, re "the same check icon and watch logic as the
show page"): `watchClassOf` (watched → `'checkicon solid'`/`'checkicon regular'`) was a private
`static` in `components/show.hl`; moved it to `watches.hl` so both pages import the SAME
definition instead of each declaring their own — `show.hl` now imports it too, nothing duplicated.
`isWatched`/`setWatched` (already in `watches.hl`) needed no change. New `follows.hl`
(`followedShowIds`) and `shows.hl` `showById` (one line) round out the read-only data access this
page and `/schedule` (ticket #6) both need.
**String comparison** (`e.release <= today`, `row.release >= x.release` for the desc sort):
hybriel#2 ("no ordering operators on strings") is filed as still-open in this ticket's brief, but
its own history shows it was fixed and merged to master 2026-09-25 — confirmed directly against
THIS app's own vendored `bin/hybriel` (`'2020-09-03' <= '2026-09-27'` → `true`), so no numeric
date workaround was needed.
**A View text node cannot hold an inline expression** (`"Ep " + row.episodeNumber + " · " +
row.title` inside a tag body) — same rule as View attributes (components/show.hl's own comment:
"a View attribute must be a literal, a member or a field path"), just not documented for text
nodes anywhere yet; it silently rendered `` instead of
erroring. Fixed by resolving the label into a plain field (`row.episodeLabel`) once, server-side,
in `unwatchedRowsOf` — the same pattern `show.hl`'s `buildRows` already uses for every other
ternary/concatenation. Not filed as a new Hybriel gap since it's the exact same restriction
already known and worked around elsewhere in this app, just not yet hit for interpolated text.
**Empty state**: signed out (no user, so no follows to show) and signed in with nothing unwatched
both render a message; used `.muted` (colorTextMuted), not `.message` (colorDanger) — the existing
`.message` class is for real errors (the login-failed banner), and an empty list isn't one (a
`decided`, small).
**Gate** (`tests/browser.mjs`, 29 → 35 checks): `tests/seed-show.hl` now takes an optional
`TRACKER_SEED_IDENTITY` and, when given, seeds a fixed user + a follow of the fixture show for it
— the only way to give `/unwatched` a known followed show before the browser ever signs in, since
the real login's `ensureUser` matches an EXISTING `users.db` row by `identity` and reuses its id
rather than minting a new one. The gate learns alice's per-app identity with one throwaway
`ident.exchange` call (consuming one code) before writing the seed. Episode 1/2 of the fixture now
carry a past `release` (`2020-01-01`/`2020-06-15`), episode 3 a far-future one (`2099-01-01`) —
proving both the release-date filter and the desc order. The new checks run signed out (no rows,
the sign-in message), then signed in RIGHT AFTER LOGIN, BEFORE the existing show-page checks that
toggle these same fixture episodes watched (order matters — otherwise `/unwatched` would already
be empty by the time it's tested): both unwatched episodes list newest-first, excluding the
future one; a click on the newest (`Second`, ep 2) removes it from the list; a reload proves the
watch is server-side, leaving only `Pilot` (ep 1).
```
node tests/browser.mjs
35 passed, 0 failed
ps -eo pid,args | grep [h]l-browser-tier
(nothing)
```
Updated `README.md` ("What it does (step 5)", Test, Files) and this file.
**Open**: nothing shown on the homepage still points to `/unwatched` — no navigation exists yet
(the homepage stays empty per ticket #1's decision; a later step, from the creator, presumably
adds it). `/schedule` (ticket #6) is next and reuses `follows.hl`/`shows.hl`/`watches.hl` the same
way.
## 2026-09-27 — ticket #4: the show page, watch checks (took over a previous half-done session)
Continued from an earlier session that had built the whole page — header, seasons/episodes,
the click-handling redesign to a flat `rows` list rebuilt server-side (`components/show.hl`'s
own header comment explains why: a nested `for` over seasons > episodes risks hybriel#86) — but
left with the gate red: clicking a check never turned it solid, no console error, and the
session ran out of budget before finding why.
**The actual bug**: the show page's own pure helpers (`buildRows`, `genreRowsOf`, `castRowsOf`,
`watchClassOf`, `initialCollapsed`) were plain instance members (`buildRows = (show, …) => {…}`,
no `static`). That shape works fine as a member's own top-level initializer (`rows = found ?
buildRows(showRow, watchRows, collapsed) : []` — this is how the page renders correctly on first
load) but throws when the SAME function is called from inside an `on server` face
(`showToggleEpisode` etc.): the server log showed `'buildRows' is not callable — it holds null`.
Found by running `.scratch/debug-click.mjs` (left by the previous session) with the server's
stdout/stderr captured — the browser console itself stayed silent (the face's error never
reaches the client at all, a separate Hybriel gap, hybriel#92 already covers that half).
Confirmed the fix by grepping other worldapi apps for the same shape:
`calendar.worldapi.org/components/calendar.hl` has `static soonOf = (u) => { … upcoming(u, …) }`
(an mpackdb-backed helper), called from its `on server calLoad`/`calSettingsLoad` faces, and
calendar's own gate is green — so `static` is the proven-safe shape for a component's own helper
that both (a) touches an mpackdb-backed import and (b) needs calling from more than one place
(its own initializer AND a later face). Declared all five helpers in `show.hl` `static`; reran
the gate: 29/29 green, including the season-bulk-toggle and the reload-proves-server-side checks.
Filed **hybriel#115 was already open** (from the previous session, about the CLIENT bundle
omitting a case for such a helper) but this session's finding is the SAME root cause reaching
further than that ticket says: the plain HL interpreter on the SERVER also treats such a helper
as `null` outside its own initializer, not just the JS-compiled client bundle. Left as one
`issues` entry pointing at hybriel#115 rather than opening a near-duplicate ticket — the
repro/observed there already generalizes (a per-instance member function wrapping an
mpackdb-backed import is unreliable wherever it is called from, not just the browser).
Ran `node tests/browser.mjs`: 29 passed, 0 failed (was 20 passed / 1 failed — the click timeout
— before the `static` fix).
**Rehearsed `tools/relink-episode-seasons.hl` against a full COPY** of the real
`storage/mpackdb/` (`.scratch/relink-rehearsal/`, since removed) — the tool's own header claimed
this rehearsal already existed here; it didn't (a leftover claim from the previous session), so
this session actually ran it:
```
relink: episodes 231584, null season before 36194, linked 36194, still null (no matching season) 0
relink: episodes.db replaced with the relinked file — re-run to confirm idempotency (0 linked the second time)
```
Ran a second time on the same copy straight after — idempotency confirmed:
```
relink: episodes 231584, null season before 0, linked 0, still null (no matching season) 0
```
All 36,194 pre-existing null-season episodes (ticket #2's count) now link to a real season by
`(show, seasonNumber)`; none left over. Not run against the LIVE `storage/mpackdb/` — that needs
the tracker container stopped first (hl:mpackdb does not coordinate two processes on one storage
directory, hybriel#21), the architect's job per the ticket ("the architect runs it live").
Updated README.md ("What it does (step 4)", Test, Files) and this file.
## 2026-09-27 — ticket #2: old data migrated (took over a previous session's rewritten tool)
Continued from an earlier session that had rewritten `tools/migrate.hl` to avoid a real data-loss
bug (see below) but never run it even once, and had left the real `storage/mpackdb/` holding data
from the OLDER, buggy two-pass version of the tool (put a bare record, then `update()` it once the
other side of a circular reference existed). Took over: moved that buggy data aside
(`.scratch/pre-migrate-buggy-backup`, since removed), kept `storage/mpackdb/users.db` (step 1's
login table, untouched by the migration), and ran the already-fixed tool for the first time.
**The fix already on disk (kept as is):** `Show.seasons`/`Season.show` and `Show.cast`/
`Person.shows` are two-way references, so the id on one side must exist before the record on the
other side can be built. The old approach — `put()` a bare record, come back with `update()` once
the far side's id exists — loses rows: reopening a table in a fresh process after an `update()` had
touched it read back FEWER rows than were ever `put()` (seen directly: `persons.db`'s `count()`
dropped 15157 → 15152 after one round of `update()` — a real `hl:mpackdb` bug, reported to Hybriel,
not something an app is allowed to patch around). The fix assigns every new id itself (8 random
bytes, `hl:crypto`) in one pass over each export file BEFORE building any record, so by the time a
record is actually built every reference is already a known id — one `put()` per row, `update()`
never called.
**Run** (`tools/migrate.hl`, `TRACKER_OLD_DATA=.../old-tracker/mongo-export`,
`TRACKER_STORAGE=$PWD/storage/mpackdb`):
```
migrate: genres 27 new, 0 already there (old 27 -> new 27)
migrate: persons 15157 new, 0 already there (old 15157 -> new 15157), show refs skipped 0
migrate: shows 9453 new, 0 already there (old 9453 -> new 9453), cast refs skipped 0, genre refs skipped 0, season refs skipped 0
migrate: seasons 6430 new, 0 already there (old 6430 -> new 6430), show refs skipped 0, episode refs skipped 0
migrate: episodes 231584 new, 0 already there (old 231584 -> new 231584), show refs skipped 0, season refs left null (pre-existing) 1
migrate: follows 128 new, 0 already there (old 128 -> new 128), show refs skipped 0
migrate: watches 11692 new, 0 already there (old 11692 -> new 11692), target refs left null (pre-existing) 3
migrate: 0 failed
```
Every count equals the old export's own document count (`old-tracker/README-RECONSTRUCTED.md`:
Show 9453, Season 6430, Episode 231584, Person 15157, Genre 27, Follow 128, Watch 11692). Ran a
second time straight after (idempotency: the tool's `oldId` index finds each row again) — every
table printed `0 new`, all rows `already there`, counts unchanged, 0 failed: the id-loss bug does
not resurface across a real close-and-reopen.
**Proof beyond the tool's own counters** (`tools/verify.hl`, new): the ticket asks for a check that
does not just trust `migrate.hl`'s own "skipped" tallies. It runs against a plain filesystem COPY
of `storage/mpackdb/` (hl:mpackdb rewrites a file's data on open, so the live store is never opened
a second time) and independently re-derives, from the persisted records alone: every table's
`count()` against the export's own document count, and — by building id sets for every table and
walking every reference field (`Show.genres/cast/seasons`, `Season.show/episodes`,
`Episode.show/season`, `Person.shows`, `Follow.show/user`, `Watch.target/user`) — that every
non-null reference resolves to a real id. It also follows one show end to end (`Raised by Wolves`,
2 seasons, 18 episodes, 1 follow, 2 season-level watches — all cross-checked back to the show).
```
$ cp -r storage/mpackdb .scratch/verify-copy && TRACKER_VERIFY_COPY=$PWD/.scratch/verify-copy ./bin/hybriel tools/verify.hl
count ok genres: 27 (expected 27)
count ok persons: 15157 (expected 15157)
count ok shows: 9453 (expected 9453)
count ok seasons: 6430 (expected 6430)
count ok episodes: 231584 (expected 231584)
count ok follows: 128 (expected 128)
count ok watches: 11692 (expected 11692)
users: 1 (expected 1, the creator)
dangling references: 0 (episodes with a pre-existing null season: 36194, watches with a pre-existing null target: 3 — not dangling, the export already had no target)
END-TO-END show 105ad9c91b14d663 "Raised by Wolves" seasons=2
season 1 (be2965a5ab7165ca) episodes=10 show-back-ref-ok=true
season 2 (0fe4e53157ed0367) episodes=8 show-back-ref-ok=true
total episodes across seasons: 18
follows on this show: 1, season-level watches touching it: 2
VERIFY PASS
```
(36194 episodes with no season = 36193 that never had one in the old export, plus the 1 the tool's
own count already named as a pre-existing dangling `season` reference in the source data — neither
is a reference this migration broke; `dangling references: 0` covers exactly that.)
The single old user (`User.jsonl`, `[email protected]`, password not taken over) is this app's
user for ident short id `az5b2` (`storage/mpackdb/users.db`, the same table step 1's login uses) —
confirmed in the copy: one record, `identity=az5b2`.
**Two Hybriel bugs found while building this** (both already filed as issues on this ticket by an
earlier session, not re-filed): `hl:fs readFile` silently truncates at 10 MiB (worked around here
with `split`, no shell, no JS, see `tools/migrate.hl`'s header comment) and `hl:mpackdb` loses rows
across `update()` + reopen (the reason this tool never calls `update()`).
**Open**: nothing shown yet — this ticket is data-only, on purpose (the next step, from the
creator, is showing the data in the UI).
## 2026-09-27 — ticket #3: no border on the header or filled buttons
Design notes from the first test (architect): "the application-header should have no bottom
border" and "the buttons also no border except they are inverted". Two changes in
`styles.hl`:
- `applicationHeader`: dropped `borderBottom`.
- `ident-selector::part(button)` (the signed-out "choose ident" button): added
`border = 'none'` — the element's own CSS gave it a 1px border the same colour as its
background (`--_accent`), invisible but still a real border in the computed style; this is
the app-side restyling hook ident's README documents (`::part(button)`), not a change to
ident's vendored `selector.js`.
Left unchanged, already correct: the native `button` rule (filled, e.g. no other filled
buttons on this page yet) already has `border = '0'`; `button.quiet` ("Log out") and the
selector's dropdown `[part="identity"]` rows keep their border (transparent background —
inverted style); `a.button` ("Log in with ident") has no border rule and browsers give `<a>`
none by default. The selector's signed-in "logged in with ident" status pill
(`::part(status)`) is not one of the buttons the ticket names and isn't a button element
(a `<span>`) — left with its border.
**Gate** (`tests/browser.mjs`, now 16 checks): added computed-style checks — the header's
`borderBottomWidth` is `0px`; `#loginbutton`'s border is `0px`; the selector's
`[part~="button"]` (inside its shadow root) border is `0px`; `#logout`'s border is NOT `0px`
(still inverted-bordered) once signed in.
```
node tests/browser.mjs
16 passed, 0 failed
```
## 2026-09-27 — ticket #1 reopened: login redone to match calendar exactly (header selector, empty page)
The architect rejected the first build of ticket #1: the sign-in sat centered on the page
(its own "Sign in with ident" card, identity + sign-out shown in a `homeCard`) instead of
the header identity selector calendar/gitoria use, and there was no `<ident-selector>` at
all. Rebuilt the login by copying calendar.worldapi.org's own files (unchanged in shape, per
the architect's instruction "Copy calendar.worldapi.org's login unchanged"):
- **`users.hl`** (new): the ident exchange + a `usersTable` (`storage/mpackdb/users.db`,
`identity` → this app's user id), copied from calendar's `users.hl` with `TRACKER_KEY` /
`TRACKER_SECRET` / `TRACKER_URL` / `TRACKER_STORAGE` in place of calendar's names (kept
the names already in this app's `docker-compose.yml`/README/DECISIONS rather than
switching to calendar's generic `IDENT_API_KEY`/`IDENT_API_SECRET`).
- **`login.js`** (new): calendar's bridge between `<ident-selector>`'s `ident-login` DOM
event and the hidden `#identcode` input, copied verbatim (creator: "login.js is correct,
as thats for externals in general").
- **`components/main.hl`**: now the header carries `userBox` (`<ident-selector>` + "Log in
with ident" / "Log out"), the `trackerLogin`/`trackerLogOut` faces and the
`trackerSignedIn`/`trackerSignedOut` client push — calendar's shell renamed to this app's
event names.
- **`components/home.hl`**: now truly empty (`View { home {} }`) — no sign-in link, no
identity/sign-out duplicated on the page; that is entirely the header's job now.
- **`components/loginfailed.hl`** (new) + **`project.hl`**: the login routes are now
`/login/callback` (function route, the button's return AND the code exchange) and
`/login/failed`, copied from calendar's `project.hl` (`safePath`, `failed()`,
`loginCallback`) in place of the old `/login` + `/callback` + `/logout` routes; the old
`/logout` POST route is gone (logout is a face, like calendar's).
- **`styles.hl`**: added the header selector styles (`userBox`, `identSelector`,
`ident-selector::part(...)`, sticky header) from calendar's `styles.hl`; dropped the
now-unused `homeCard`/`accountBar`/`userName` rules from the rejected centered sign-in.
**Gate rewritten** (`tests/browser.mjs`, 12 checks): replaced the old gate (which drove
ident's full email-code UI through a vendored dev copy of ident at `.scratch/ident-dev`)
with calendar's own pattern — `tests/identkit.mjs` (copied from calendar unchanged) starts a
throwaway ident (a fresh copy of `/media/STORAGE/projects/ident.worldapi.org`'s code, no
`.env`/storage copied, codes to a mail sink, never the live ident), signs in over its REST
API and registers this app, then the gate proves the header in a real headless Chrome:
signed out → `#selector` + `#loginbutton` in the header, **`main` is empty** → sign in
(`/login/callback?ident_code=`, the same exchange the selector would trigger) → `#selector`
shows `class="in"`, `#logout` appears, **`main` still empty** → sign out → signed out again
→ a reload stays signed out. No console errors.
```
node tests/browser.mjs
12 passed, 0 failed
```
`./deploy.sh --dry-run --target .scratch/deploy-preview --url http://127.0.0.1:0/`: gate
passed as step [1/4], rsync preview held none of `storage/`, `.sessions/`, `.env`,
`.scratch/`, `server.*`, logs. Removed the stale `.scratch/ident-dev` (the old gate's vendored
ident copy) and `.scratch/browser-gate` (the old gate's storage) — unused by the new gate.
**Still open**: the app is registered in ident's dev copies only (this gate's own throwaway
ident). The LIVE ident registration (`TRACKER_KEY`/`TRACKER_SECRET` in `.env` next to
`docker-compose.yml`) is the architect's first-deploy step (README "Deploy"), not built here.
## 2026-09-27 — ticket #2: data migration — blocked twice on sandbox access, nothing done
Two workers in a row (`s-20260927T1029-f57b0f`, then this session) could not start: the
migration source the ticket names, `/media/STORAGE/projects/old-tracker/mongo-export/*.jsonl`
+ its `README-RECONSTRUCTED.md`, is not reachable from this worker's sandbox. Verified fresh
this session:
```
$ ls -la /media/STORAGE/projects/
antcolony-docs hybriel ident.worldapi.org tracker.worldapi.org # no old-tracker
$ mount | grep STORAGE
/dev/nvme1n1p6 on /media/STORAGE/projects/tracker.worldapi.org type btrfs (rw,...)
/dev/nvme1n1p6 on /media/STORAGE/projects/hybriel type btrfs (ro,...)
/dev/nvme1n1p6 on /media/STORAGE/projects/ident.worldapi.org type btrfs (ro,...)
/dev/nvme1n1p6 on /media/STORAGE/projects/antcolony-docs type btrfs (ro,...)
# old-tracker is not among the mounted subvolumes at all
```
So this is not a data problem (the export is "complete" per the ticket) — it's the worker
sandbox for this ticket that never got `old-tracker` mounted (read-only, like `hybriel` /
`ident.worldapi.org` / `antcolony-docs` above). Nothing was built or changed this session:
no `storage/`, no `tools/`, no id-mapping — writing the migration tool blind, guessing the
JSONL shape from the ticket text alone, risks silently wrong data and was avoided on purpose
(see the project's "Build it properly" rule). Filed as an antcolony issue so the next worker's
sandbox includes `old-tracker` before another attempt.
## 2026-09-27 — ticket #1: the empty shell, ident login
Built the app from scratch: vendored `bin/hybriel` + `plugins/` + `shared/tokens.hl` from
`ident.worldapi.org` (newest local build, hybriel master 837fe120, no local patch — see
README "Vendored Hybriel"). `project.hl` (routes `/login`, `/callback`, `/logout`, `/`),
`components/home.hl` (the empty homepage), `components/main.hl` (shell), `styles.hl`
(accent green `#4ec9b0`, per `antcolony/README.md` "look and style" / `CONCEPT.md`).
`docker-compose.yml` / `Dockerfile` / `deploy.sh` following ident's own house pattern
(port 45008, container `tracker.worldapi.org`).
**Login**: the ident login BUTTON flow (ident.worldapi.org README "How apps use ident"),
server-side exchange (`POST <ident>/api/exchange`), this app's own framework session
(`session.user = { identity }`, cookie `trackersid`). ident hands over only the identity's
public **short id** — no display name yet (ident#11, properties handover, is on hold) — so
"your name" on the homepage is `Signed in as <shortid>`, the same identifier every other
worldapi app would show; see `decided` in the report.
**Lesson (Hybriel)**: a face that mutates `session.user` (e.g. `signOut`) does NOT
automatically re-render the page — a component's reactive members (`signedIn`, `identity`,
…) are computed once at mount from the framework's read-only session snapshot; a client
handler that calls `emit server X()` must ALSO flip the affected members itself afterwards
(ident's own `components/home.hl` `on doSignOut(e)` does exactly this: `emit server
signOut()` then `signedIn = false` …). Missing that made the sign-out button silently do
nothing client-side (the ack came back `ok:true`, the server-side session really was
cleared — proven by a reload — but the DOM never updated) until copied.
**Gate** (`tests/browser.mjs`, 13 checks): a fresh copy of `ident.worldapi.org`'s code (no
`.env`, no `storage/`, no `.sessions/` — verified empty of secrets/data before first use)
runs as this gate's own ident, on its own storage and its own mail sink (no live ident, no
real mail). The gate registers this app in that dev ident once over the `/__hl/emit` API
(the same one-time step a person does by hand in ident's `/apps` page), then drives the
REAL login-button flow in a real headless Chrome: signed out → click "Sign in with ident" →
ident's email form → the mail-sink code → the first-login optional-names form (Skip) → the
one-identity choice (one click) → back on tracker, "Signed in as `<shortid>`" → "Sign out" →
signed out again → a reload stays signed out. No console errors. Ran twice for stability, 0
failures; no leftover Chrome/hybriel processes after either run.
```
node tests/browser.mjs
13 passed, 0 failed
```
`./deploy.sh --dry-run --target .scratch/deploy-preview --url http://127.0.0.1:0/` (a local
directory, per ident's own convention for testing deploy.sh without touching Byrodin): gate
passed, rsync preview held none of `storage/`, `.sessions/`, `.env`, `.scratch/`, `server.*`,
logs — confirmed by the same grep the script refuses on.
**Open**: the app is not registered with the LIVE ident yet — `TRACKER_KEY`/`TRACKER_SECRET`
are unset until the architect's first deploy does that (README "Deploy"); until then
`/login` on the live site answers a plain error page instead of redirecting (`/` itself
still renders). This is normal for a brand-new app, the same as ident's own SMTP `.env`
before its first deploy — not something this ticket's worker can set (no `.env` access,
and it is a LIVE ident registration).

Branches

mainmain branch

Latest commits

2667da05tracker#30 (mission 025): /my/ pages from slim cached title cards, episode rows and watch sets (after the jobs /my/series 1.8 s -> 0.06 s, /my/unwatched 4.1 -> 0.18 s); timeline page shows its name once; franchise widget under the poster/title; movies with TV leftovers (First Contact) go through the details repair; tools/count-tmdb-ids.hl; gates 327/0, 52/0, 32/0mre2026-10-02 16:07
3909810dantcolony#40: mission references in README/STATUS/docs point to the moved missionsmre2026-10-02 14:48
e63b1d28antcolony#40: history (LOG.md), worker briefs (missions/) and reports moved here from antcolony, numbered per project; old numbers in antcolony docs/mission-map.mdmre2026-10-02 14:45
1cda451dtracker: Hybriel master 190aa11d (#127 both shapes, GC correctness fc838894) — conductor adopts despite /my/series 2.4x after jobs (memory 8.0 → 1.7 GB boot); see reports/071mre2026-10-02 14:16
c1fa2f2etracker (mission 071): Hybriel master 190aa11d measured on the real copy vs the live binary 8590df63 — NOT adopted (after the first-start jobs /my/series 2.4x slower, /series 1.6x, RSS swings 7.4-12.2 GB; fresh it is flat at 1.7-2.2 GB and /my/unwatched faster), vendor stays 8efba065, candidate kept in .scratch/w071/vendor-190aa11d; tests/kinds.mjs: collection seed off (its TMDB request broke check 1 in 1 of 4 runs); tests/realdata-071.sh + realdata-071-bench.mjs + tools/realdata-071-table.py; README + STATUS (numbers, how to repeat); gates 325/0, 32/0, 50/0mre2026-10-02 14:15
8081350atracker docs (mission 070): README (summary, Config HL_GC_BYTES — kept at Hybriel's default, the 256 MiB setting is taken out of docker-compose.yml again: the jobs grew to 12+ GB with it too, see STATUS), Test (three gates), Deploy (first start ~50 min: kinds then seed, restart once after collections done, memory numbers), Vendored Hybriel 8efba065 + #48 audit, Files; STATUS mission 070 entry (merges, migrated counts, lambda audit, gates, RSS old vs new, how to repeat, open points); docs/kinds.md + docs/franchises.md job order; tests/realdata-070-*.sh, tools/count-migrated.hl, tools/ref-params.py, tools/lambda-audit.pymre2026-10-02 07:52
1ad19c8ctracker: re-vendor Hybriel master 8efba065 (#126 GC by bytes, #48 lambda parameters copy) (mission 070): bin/hybriel sha256 50361e95…, plugins core crypto data fetch fs http http1 mpackdb proc smtp time web; lambda audit: 13 lambdas change a passed record/list (11 through a local alias), no caller relies on it — unchanged; 319 read-only lambda parameters get & (no copy per call: /my/schedule 2.35 → 0.46 s, /my/unwatched 13.7 → 5.7 s on the real copy); /my/unwatched one merge sort instead of n² inserts; docker-compose HL_GC_BYTES=268435456; deploy.sh runs kinds.mjs + franchises.mjs too (default ports 8700–8710); tests/realdata-070.mjs; gates browser 325/0, kinds 32/0, franchises 50/0mre2026-10-02 06:15
46b21389tracker (mission 070, conductor): the sync never destroys migrated data — the one-time summary step MOVES a copied summary to migratedSummary (marker summaries-moved.txt) instead of clearing it; a migrated record's first title/genres/homepage/tagline TMDB replaces → migratedTitle/migratedGenres/migratedHomepage/migratedTagline, a migrated season's/episode's title/summary → migratedTitle/migratedSummary (set once); tests/peek-shows.hl prints them; gate 325/0mre2026-10-02 03:10
749019b0Merge t19 (tracker#19 franchises + timelines) into main (mission 070): conflicts README/STATUS/show.hl/project.hl/browser.mjs/faketmdb.mjs, both sides kept; franchise/timeline pages get #20's typed heading (Franchise | …, Timeline | …), their title links via titlePath (/movies|/series|/shows); the collection seed waits for repair, kinds and credits too; franchises.mjs URLs + 2 new checks; gates browser 323/0, kinds 32/0, franchises 50/0mre2026-10-02 03:05
daf49feaMerge t20 (tracker#20 typed headings + #21 series/shows split) into main (mission 070): conflicts README/STATUS/show.hl/project.hl/search.hl/components/search.hl/browser.mjs/faketmdb.mjs, both sides kept; clock order backfill → repair → kinds → credits; withDetailsFields stores tmdbType + kind; gates: browser.mjs URLs → /series|/movies, typed h1 selectors; kinds.mjs repair/credits off, fixture name = seed name; browser 323/0, kinds 32/0mre2026-10-02 02:59
20e09d89Merge t18 (tracker#18 delta sync) into main (mission 070): conflicts README/STATUS/show.hl/project.hl/browser.mjs, both sides kept; pageShowOf summary = summaryOfmre2026-10-02 02:49
f83571c3tracker#18 (mission 067): daily sync by change lists — TMDB /tv|movie/changes (since the stored day, paged) + TVmaze /updates/shows → only our changed titles (followed: full step, unfollowed: light step — changed seasons, no TVmaze), full walk on first run / gap > 14 days / failed list; show record refreshed (title, tmdbSummary, tagline, status, genres …; renamed titles re-indexed); summary = the creator's own text (page: summary > tmdbSummary > tvmazeSummary), one-time clear of copied summaries (9,647 on the live copy); gate 261, tests/realdata-018*.mjs, README + STATUSmre2026-10-02 00:53
1ed5457etracker#20 + #21 (mission 068): typed headings "<Type> | <name>" in type colours; TV titles split into Series (/series) and Shows (/shows) by TMDB type + Reality/Talk/News genres — kind stored by sync/import/adult backfill + new kind backfill (resumes), /movies/<slug>, /shows/<slug> of a series/movie → 301, /my/series + /my/shows, home 5 tiles + 3 rows, search/filmography labels; gates kinds 32 + browser 266, tests/realdata-068.mjs, tools/count-kinds.hl, docs/kinds.md, README + STATUSmre2026-10-02 00:23
f2b00674Merge t26 (tracker#26 + #28) into main (mission 062): short ids for every new person (castPersonId, guest route), guest stars stored on the title and created as people only when opened (/person/tmdb/<id>?show=<id> → 302), lean watch/follow clicks (showRow a small object, cast/crew from the slug, watches cached per user, face rows only after a season toggle); gate 311, tests/realdata-062.mjs, README + STATUSmre2026-10-01 23:49
0553b51ftracker#19 (mission 066): franchises and timelines — tables, /franchises, /franchises/<slug>, /timelines/<slug> (Timeline | Release sort, series by last episode), the Prequel | Timeline | Sequel widget with the franchise above, the creator's editor, TMDB collection seed in the app (resumes, paced); gate tests/franchises.mjs 48/0 + browser.mjs 266/0, tests/realdata-066.mjs, docs/franchises.md, README + STATUSmre2026-10-01 23:34
fa1f9dfatracker#29 (mission 063): unwatched check muted grey outline + check (accent only on hover), watched stays solid — no code regression, the accent outline read as ticked; gate checks real checks visibly (computed style + screenshot pixel) on /my/unwatched, show, movie, /my/movies; gate 266, tests/realdata-063.mjs, README + STATUSmre2026-10-01 19:10
10bb3f93tracker#28 (mission 061): full cast (all seasons, main cast by episodes, guest stars) + crew (created by, directed by, written by, screenplay, story, music) — stored by the details completion, the daily sync, the search import (one details request) and a background credits job (resumes, RSS limit); show page collapsed after 20 with client-side Show all; showBySlug via a slug map; gate 249, tests/realdata-028.mjs, README + STATUSmre2026-10-01 16:12
d8b12d67tracker#27 (mission 060): short ids for movies, series and persons — old 702 kept (data/old-short-ids.json), new random [a-z0-9]{5} unique across both, claimed at creation, background backfill (resumes), shown under poster/photo, /<shortId> → 301; gate 259, tests/realdata-060*, README + STATUSmre2026-10-01 14:59
25a50bc4tracker#26 (mission 059): titles from a filmography are completed — on open (skeleton, step-wise face showComplete, no reload) and by the in-app details repair (resumes, TMDB-paced, series in parts); cast from TMDB credits; gate 231, tests/realdata-026.mjs, README + STATUSmre2026-10-01 14:37
f14db671tracker#22-#25 (mission 058): episode air dates, season check = all episodes watched, movie watched check (+ /my/movies count), /genres/<genre> pages (movies + series, newest first, paginated); gate 238, tests/realdata-058.mjs, README + STATUSmre2026-10-01 14:06