Codify the standing rule at the top of the Tests section: the dev DB is
the test environment, and the invariant every test must uphold is the
testuser experience — its settings, scans, and reading data must be
exactly as the suite found them. The helpers exist to enforce this;
when a test needs protection the helpers don't provide, expand the
helpers rather than hand-rolling.
New known-gotchas from the 2026-10-21 incident cluster:
7. scanner re-extraction is write-after-success — audio_metadata is
guarded on both writers; new scan-owned column writes need the same
guard.
8. templ double-escapes style attribute values (sanitizer + EscapeString)
— emit quote-free CSS via unquoted url() tokens through SafeCSS.
9. vp test runs the vite config — build-hook plugins fire in the test
context too; outDir-mutating plugins need an "am I a build?" guard.
The testuser-experience rule: the suite must leave shared settings
exactly as it found them. setSystemSettingRestored snapshots the prior
system_settings row, writes through the sanctioned API path (the only
one that reloads the settings registry — raw SQL leaves the cache
stale), and registers a cleanup restoring what WAS there, never a
hardcoded value.
setExtractorVersion now rides it, so the ev-pin tests no longer
restore scan_extractor_version to a hardcoded 1: a deliberate dev bump
of the extractor version survives a suite run instead of being
silently cancelled (the live dev server obeys that setting on its next
poll and would stop re-extracting parser fixes). The now-redundant
explicit t.Cleanup restores are removed from
TestExtractorVersionReextract and pinExtractorVersionOne.
Three tests in scanner_audio_guard_test.go pin the routes the wipes
traveled; each drives a real scan/rescan/edit over ffmpeg-rendered
fixtures and asserts the payload survives with duration and manifest
intact:
- TestRescanMediaItemPreservesAudioMetadata: folder book through the
focused payload writer.
- TestRescanMediaItemPreservesAudioMetadataSingleFileM4B: the incident
class verbatim — single-file m4b (no layout key) routes through the
generic updater; fixture rendered with ffmpeg AAC/ipod
(writeTestTrackM4B).
- TestMetadataEditorSavePreservesAudioMetadata: the metadata editor's
full-row save through PUT /api/media-items/:id.
Verified to bite: with the handler fix stashed, the editor test fails
with "metadata edit must never NULL audio_metadata"; with it, green.
PUT /api/media-items/:id (the metadata editor) writes the FULL row via
UpdateMediaItem, whose query sets audio_metadata unconditionally — and
audio_metadata has no editor field, so the params literal left it nil
and every save NULLed the payload. A plain metadata edit (genre tweak,
title rename) dropped any audiobook back to "1 second long" with no
track manifest: the same incident class as the scanner wipe, sitting
on the most-used admin path in the app.
Carry existing.AudioMetadata through the write. Write-after-success:
only the audio chain (probe-backed extraction /
UpdateMediaItemAudioBookPayload) ever replaces that column.
The 2026-10-21 wipe: a re-extraction pass NULLed audio_metadata on the
m4b audiobooks — every affected book dropped its chapters, track
manifest and duration and displayed as 1 second long (the reader also
lost its manifest, since single-file books synthesize it from the
payload).
Guards on both writers of that column:
- updateMediaItemWithMetadata (generic full-row pass): when the new
extraction carries no audio payload, keep existing.AudioMetadata.
Single-file m4b books route here, and a failed/absent extraction
leaves AudioMeta nil.
- updateAudioBookItem (focused payload write): an empty payload (no
tracks, no chapters, no duration) over a populated row is a failed
extraction — refuse the write with a loud warning and let the next
scan retry.
audioPayloadIsEmpty() pins the empty-payload predicate (unit-tested):
the column is only ever replaced by fresh data from a real audio
extraction, never by absence of one.
manifest.json has referenced /static/icons/icon-{192,512}.png since
April, but the directory was never committed — ~100 404s per browsing
session and a broken PWA install. Rendered both sizes from the
existing favicon book glyph on the Tokyo Night palette
(#1a1b26 background, #7aa2f7 stroke).
audio_reader.templ was the only template emitting <!DOCTYPE html/>
(trailing slash is a parse error), which put the audio player page in
quirks mode — document.compatMode was BackCompat while every other
page is CSS1Compat. Quirks layout rules are a standing hazard for the
player's h-screen/flex page even where it happens to render. Drop the
slash.
templ's codegen wraps style attribute expressions in a sanitizer AND
templ.EscapeString, so any quote character — even inside templ.SafeCSS
— reached the DOM as literal ' text. The hero backdrop's url('...')
never parsed: no blurred cover on any book detail page, and Firefox
resolved the mangled token as a relative URL, fetching /media/& (500)
once per page load.
Fix: emit quote-free CSS — an unquoted url(...) token built through a
new cssURL() helper that percent-encodes the characters that would
break it (space, quotes, parens, backslash). Stored cover paths are
already percent-encoded URL paths, so the escaping is defensive only.
Clicking the bottom progress slider left it focused, and the reader's
keydown guard treats INPUT focus as typing — so arrow keys stepped the
range input instead of turning pages. With step="any" each press moved
~1% of the book and key repeat crossed the whole book in seconds; the
debounced progress saves then persisted every wild landing (positions
scattered from page 5 to 285 to 100%, "progress no longer saves
correctly after page turns").
Two layers:
- reader.templ: all three progress sliders blur on pointerup, so a
mouse scrub no longer leaves focus behind.
- reader.ts handleKeydown: ArrowLeft/ArrowRight on a focused progress
slider preventDefault, blur, and fall through to the page-turn
handlers; every other key still belongs to the control, so a
keyboard user who Tabbed to the slider keeps arrows for it.
The cleanStaleChunks plugin lives in the shared vite config, so its
closeBundle hook also fired under `vp test` — where generateBundle never
runs, produced stayed empty, and the sweep deleted every hashed chunk
in web/static (28 files → 3). The entries still referenced their old
imports, so every page 404'd and Alpine never started: the whole reader
and media system appeared dead after merely running the test suite.
Guard: skip the sweep when produced.size === 0 — not a build run, no
build output to clean. Verified: build → 28 files, test → 28 files;
back-to-back builds stay byte-identical.
Test infrastructure uses the vite-plus stack: vp test (bundled Vitest 5),
jsdom environment configured in vite.config.ts, colocated web/src/*.test.ts
files that tsc type-checks, vp fmt formats, and the bundler never imports.
- markdown.test.ts pins the notes XSS defenses: parser-level html escaping,
DOMPurify FORBID_TAGS, scheme checks, footnote rendering, html->markdown
round-trips incl. the headingless-table rule.
- filter_builder.test.ts pins the deep-link wire contract (ruleToFlatParam),
smart-serialization rules (isFlatable), and pruneIncomplete semantics.
- password_validation.test.ts covers the rules fetch + fallback and the
full submit-gating matrix through the real DOM flow (debounce included).
- api.test.ts pins response/error contracts incl. the HTTP-status and
Unknown-error fallbacks and the 404-as-null progress contract.
- storage.test.ts covers token/theme/library round-trips, the __all__
sentinel, and the SSR cookie.
- toast.test.ts pins config completeness for every ToastType (the missing
'warning' entry regression) and render behavior.
Found + fixed while writing these: the no-useless-escape fix in
d7d14957 had de-escaped the ']' in hasSpecialChar's character class,
closing it early — special-character validation silently stopped
matching most specials. The new password suite caught it immediately.
jsdom chosen over happy-dom after empirical comparison: DOMPurify's
sanitizer output is mangled under happy-dom (forbidden tags survive),
which would have made the security suite test nothing.
Note: AGENTS.md gains a 'Web (TypeScript) tests' section documenting the
workflow and the happy-dom gotcha.
Matches the error/success/info entries visually. bg-yellow-500/90 is
generated by the Tailwind scanner from this file itself (content globs
include web/src/**/*.ts), same as the sibling bg-*-500/90 classes — no
safelist entry needed.
tsc --noEmit and oxlint web/src are now both clean; vp check passes.
- htmx: one canonical Htmx interface in shims.d.ts (ajax/process/trigger
per the htmx 2.x API). alpine.ts had a second, conflicting Window.htmx
declaration (TS2717) — removed; bookPicker/collections lose their
(window as any) escapes now that Window.htmx is properly typed.
- tsconfig lib ES2020 -> ES2021 + DOM.Iterable: replaceAll exists
(escapeHtml) and for..of over querySelectorAll results is legal.
Type-checking only — vite emits at esnext regardless.
- custom-section-builder: selectedBooks holds identity fields only;
typed as Pick<BookInfo, ...> instead of forcing partial API shapes.
The API-critical BookInfo contract is untouched.
- toast: real bug found by the type system — setup.ts passes 'warning'
but getToastConfig had no entry, so the warning toast threw TypeError
at runtime instead of displaying. Added the type + config. Uses
bg-yellow-500 (verified present in scanned templates => generated by
Tailwind); the one intentional behavior change in this commit.
- book-detail: instanceof narrowing replaces an as-cast; unused
AnnotationSource import removed.
- password_validation: explicit HTMLElement | null for the checklist
list (querySelector('ul') infers HTMLUListElement, conflicting with
the container-as-list branch).
- stats: tooltip parsed value typed via annotation (bar=scalar,
line={x,y}); dead getActiveRange() removed (was already tree-shaken
out of every bundle).
- library-switcher: onSwitch honestly typed void | Promise<void>
(series.ts passes a sync navigator).
- 20 unused catch bindings converted to optional catch binding.
Verification: fresh build is byte-reproducible; every chunk is
identical to the pre-fix build except main.js (instanceof narrowing,
regex fix, stats refactor), reader.js (destructure), and the shared
storage chunk (renamed — carries toast.ts). reader/main diffs audited
hunk-by-hunk against the intended edits.
The extractor/force-rescan tests assert manifest stamps against the
default version (1), but the setting lives in the SHARED dev DB — any
prior run (or a manual admin-UI bump) can leave it at 2, failing every
stamp assertion through no fault of the code. Each ev-sensitive test
now pins the setting to 1 at start and restores 1 on cleanup; proven
against a deliberately bumped setting (tests pass, row restored).
Force rescan contract change (owner-directed): the editorial-keep skip
is retired. Force rescan now re-derives EVERY field on EVERY book from
all sources — multi-file books included (the folder/zip flows ran the
full chain on create and on manifest-flagged edits, but skipped the
sidecar apply under force). metadata_overrides is the only protection:
ApplyMetadataOverrides re-applies manual values inside the writer, so
manual edits survive; everything else re-derives.
- A force pass also STAMPS each re-derived book's manifest with the
current parser version (success-only — same truthfulness rule as
Phase 10) and closes its open processing issues (Phase 11 self-heal).
Force rescan and the ev-bump trickle now converge: either leaves the
library fully current; the deploy order is bump-first, then Rescan
All (documented in the plan and contract_pins).
- Single-file force rescan already ran the full chain — unchanged,
now pinned.
NEW filter field manually_edited ("Manually edited", boolean):
cardinality(mi.metadata_overrides) > 0 / = 0. Composed-engine (qf)
only — the builder's Book facts group gains it with is true/false;
no value menu, no suggestion surface; family lens inherited.
Tests: multifile force applies sidecar metadata + stamps + override
protection; single-file pin; the filter both directions scoped to the
test library; adjacent suites (Audiobook, Calibre, ExtractorVersion,
CrossLibraryMove, ProcessingIssues, FilterBuilder, SchemaIdempotency)
green. The extractor mechanism itself is untouched — the owner bumps
scan_extractor_version manually via the admin UI (plan Phase 13 §3).
processing_issues only recorded cross-library move rejections; files
that failed to import vanished into an error counter and stdout. The
ledger now records every failure class and heals itself:
Schema (in place, pre-release boundary):
- media_item_id NULLABLE + file_path TEXT (deliberately unbounded —
an over-long path is itself a failure mode; a VARCHAR(500) ledger
column died on the exact 22001 it existed to report — caught by the
new test). Row-less failures key on (library, file_path); item-backed
on (item, type). Widened dedup = COALESCE-expression unique index;
the old table UNIQUE stays (NULLs pairwise-distinct for row-less).
- Conflict-target upsert: repeat failures UPDATE the same row and
REOPEN resolved ones — never duplicates.
Wirings:
- CreateMediaItem failure (single-file + multifile createAudioBookItem)
→ row-less import_failed (error).
- Extraction hard-fail with fallback import → item-backed
metadata_extraction_failed (warning).
- Change-path re-extraction failure (the ev/sidecar guard) → item-backed
metadata_extraction_failed; LogProcessingIssue keeps its signature, a
row-less helper joins it.
Auto-resolution:
- Successful import → ResolveOpenProcessingIssuesForPath (library,
file_path).
- Successful (re)extraction → ResolveOpenProcessingIssuesForItem
(single-file change path, multifile editorial apply).
- Unchanged-skip scans do NOT resolve: extraction never re-ran, so the
degraded metadata is still degraded — the page heals exactly when the
book does.
Family lens on ALL issues surfaces (list, stats, resolve, delete, SSR
page): explicit out-of-set library id → 403, never silent. This
tightens one pre-lens behavior — a nonexistent library id returned
200-empty, now 403 (pinned, changelog entry in docs/contract_pins.md).
List is LEFT-JOINed so row-less rows appear (media_item_id "",
title falls back to the path).
Tests: full lifecycle state machine through real scans (fail → appears,
repeat → no dups, success-at-path → resolved, LEFT-JOIN list shape);
family lens 403s with a fresh non-shared admin; all 7 pre-existing
issue suites green (two expectations widened to the lens, changelogged);
adjacent suites (CrossLibraryMove, Audiobook, Calibre,
ExtractorVersion, SchemaIdempotency) green; unit suites green.
setupTestServer built the settings registry (and wired it into
systemSettingsHandler) but never called authHandler.SetSettings, unlike
main.go:112. The auth handler therefore ran with a nil registry and
defaultTimezone() took its nil-fallback branch, returning "UTC" for every
Register/AdminCreateUser in tests regardless of the system default_timezone
setting — so the 740213f0 inheritance pin in TestAdminCreateUser could never
pass and failed every full-suite run.
Also replace the TestAdminCreateUser "stable-ish" uniquifier (MaxConns is
pinned to 1 by the harness, so it produced the same user every run and
depended on the next run's cleanup sweep) with a real uuid suffix.
Parser fixes (7e4c97f6, genre model v2) never reached rows ingested
under an older parser: the sidecar never changed, so every fingerprint
said 'unchanged'. The manifest JSONB now carries the reserved 'ev'
stamp — the parser version that produced the item's row.
Semantics (all pinned by tests):
- ABSENT ev decodes as 1 ('v1-or-unknown'); stamping is truthful-only:
create paths + post-extraction stores stamp the setting's current
value; deploy-day backfill and the extract-failure guard stay UNSET
(a row never claims a version its data can't prove — a mid-backfill
bump self-heals next poll); repoint reconciliation preserves the
stored ev (stat-only refresh re-extracted nothing).
- NEW setting scan_extractor_version (seeded + registry + getter,
int default 1, category scanner): a manifest whose ev != the setting
re-extracts through the SAME change path (full extraction →
updateMediaItemWithMetadata → fresh stamped manifest). Sidecar-less
books get an entries-less manifest as their ev home (create: stamped;
first-sighting: unset).
- Catch-up bounded: evReextractBudgetPerScan=100, atomic (watcher pass
can overlap a poll scan), reset per ScanFolders — amortized +
resumable across polls; no force rescan, no flood.
- Multifile books included: the ev gate rides multifileSidecarChanged's
shared ladder; ev-driven manifestChanged drives the editorial apply
+ stamped re-store.
Settings API: the unified GET/PUT /api/system/settings now lists the
key (registry-driven, auto-surfaces in the admin UI). Test harness:
the registry is now Loaded + wired into the SystemSettingsHandler
(main.go parity) — settings writes through the API were 503s before.
Verified: unit (ev normalization, budget CAS, GetExtractorVersion
cache paths, registry clamping); integration state machine —
create-stamp ev=1 → equality skips (corruption probe persists) →
wiped scan-state backfills UNSET without re-extract → bump re-extracts
ONCE (sidecar value restored, ev=2) → stable (second corruption
persists); multifile twin green; adjacent suites (Audiobook, Calibre,
CrossLibraryMove, SchemaIdempotency) green; live app restart re-scans
1198 files with 0 changed (deploy-day rule), setting listed at
/api/system/settings.
The client's operator map passed the display string as BOTH the map
key and the OpMeta.key, so the menu clicks transmitted 'doesn't
contain' / 'isn't' / 'is empty' / 'at least' / … literally. The server
rightly rejected them (unknown operator → 400), the count fetch's
catch hid the counter, and flipping away from such an operator never
came back. Only self-named ops (contains, is, has, between…) worked.
- Map keys now DERIVE from meta.key (Object.fromEntries over the op
list) — wire value and label structurally cannot diverge again.
- New scripts/filter_builder_harness.js: a playwright harness that
drives the real drawer in headless Chromium and sweeps every author
operator, capturing filter-count responses and console errors — the
first UI-level regression net for the builder (the server tests
could never see a client-side mis-key).
- Verified live: contains 8 → doesn't contain 60 → is 8 → isn't 60 →
is empty 2 → refill 8; zero console errors across the sweep.
The count fetch reads the library select at fetch time, but nothing
re-triggered it on a scope change — the number stayed stale (8 across
all books) after switching to a library where the rules match nothing.
The picker's change event now reschedules the count, same as rule
edits do.
The browse-on-click blank sentinel was trimmed to empty (correct), but
the typed query had already been bound unconditionally — the id-source
suggestion query has no score CASE to reference it, leaving an unused
parameter Postgres rejects (42P18). Every click of an empty library /
collection / added-by box answered 'No matching values' via the error
path.
- ComposedFieldValues binds the typed query ONLY in branches that
reference it and only when non-blank; the score expression is
'0 AS score' for browse (no param) and the similarity CASE when
typing. composedIDSourceValues takes the raw typed string and binds
it only when the name filter is built. userBind already followed
this discipline; audited every bind site in the file (fuzzy helpers,
composed search, count) — all interpolate immediately.
- Tests: library browse under rules (exactly the surviving library,
count pinned), series browse blank, added_by browse blank (no
error), collection browse blank (offered with count) and under a
zero-match rule (never offered). Verified end-to-end against the
live dev data: author+Sapkowski AND narrator+Kenny → Audiobooks (8).
- Blank typed queries rode the STATIC autocomplete path for the classic
six fields, where a single space similarity-matches into nothing —
clicking an empty value box with no other filters said 'No matching
values' in a library full of values. Blank = browse now always rides
the composed engine (which trims and answers the top values); static
keeps its pinned contract for real typed queries. Regression-pinned:
browse with no rules lists the field's values.
- The value popup now drops BELOW the input (top: 100%), not over it.
- Field-menu group headers lost position:sticky — the search box alone
stays pinned.
- The flash was a race: on localhost the suggestion fetch resolved
between the focus (mousedown) and the click event, so the just-opened
menu was closed by its own click-outside handler — the input sits
outside the menu div. Clicks anywhere inside the value cell now
never close the menu; outside clicks do.
- The menu opens immediately in a loading state; an empty browse
renders 'No matching values' instead of nothing. Plain free-input
fields (page_count etc.) open no menu at all.
- Field menu: the filter input moved outside the scroll region — it
sticks while the grouped list scrolls beneath it.
- Per-rule connectives: the chip before each rule flips that rule
between AND and OR — OR-joined neighbors merge into an invisible
OR-block, AND pulls them back out. The tree stays the single source
of truth (implicit blocks are a client marker the server ignores),
so deep links, saved filters, and the composed engine are untouched.
Group-mode pills are gone; explicit groups remain as pure structure
(+ group). Compile limit, same as Excel: flat chips express
AND-of-OR-groups; a top-level OR of AND-blocks still needs nesting.
- Browse-on-click: clicking an empty value box now fetches the field's
top values UNDER the other rules (blank typed = browse, not a
similarity filter) — 'assign an author, click Publisher, see their
publishers' without typing. Server: the composed suggestion engine
trims the typed query; client: the >=2-char guard no longer blocks
the empty query, browse limit 15 vs 8 while typing.
- Tests: the browse case pinned (blank sentinel + rules → the field's
values under those rules, zero-match values absent).
The ALL/ANY pills required mapping ALL→AND, ANY→OR that nobody should
have to learn. The pills now say AND / OR directly (wire values stay
all|any — server contract untouched), and the AND/OR text between
sibling rules became a clickable toggle for its parent group, so the
switch lives exactly where you're looking.
- Root bug: incomplete rules (the ready rule, every new group's first
rule) were sent to the server — the live-count fetch toasted a 400
and every suggestion fetch failed silently (why author typing showed
no dropdown). Fix: completedTree() prunes incomplete rules and empty
groups; the count fetch skips when nothing is complete, suggestion
fetches send only the other COMPLETED rules, and serialization wires
only completed rules (an unfinished rule was never a filter).
- Group remove button was x-show'd away on depth-0 groups — exactly
where user groups live. Always visible now.
- Field menu: filter input at the top (Enter prevented), no-match
state; group headers restyled as flush-left tinted sticky bands with
options tabbed in beneath (user's layout pick).
- toast.ts: filter-count fetches join the suggestion-fetch exemption
(a failed background hint never interrupts).
- Field registry (services/filter_registry.go): ~55 typed fields —
every book column, file/lifecycle facts, the three HOLDING
STRUCTURES (library, collections with excluded-aware membership,
series with FULL primary+alternate_info membership), and the whole
user-relative class (status, progress, last read, my rating, review,
notes/highlights presence, note/highlight text, time spent, pages
read, collections, added by) reading the VIEWER's rows. Operators
are typed constants; negation is an operator.
- Composed engines (services/filter_query.go): one Go clause builder
over raw pgx feeds the search (identical SearchMediaItemsUnifiedRow
shape), the field-aware suggestions, and a true-COUNT live count.
Mandatory WHERE head everywhere (family lens + integrity facts).
Values are always bound parameters — no user string reaches a
field/operator position, ever.
- Wire: the qf JSON param (validated strictly — unknown
fields/operators/malformed values are 400); qf wins over flat params
on hand-crafted URLs; flat deep links keep working unchanged.
- Endpoints: GET /api/media-items/filter-count (live count);
field-values autocomplete upgraded in place — any registry field
key, qf binds the OTHER rules (symmetric under-filter suggestions,
zero-match values never offered); classic six without a tree keep
their static pinned contract.
- UI: the drawer becomes the rule builder — the ONLY filter UI (templ
markup, Alpine tree state, sanctioned island; grouped field menus,
operator menus with red negations, field-aware value menus, live
count, chips per rule with red negations, URL rehydration, smart
serialization flat-when-simple else qf).
- Tests: composed semantics (AND/OR/negation/nesting/enums/bools/
numbers/dates/empty), every user-relative field class with the
my-data weld (two readers, different worlds), note-text substring
pin (catharsis ≠ Carthage), family lens (hidden libraries nowhere,
403s), suggestion-under-rules exactness, qf validation + precedence,
flat-vs-tree equivalence; registry completeness unit test.
- Remove the Phase 7 facet-panel UI: SSR route, facet_panel.templ,
FacetCountsForView, drawer panel block, panel TS + .facet-* CSS.
The ENGINE stays (FacetCountsUnderFilters, all-facets endpoint,
drift guard) — it becomes the Phase 9 builder's machinery.
- Fix the All-Books bug: field-values autocomplete (author/genre/
tags/narrator/series/language) treats library_id as OPTIONAL —
absent aggregates every VISIBLE library (family lens in-query);
explicit out-of-set library_id is now 403, never a silent pass.
The four scalar Search*Values queries get the nullable shape.
- toast.ts: suggestion fetches are exempt from error toasts
(console-only) — a failed background hint never interrupts.
- TestFacetPanelPartial deleted with the behavior; TestFieldValuesAll-
Libraries pins the fix (200 + cross-library aggregation, family
lens, 403). Contract-pins changelog entry added.
The 9 Batch A facets could not be SET from the web UI, the facets
endpoint had no in-app consumer, and the registry's libraryTypes hints
rendered nowhere. This phase closes all three.
- ONE UNION-ALL sqlc query (FacetCountsUnderFilters) computes all 14
listable facets in a single round trip: each arm omits its OWN
facet's filter block and carries every other registered facet's
block (copied mechanically from SearchMediaItemsUnified) — classic
faceted semantics: with filters active each facet counts under the
others; with none, blocks are neutral and the counts are the v1
global-per-library ones. One query serves both cases. The 14
per-column Batch C queries are deleted.
- Endpoint: no facet param → the new all-facets shape
({facets: {param: [{value, display?, count}]}}), filters bound via
the shared ParseMediaSearchParams, library_id optional (the
All-Libraries view, family lens still enforced); facet=X keeps the
pinned single-facet shape (now routed through the same query with
neutral filters).
- SSR partial GET /media-items/facet-panel: templ renders the drawer's
value lists — registry order, Batch D library-type scoping
(FacetsForLibraryType: manga_type on manga, reading_direction on
manga+comics, narrator on audiobooks; All view = union), prettified
genre labels + language display names, active values marked. Family
lens: out-of-set explicit library_id is 403.
- Drawer wiring: the panel fetches on drawer open and after every
apply with the CURRENT filter query string; a value click toggles
that facet's comma-joined filter (hidden input where the drawer has
no field — same as deep links) and re-applies the form; chips and
URL stay in sync. Values are server-rendered (rendering doctrine);
bookshelf.ts owns only the click/fetch orchestration.
- prettifyGenre moves to internal/utils as the canonical
implementation (services' display enrichment shares it; templ wraps
it; the TS port in bookshelf.ts is unchanged).
- Drift guard extended: FacetCountsUnderFiltersParams must carry every
registry facet's param, the bools as pgtype.Bool.
Tests: under-filter exactness (own-block omission, OR-within,
AND-across), family lens (hidden values absent; 403 on both shapes),
all-shape + pinned-shape coexistence, partial rendering with Batch D
scoping + active marking + 403. Full container suite green except the
documented pre-existing TestAdminCreateUser red.
Two suite failures, neither caused by scanner code (the Phase 6 diff
touches no handler):
- search_test invalid_library_id expected the pre-family-lens 404;
7e586df4 shipped the AGENTS-mandated 403 for every out-of-set id
(existence must not be enumerable) and docs/contract_pins.md:195
documented it — this one subtest's expectation was missed. Pin
widened to the shipped 403.
- TestWebSocketPingPong deleted: it asserted a text-level ping→pong
exchange no server code has ever implemented and no client has ever
sent; it passed vacuously through its read-timeout path until
interleaved scan_progress broadcasts from the live server's scans
reached the socket first. Keepalive is the standard protocol-level
ping the hub already sends. Never-built behavior is not a pinnable
contract.
Both recorded in docs/contract_pins.md. Full container suite is green
except the documented pre-existing TestAdminCreateUser red.
Reverses the Phase 4 scope cut: for EXISTING folder/zip audiobooks a
sidecar edit was invisible to everything, including force rescan —
updateAudioBookItem writes only scan-derived fields, and these formats
often have no useful embedded tags, so the sidecar is the only source.
- Change fingerprint gains the manifest compare in both processors:
changed := size differs || track payload differs || manifest drifted.
The ladder runs against the bulk-loaded map entry (stat pairs only,
zero extra queries); NULL stored manifest backfills without
re-extraction (deploy-day rule, same as single-file).
- Manifest-changed applies the once-extracted book metadata to the
editorial columns (updateMediaItemWithMetadata, overrides protected)
and re-runs the playlist hook. The apply runs BEFORE the focused
payload write: the generic update's format-group pass derives mime
from the sentinel (none) and chapters from metadata.ChapterCount
(unset) — the payload write must win (rescanMultiFileAudioBook's
ordering). Track-only changes keep the existing editorial-keep
update; force rescan keeps it too.
- Manifests seeded at multifile create (folder names derive from the
sentinel — the exact representative the extractor reads — via
detectSidecarChangeWithRepresentative, so manifest and extractor can
never disagree; a directory-basename cover is ignored by the
extractor and therefore not tracked).
- Zip parity fixes: the fast path now lifts archive state like the
folder path always did (a returned zip book stayed archived
forever), and the change branch handles the unchanged-archived case.
- Single-file sidecar change path: extract-failure guard — the old
non-nil empty &MediaMetadata{} fallback bypassed the update path's
filename-title guard (preExtracted == nil only) and would blank the
title on a failed extraction; the row is now left untouched and the
manifest refreshed. Empty-success extraction (sidecar removed from a
tag-less book) now passes nil so the create-equivalent chain (tags →
filename) re-derives instead of blanking.
- Multifile changes now count in the scan summary (changedItems);
they were previously uncounted.
Tests: sentinel-name derivation unit pin; folder sidecar edit
(title + collection sync + fresh manifest + steady-state skip),
track-only change keeps editorial, zip sidecar edit + archive
restore, sidecar removal falls back to filename (the wipe-guard's
testable sibling). Full container suite: green except the documented
pre-existing TestAdminCreateUser red.
Part 2 of the faceted plan, Phase 5 (steps 5-6): the scan that used to
print ~4 lines per unchanged file now prints ONE summary line, and a
poll tick landing mid-scan no longer stacks a second full scan.
- Quiet scans (§4): the walk loop is silent — inserts, changes, and
errors log inside processMediaFile; the duplicate
'Successfully processed' branch (both branches printed it) and the
per-file 'processing / already exists / same size, skipping' chatter
are gone. Summary carries the trigger and a changed counter:
'Scan of <folders> (trigger: poll): N files, N changed, N new, Xs,
N errors'. Triggers threaded through every enqueue site: poll
(startBackupScan), watcher (dirty-dirs flush), admin (scan API),
initial (startup scan).
- Enqueue coalescing (§5): Worker.HasActiveScanForLibrary (JobResult
gains additive job_type so only SCAN jobs coalesce — an import
sharing the library id never blocks one). enqueueLibraryScan skips
silently; the admin scan API answers 409 with the active job id.
Enforces the one-scan-per-library invariant the bulk scan-state map
leans on.
- Watcher capability probe (§4): ONE Add on a root folder; failure
disables the watcher with a single startup line (replaces the
per-directory warning spam on inotify-broken deployments). The
watcher stays an optional latency reducer over scheduled scans.
- Per-directory sibling cache (§6): hasSiblingBookFileCached caches
the ReadDir per directory for one scan — loose-image manga re-stats
thousands of images sharing a handful of directories.
Tests: TestHasActiveScanForLibrary (unit — running/pending coalesce,
terminal jobs don't, non-scan jobs don't, empty library never, nil
worker safe). Full scanner suite green in the container.
Part 2 of the faceted plan, Phase 4 (steps 3-4). Size-only detection
was blind to metadata-sidecar edits; a size-stable book is changed by
EDITING its metadata.opf/metadata.json/cover — previously invisible
until a force rescan.
- Sidecar manifest fingerprints (media_scanner_manifest.go): the
manifest stats exactly the names the extractors consume (metadata.opf,
metadata.json, findSidecarCover patterns incl. <basename>.<ext>), so
manifest and extractors can never disagree. Shape v1:
{v, entries:[{n, s, m (unix nanos — second granularity would hide
edits landing in the backfill scan's own second), h (sha256 — covers
included: the repoint ladder needs a baseline hash for any stat-
drifted sidecar; a hashless cover turns harmless mtime drift into a
re-extraction, the exact flood this design kills)]}.
- Evaluation ladder in cost order: stat pairs identical → skip with
ZERO bytes read; stat drift → hash that sidecar (equal = host repoint
reconciliation, stat pair refreshed, no re-extraction; differing =
changed); NULL stored manifest → backfill only, NO re-extraction
(deploy-day rule: the first scan after release must not look like an
mtime flood).
- Change path: re-read full row, extractMetadata once and pass it
through updateMediaItemWithMetadata (preExtracted — metadata_overrides
stay protected), run syncPlaylistCollectionsAfterCreate (Correction 2:
a reading-order entry added to an EXISTING book now creates its
collection; add-if-missing is idempotent), store a fresh full
manifest. Creates seed their manifest at insert time.
- Multifile folder/zip books intentionally keep Phase 3 behavior: their
change path does not consume sidecar extractors, so gating them would
flip books 'changed' without picking anything up (plan correction —
the plan's §3 multifile note assumed otherwise).
- UpsertMediaScanState + GetMediaScanState queries.
Tests: media_scanner_manifest_test.go (ladder unit tests — fast path,
repoint, same-size drift, appear/vanish, unhashable legacy entries);
TestScanManifestLifecycle (seeding, unchanged-skip stability, repoint
reconciliation with stat-pair refresh assertion, same-size edit
consumed, playlist sync on the change path + idempotency, no-leak
guard). Container runner.
Part 2 of the faceted plan, Phase 3 (steps 1-2). The periodic scan
issued one GetMediaItemByFilePath per scannable file (~3,094 queries
per scan in prod, nearly all answering 'unchanged, skip') plus one
GetLibraryByFolder per file for folder resolution.
- NEW media_scan_state table: the scanner's per-item fingerprint
cache. Deliberately NOT a media_items column — the library-scoped
list emits raw sqlc rows as JSON (7537063c), so scanner bookkeeping
must never leak into API payloads nor churn the model or pins.
Manifest JSONB shape v1 documented in the schema (n/s/m/h; ev
reserved for the Phase 4+ extractor-version mechanism).
- GetScanStateByLibrary: lean bulk projection (id, file_path,
file_size, archived_at, missing_scan_count, file_sha256, manifest).
archived/hash columns ride along so the skip path's archive-restore
and hash self-heal checks stay DB-free too (refines the plan's
4-column sketch; same goal, narrower invariant).
- ScanFolders: per-run reset, per-folder library cache, one bulk load
per distinct library. processMediaFile: map hit answers existence +
size + flags with zero DB reads; force-rescan and size-change
branches re-read the full row (the change path may be expensive);
write-through on create/size-change keeps later passes honest.
- Multifile folder/zip books keep their full-row lookup: it is one
query per BOOK ROOT (not per track) and needs audio_metadata for
the duration-probe check. They join the map in Phase 4, when the
manifest compare replaces their fingerprint check.
- nil map (watcher's scanDirectory path, bulk-load failure) falls
back to per-file lookups — behavior-preserving; the failure case
logged once per scan.
Test: TestScanStateBulkLoad (unchanged rescan leaves updated_at
stable; API-created rows are seen by the bulk load; size changes
still update in place). Container runner.
Frontend half of the faceted-navigation Phase 2. 'Everything filterable
should be clickable': every displayed facet value links to the bookshelf
pre-filtered by its raw stored value; display forms are prettified at
render, never in the DB.
- templates/facet_links.templ: FacetLink (inline links) + FacetBadge
(badge links) to /bookshelf?<facet>_filter=<value>; the raw value
rides the param while the shown text is prettified, so matching and
display can never disagree. Series/tags keep their dedicated pages
by design; the card media-type icon stays unlinked (a derived kind,
not a column value — deep-linking it would send wrong values).
- book_detail.templ: publisher, language, edition, genre, imprint
cells + narrators line + age-rating/story-arc badges are now links.
- prettifyGenre (templates/utils.go + unit tests): title-cases only
entirely-lowercase values; small words stay down except first
position; & preserved. TS port for chip labels documents the twin.
- bookshelf.templ + bookshelf.ts: active-filter chips bar — Alpine
renders chips from client-local form state (sanctioned island);
one bubbled htmx:afterRequest hook keeps chips + shareable URL in
sync; removeFilterChip clears the facet and re-applies. URL deep
links now accept all 18 facet params (hidden inputs where the
drawer has no field).
Pins changelog: Batch B/C entry in docs/contract_pins.md. Rendering
ownership guard (make lint-rendering) green.
Phase 1 + backend of Phase 2 of the faceted-navigation plan. One
declarative facet registry (services/facet_registry.go) drives the
bookshelf filter system; a drift-guard unit test reflects over the
generated SearchMediaItemsUnifiedParams so a registry entry and its
sqlc.narg block can never disagree silently.
- Nine new filter params on SearchMediaItemsUnified: story_arc,
imprint, edition (scalar-fuzzy), age_rating, format_group,
manga_type, reading_direction (exact), explicit, abridged
(presence-bound bools). Comma = OR-within multi-select on every
text facet; AND-across-facets unchanged.
- Genre model v2 (user rule: a subgenre is not a genre, it's a tag):
genre_lexicon.go rebuilt as shelf set + subgenre->shelf map;
pickSmartGenre picks direct shelf > mapped subgenre > Fiction floor
> empty (unknown-only stays tags-only). Type-words (Manga, Comics,
Graphic Novel) join the demote list. Both 'adventure' and 'action &
adventure' are shelves — provider data uses both and the verified
traces pin them.
- Language filter UX: comma elements normalize through the ingest
language map, so display names ('English') match rows stored as ISO
codes ('en'). The language autocomplete now serves display names.
- NEW GET /api/media-items/facets?facet=X&library_id=... — registry-
driven value+count listing (14 static per-column queries, family-lens
join); language values carry display names; bool/range facets 400.
- Family-lens gates: an explicit hidden-library library_id is now a
403 on the search endpoints (was a silent empty set) — JSON branch
and HTMX grid partial both.
- handlers.ParseMediaSearchParams is the SINGLE param parser for the
search JSON endpoint and the HTMX grid partial; the HTML branch
previously dropped publisher_filter and narrator_filter silently.
Pins: docs/contract_pins.md changelog entries (additive request
params; new facets endpoint). Tests: facet_registry_test.go (shape +
SQL drift guard + library-type union), series_parsing_test.go genre-v2
cases, facets_integration_test.go (param filtering e2e, counts,
language display, 403s) — container runner.
Completes the A2c acceptance scenario: click an author on a book
detail page -> land on the bookshelf filtered to everything that
author wrote (including their share of co-authored books).
- Media detail + unscoped list resolved maps expose subtitle,
explicit, authors, publishers (the library-scoped raw branch and
the search endpoint already did — raw rows follow the SELECT).
- SearchMediaItemsUnified SELECT carries the new columns; the search
endpoint serializes the row directly, so clients get them there too.
- book_detail.templ renders each author as a link to
/bookshelf?author_filter=<name>.
- Bookshelf honors filter params in the URL on load (author_filter,
series_filter, genre_filter, publisher_filter, tags_filter,
narrator_filter, language_filter): pre-fills the form, applies it to
the first grid load. Comma-joined values are multi-select via the
server-side string_to_array split. Filtered views are shareable.
Pins widened: search rows (+7 keys), resolved keys (+4); changelog
entries in docs/contract_pins.md. Android follow-up plan written to
~/.opencode/plan/bookhoard-app-v2-compat.md.
Phases 4 + 5 of the metadata overhaul (plan §4, §5) plus §3 handler
wiring.
Filters (§4.1): author_filter/publisher_filter now match per-element
against the authors[]/publishers[] arrays — a co-authored book matches
any one of its authors — with comma-joined params acting as
OR-within-facet multi-select (string_to_array + cross unnest). Legacy
single-value params unchanged on the wire; series/genre/language
filters stay on their varchar columns; narrator filter already
per-element.
Collections API (§3.3-3.5): collection detail carries viewer +
hide_finished into the listing query and exposes up_next (visible-
library-enforced reading-queue head); PUT
/api/collections/:id/items/order (owner-only, full-order rewrite);
UpdateCollectionRequest gains hide_finished/show_up_next ([]
string last-value-wins for the hidden-false-sibling checkbox pattern)
merged server-side OVER stored view_settings so imported_from_series
provenance survives edits; EditCollectionModal renders the toggles
from view_settings.
Dashboard (§5.5): sections JSON carries is_playlist/complete
(dashboard.ts is the AGENTS-documented JSON-render exception; first
paint stays templ). Fully-read playlists render '✓ Playlist complete'
instead of an empty rail; imported playlists get a 📖 Playlist badge.
Frontend: book detail shows subtitle + full authors + Explicit badge;
edit modal gains subtitle, comma-separated authors/publishers (client-
split to arrays exactly like narrators), step=any series number,
explicit checkbox; collection book cards get playlist move ↑/↓ buttons
persisting the full order via the reorder endpoint (templ partial
re-render on failure — DOM order is client-local, API order is truth).
AGENTS.md gotcha #3 amended per plan §3.7: user collections are
dashboard sections by default; the show_on_dashboard premise is
retired.
Verified: full affected integration sweep green (pins widened for
up_next/is_playlist/complete — changelog entries in
docs/contract_pins.md); unit suite green; go build + templ + vite
green.
Phase 2b + 3 of the metadata overhaul (plan §2.4, §3.1-3.7).
Auto-playlist collections (§2.4): the scanner turns obviously-playlist
sidecar series (explicit '(Reading Order)'/chronological/publication/
in-order/timeline markers only — precision over recall) into curated
static collections: upsert by view_settings.imported_from_series then
name (rename-safe), position set ONLY on insert (clean integers; 1.5/1a
stay unpositioned rather than guessing), membership add-if-missing,
never removed or re-ordered by rescans, provenance in description +
view_settings, show_on_dashboard true (dashboard-eligible by design),
'scanner.auto_playlist_collections' setting (default on, Settings
Registry + schema seed).
Ordering (§3.1-3.2): collection listings order position ASC NULLS LAST
then added_at ASC (was added_at DESC — walked playlists backwards);
PUT /api/collections/:id/items/order applies full-order rewrites
(owner-only, service-gated).
hide_finished + Up Next (§3.3-3.4): GetCollectionItemsForDashboard
gains viewer_user_id + hide_finished (LEFT JOIN reading_progress,
COALESCE percentage < 0.999) and viewer_percentage rides the row for
card graying; GetCollectionUpNext serves the collection's reading-queue
head with the family-lens visible-library array enforced.
Dashboard (§3.5): user collections are dashboard-ELIGIBLE BY DEFAULT —
the show_on_dashboard gate (never writable for user rows since the
column shipped) is gone from GetUserCollectionsForDashboard; hiding
stays dashboard-prefs (hidden_collections) territory; hide_finished
filtered-everything now emits a Complete section instead of silently
vanishing; sections JSON carries is_playlist/complete.
Series page ordering hardening (§3.6): GetSeriesBooks tiebreaks by
title ASC — NULL/tied numbers previously returned heap (scan) order,
which is what the user saw.
Frontend custom-section-builder note: the builder has sent
show_on_dashboard: true since its creation; the API never bound it.
User-row visibility is now intentionally prefs-governed, so the field
stays unbound by design (documented in code).
Pins widened (docs/contract_pins.md): collection detail + up_next;
dashboard sections + is_playlist/complete; new reorder endpoint
recorded.
Verified: TestPlaylistCollectionsEndToEnd (new, full e2e: sidecar
series split, collection auto-creation with position 7, rescan
idempotency, owner-only reorder, finished-filter + completion marker +
Up Next drain) plus the full affected integration sweep green.
Phase 2a of the metadata overhaul (plan §2.1-2.3, §2.5-2.9).
The headline bug: ABS writes sidecar series as an array of STRINGS
(["The Witcher #3", ...]); the parser only accepted {series, sequence}
OBJECT arrays, so every ABS-matched book silently lost its series (found
live: 9 sidecars, zero series ingested; only folder-structure Malazan
survived). parseSeriesCandidates now normalizes string arrays, object
arrays and Libation plain strings.
Three-way series assignment (plan §2.2):
- primary bibliographic series -> series/series_number columns
- additional real series (Drizzt: Icewind Dale #1 = Legend of Drizzt #4)
-> alternate_info JSONB, comic-established shape
- ordering constructs ('The Witcher (Reading Order)') -> PlaylistSeries
candidates; NEVER the series column. Detection is explicit-marker-only
((reading order|chronological|publication order|in order|timeline) as
parenthetical/suffix) — precision over recall; a series merely
containing 'timeline' mid-name stays a series.
Sequence policy: leading-number extraction ('Book 2' -> 2, '2,
Dramatized Adaptation' -> 2, '1a' -> 1); lossy extractions preserve the
raw text in alternate_info.series_sequence_raw.
Smart genre (deterministic lexicon): compound Audible blobs split
('Epic, Action & Adventure, Military, Sword & Sorcery' -> 4 candidates),
alias-normalized, format noise demoted ('Audiobook' never wins),
most-specific canonical genre takes the genre column, all candidates
stay in tags.
Language: display names normalize to ISO 639-1 on ingest; ROOT-CAUSED
the update-path drop — CreateMediaItem call site never set Language
(explicit NULL over the column default), the update path did; create
now carries it.
Title-derived series fallback (§2.7): 'Baptism of Fire: The Witcher,
Book 3' 2-level folders split conservatively (only when the post-colon
tail matches ', Book N'), subtitle parsed for series+number, blanks-only
precedence, matching-name number fill, zero display mutation. Works
with no sidecar and no network.
Ingest fixes: subtitle/explicit land in their columns (audio tag
subtitle stops joining into title); full authors[] array (was [0]-
only); sidecar descriptions sanitized via bluemonday UGC policy
(vendored).
metadata_overrides: verified scan-update path already applies
ApplyMetadataOverrides; subtitle/explicit/authors covered by the Phase 1
override adaptation.
Tests: new series_parsing_test.go (string-array sidecar e2e, Drizzt
alternate_info, lossy raw preservation, playlist classification,
title-derived fallback incl. no-mutation + blanks-only, genre lexicon,
language map); audio-tag subtitle pin updated to column behavior;
calibre/ABS 2.5-series_index pins widened per float change (changelog'd
in phase 1).
Deferred to 2b: auto-created playlist collections (§2.4) — settings
toggle, collection upsert, ownership.
Phase 1 of the series/collections/metadata overhaul
(~/.opencode/plan/series-collections-metadata-overhaul.md).
Schema (in place, idempotent):
- media_items.series_number INTEGER -> DOUBLE PRECISION: real-world
ordinals are decimals (0.5, 1.5 novellas); INTEGER silently truncated
them (calibre 2.5 pinned as 2 in tests).
- + subtitle TEXT, explicit BOOLEAN (sidecar ingest; ABS explicit /
Audible isAdult).
- + authors/authors_search/publishers/publishers_search TEXT[] NOT NULL
(narrators A2a precedent). Arrays are write truth; author/publisher
varchars become trigger-regenerated display strings
(sync_author_publisher_display) so every varchar consumer (trgm
search, facets, API, OPDS/Kobo) keeps working unchanged.
- Row-level NULL normalization for the new arrays in
set_library_type_name (A2 lesson: Go nil -> explicit NULL -> 23502).
- Backfills ARRAY[author]/ARRAY[publisher] — deliberately no
comma-splitting ('Tolkien, J.R.R.' would corrupt; scanner kept only
authors[0] historically).
- collection_items.position INTEGER + (collection_id, position) index
for playlist ordering.
Go: sqlc regen; MediaMetadata/handler params gain Subtitle/Explicit/
Authors/Publishers (legacy singles still accepted, arrays win);
SeriesNumber int32 -> float64 end to end; metadata_overrides adapted
with legacy override keys ('author'/'publisher') preserved so existing
user overrides survive; bulk author edit writes the array.
Pins widened (docs/contract_pins.md): raw sqlc row key set + the six
new columns; changelog entry added.
Verified: unit suite green; integration sweep (bulk/smart collections/
scanner/stats/sync/queue/calibre/pins/audiobooks) green against the
shared dev DB after in-place migration. Pre-existing unrelated red:
TestOwnershipAndAdminTestSuite/TestAdminCreateUser timezone inheritance
fails identically at the pre-change commit (live shared state).
@view-transition { navigation: auto; } lets Chrome 126+ / Safari 18+
crossfade same-origin navigations — sidebar links and the series
page's library-switch navigation get a subtle fade with zero
JavaScript, softening the switch to full SSR navigation. Pure
progressive enhancement: Firefox and older browsers simply navigate
instantly. style.css regenerated via build:css:prod.
web/static/*.js is gitignored and built at image-build time, so the
hardcoded ?v= on the module script tag is the ONLY mechanism that
evicts a stale bundle from browsers. Shipping the series twin deletion
without a bump would let cached main.js (with the deleted renderer)
keep running on prod, making the paginator fix appear not to work.
Found while shipping; easy to miss because the bundle never appears in
git status.
The bookshelf select (#library-select) sits inside the htmx filter
form and the SSR route honors ?offset= — wiring a change handler or
hx-trigger="change" on it would let header.ts's synthetic restore
event re-fetch the grid at offset 0, the same stomp class that broke
the /series paginator. Comment-only guard; regenerated
bookshelf_templ.go carries it into the emitted markup.
renderDashboardCollections/renderSectionHTML re-render dashboard
sections from /api/dashboard/sections JSON on library switch, settings
save, and refresh — deliberately kept in TypeScript to preserve the
switch fade on the app's most interactive page. First-paint section
markup remains templ (Dashboard handler in internal/router/frontend.go).
Recorded in the Rendering ownership section as a sanctioned exception
so it is not silently 'fixed' without restoring a templ-partial fade
path, and so the pattern is not cloned to new pages. (User-directed:
the dashboard keeps its fade; every other page stays templ-rendered.)
New integration test mirroring a prod bulk import: a dedicated test
library with 30 series whose items share ONE identical created_at
(single-transaction NOW()), the exact data shape that made prod
pagination look broken.
Pins, via the real SSR /series route:
- page 1 renders exactly 24 deterministic series (01-24) with chip
"Page 1 of 2", no Previous, and a Next href carrying library_id and
page (templ escapes & as & inside attributes)
- page 2 renders the remaining 6, disjoint from page 1, chip
"Page 2 of 2", no Next, Previous href correct
- missing page param defaults to page 1; page=0/-3/two clamp to page 1
Guards both the SQL tiebreaker and the /series handler contract;
would fail on any return of the stomp or an unstable sort.
GetDistinctSeries ordered by MAX(created_at) DESC alone. Bulk imports
inserted in a single transaction share one NOW() timestamp, leaving
large numbers of (library_id, series) groups tied on MAX(created_at);
Postgres does not guarantee tie order between executions, so
LIMIT/OFFSET pages could repeat or drop series across page loads.
Add the GROUP BY keys (library_id, series) as tiebreakers so the sort
is total and deterministic. Comment documents why the tiebreaker is
required so it is not dropped as noise. sqlc regenerated (interface
unchanged; the comment flows into querier.go docs).
"Page N of M" now renders as small secondary-color text on the page
background rather than an accent chip (user preference; the old TS
twin happened to render this way and the chip was only being seen
after JS took over). With the twin deleted, this templ markup is the
only paginator rendering, so the look is stable.
John O'Keefe