Skip to content

v1.12.8#781

Merged
ajslater merged 342 commits into
mainfrom
develop
Jun 15, 2026
Merged

v1.12.8#781
ajslater merged 342 commits into
mainfrom
develop

Conversation

@ajslater

@ajslater ajslater commented Jun 15, 2026

Copy link
Copy Markdown
Owner
  • Fixes
    • Fix read/unread filter leaking one user's read state to other users.gi

ajslater added 30 commits April 8, 2026 19:11
@ajslater
make scope private
384d6f0
@ajslater
update devenv
929ac98
@ajslater
update deps. migrate to unhead v3
3f23d63
@ajslater
update deps
ca2b99a
@ajslater
fix creating reader global settings
ac23fb6
@ajslater
fix caching
b118359
@ajslater
rename codex build-dist to codex-ci
a9bab2a
@ajslater
fix image name. make gha steps depend on each other more.
498a2d0
@ajslater
fix gha syntax errors
00307e2
@ajslater
Merge branch 'main' into develop
0e62ca4
@ajslater
names for gha steps
1a0865a
@ajslater
use ghcr.io for python-debian base
8d73870
@ajslater
update deps
ce3c806
@ajslater
format dockerfile
aea24a4
@ajslater
picopt treestamps
928084a
@ajslater
fix custom covers not importing. v1.10.11
34325e8
@ajslater
fix custom covers count in admin view
8364163
@ajslater
bump news for custom cover count fix
84a958e
@ajslater
update deps
d28662c
@ajslater
Merge branch 'main' into develop
97daa11
@ajslater
codex identification in server tag and opds generator tag
9abd4fd
@ajslater
update deps
67a823f
@ajslater
force no entries on opds start page
2791819
@ajslater
common opds start page mixin. emtpy group objects on start page
fe6ef60
@ajslater
update deps. typechecking.
68f1c65
@ajslater
api change q to search
c2a37bf
@ajslater
standardize search param as 'search' instead of 'q' or other variations
62baa9e
@ajslater
remove errant icecream
045293d
@ajslater
clear settings on backend
44a9d32
@ajslater
Squashed commit of the following:
ceba72d 
    Fix clear settings null bug, add global settings clear button

    - clearComicSettings was setting book.settings to null, causing
      Object.entries() to throw TypeError downstream in getBookSettings
    - Add null guard in getBookSettings as defense-in-depth
    - Add null guard in isClearDisabled computed
    - Add clearGlobalSettings action and clear button to Default Settings panel
    - Compare against READER_DEFAULTS to determine if global clear is disabled
ajslater and others added 27 commits May 20, 2026 23:19
@ajslater
Merge branch 'main' into develop
19dee26
@ajslater
update devenv
a5b78f5
@ajslater
update deps
a0bb64a
@ajslater
update devenv
3d76820
@ajslater
move script
9618638
@ajslater
reduce complexity
94aff1f
@ajslater @claude
fix(frontend): keep modulepreload URLs absolute on lazy-loaded routes
c4e5e3d 
vite-plugin-dynamic-base 1.4.1 dropped the leading `/` it used to add
inside template-element replacements. The codex publicPath stripped
APP_PATH's leading slash with `.substring(1)` to avoid a double `/`
with the old behavior; under 1.4.1 the strip made chunk URLs
relative, so deep routes like /admin/libraries fetched chunks from
/admin/static/... and got HTML from the Django catch-all, breaking
the admin panel and any lazy-loaded route.

Bump version to 1.12.6 and add news entry.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
@ajslater
update devenv
986cb72
@ajslater
change sentinel to None to make cron checking more sensible
36a6b12
@ajslater
Merge branch 'main' into develop
dc1e004
@ajslater @claude
fix(librarian): allow .codex-cover dotfiles through ignore filter (#776)
dc0d86f 
a709a74 added a "." prefix to _IGNORED_BASENAME_PREFIXES so the
poller's walker, the watchfiles filter, and expand_dir_added all
skip hidden files before any cover predicate runs. User-supplied
folder covers named .codex-cover.jpg / .png / .webp / ... are
dotfiles by name but legitimate covers, so the dotfile filter
swallowed them and broke custom folder covers from v1.12.5 on.

is_ignored_basename now exempts a basename when match_folder_cover
claims it. .DS_Store, .git, .Trashes, .codex-cover (no extension),
.codex-cover.txt etc. still get filtered.

Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
@ajslater
bump news
dfa8445
@ajslater
update deps and bump version to v1.12.7
4acc784
@ajslater @claude
fix(auth): allow Remote-User SSO for DRF API endpoints (#777)
41ef2d6 
DRF's ``SessionAuthentication`` enforces CSRF on unsafe methods, which
rejects proxy-authenticated API clients that legitimately carry no CSRF
token. Add ``HttpRemoteUserAuthentication`` (subclass of DRF's stock
``RemoteUserAuthentication`` reading ``HTTP_REMOTE_USER`` instead of
``REMOTE_USER``) and conditionally prepend it to
``DEFAULT_AUTHENTICATION_CLASSES`` when ``CODEX_AUTH_REMOTE_USER`` is
enabled, so proxy-forwarded ``Remote-User`` API requests authenticate
through the existing ``RemoteUserBackend`` without hitting CSRF.

Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
@ajslater
bump news
45fb947
@ajslater
update deps
d780a8b
@ajslater
Merge branch 'main' into develop
2016704
@ajslater
update devenv and deps
57d5cf3
@ajslater
move VPullToRefersh out of labs
5918d0b
@ajslater @claude
fix(frontend): convert v-dialog em widths to px so dialogs aren't sli…
541907b 
…vers

Vuetify 4's v-dialog defaults to location 'center center', which activates
the connected overlay location strategy. That strategy caps width with
parseFloat(props.maxWidth), so "20em" parses to 20 and the em is dropped,
collapsing dialogs to a ~20px sliver. Convert every em-based max-width/
min-width on a v-dialog (or forwarded via $attrs) to unitless px at 16px/em.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
@ajslater @claude
fix(bookmark): scope read/unread filters per user (v1.12.8)
c536678 
The UNREAD filter used Q(bookmark=None), a multi-valued relationship
test for "comic has no bookmark from anyone" rather than "no bookmark
for me". Once any user finished a comic it dropped out of every other
user's (and anonymous visitors') unread view. UNREAD now negates the
per-user finished predicate (~(my_filter & finished=True)), compiling
to a per-user NOT-EXISTS subquery.

Also harden get_my_bookmark_filter: an anonymous visitor with no
established session_key resolved to `session_id IS NULL`, which matches
every authenticated user's bookmarks. Return a match-nothing filter
instead; a session key is minted only when a bookmark is written.

Adds tests/test_bookmark_filter_isolation.py covering cross-user and
anonymous isolation for READ/UNREAD/IN_PROGRESS.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
@ajslater
shfmt
1f867c1
@ajslater
shfmt
e5a2a41
@ajslater
update devenv
5dd3a0d
@ajslater @claude
fix(types): narrow OPDS redirect settings to Mapping for ty
be211c5 
Replace a stale `ty: ignore[no-matching-overload]` with an isinstance
guard so `settings` is a Mapping before `_copy_params_into`, clearing
the `invalid-argument-type` diagnostic from `make ty`.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
@ajslater
remove gitignore
d7570ec
@ajslater @claude
fix(test): call super().tearDownClass() in BaseTestImporter
a8aca7b 
Skipping super() left the class-level atomic from setUpTestData
uncommitted-but-unrolled-back, leaking the codex_init() "admin"
superuser into later test classes and causing UNIQUE constraint
failures (e.g. test_bookmark_filter_isolation) in full-suite runs.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
@ajslater ajslater merged commit 72fad68 into main Jun 15, 2026
4 checks passed
ajslater added a commit that referenced this pull request Jun 15, 2026
@ajslater
Squashed commit of the following:
5ad35f1 
commit 5e6f4dfc55ee4ad3f8fab48b7b33ffe02f17d8a4
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:26:21 2026 -0700

    update version 2.0.0a13

commit 59d607d6c886cccb093fbc07dea5a2ef97ab1dc0
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:25:51 2026 -0700

    test(onlinetag): exercise replay path in unmatched-write test

    The candidate carried issue_id 123, routing the resolve through the
    explicit-id fetch (real fetch_tags_by_explicit_id) instead of the
    replay session the test sets up and claims to verify. That crashed
    on the empty touched .cbz before any assertion ran. Drop the issue_id
    so the candidate falls back to the replay path guarded by _first_tags.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit dfa2349e8b821fbaa93b53fb4639bd2d66db8522
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:19:14 2026 -0700

    fix(mail): ASCII hyphen in noqa comments to dodge complexipy panic

    complexipy's extract_comment_marker slices noqa comment text by a
    fixed byte offset (16); a multi-byte em-dash straddling that offset
    lands mid-char and panics the Rust core, aborting the whole run.
    Swap the two em-dashes in the noqa comments for ASCII hyphens.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit 1799ce937559d62a667a0bf66f819f1593c93c5e
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:16:34 2026 -0700

    decomplexify test for radon

commit 6213c0afe641942e9e1d28f090970c6da524ceb4
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:14:19 2026 -0700

    remove unused ty ignore

commit dc59f1dc9560e177f9e0c36ea4552208014746d6
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:13:38 2026 -0700

    remove redundant definition

commit 5d7a204d5069783c8638bc8f76f84bf15840239a
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 23:10:36 2026 -0700

    Squashed commit of the following:

    commit 1274bad663bf382d76453e65a6fc935b8b3d4559
    Merge: a8aca7b33 72fad6861
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 22:57:33 2026 -0700

        Merge branch 'main' into develop

    commit 72fad6861b0af14731847b5444f097fdec699fda
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 22:06:27 2026 -0700

        v1.12.8 (#781)

    commit a8aca7b3341e7afdee8e99829be88d77e4d988c5
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:50:21 2026 -0700

        fix(test): call super().tearDownClass() in BaseTestImporter

        Skipping super() left the class-level atomic from setUpTestData
        uncommitted-but-unrolled-back, leaking the codex_init() "admin"
        superuser into later test classes and causing UNIQUE constraint
        failures (e.g. test_bookmark_filter_isolation) in full-suite runs.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit d7570ecf99e1a15449dec2f7a0550122b9d405df
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:16:59 2026 -0700

        remove gitignore

    commit be211c5724d5d7cc25ca367813fa34244097f372
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:12:03 2026 -0700

        fix(types): narrow OPDS redirect settings to Mapping for ty

        Replace a stale `ty: ignore[no-matching-overload]` with an isinstance
        guard so `settings` is a Mapping before `_copy_params_into`, clearing
        the `invalid-argument-type` diagnostic from `make ty`.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit 5dd3a0dd9feb18aa4b13c15b000c048befdeceef
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:09:05 2026 -0700

        update devenv

    commit e5a2a41881d35e2d8716bd6dfefdd64fa3792b5c
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:07:34 2026 -0700

        shfmt

    commit 1f867c1497e4edc14315d63c7eeb8c3e2b0f07bd
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:05:38 2026 -0700

        shfmt

    commit c5366787b56212e12a69c5aaab1776474962a31b
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 21:02:56 2026 -0700

        fix(bookmark): scope read/unread filters per user (v1.12.8)

        The UNREAD filter used Q(bookmark=None), a multi-valued relationship
        test for "comic has no bookmark from anyone" rather than "no bookmark
        for me". Once any user finished a comic it dropped out of every other
        user's (and anonymous visitors') unread view. UNREAD now negates the
        per-user finished predicate (~(my_filter & finished=True)), compiling
        to a per-user NOT-EXISTS subquery.

        Also harden get_my_bookmark_filter: an anonymous visitor with no
        established session_key resolved to `session_id IS NULL`, which matches
        every authenticated user's bookmarks. Return a match-nothing filter
        instead; a session key is minted only when a bookmark is written.

        Adds tests/test_bookmark_filter_isolation.py covering cross-user and
        anonymous isolation for READ/UNREAD/IN_PROGRESS.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit 541907b6708dfbbde197432960c6adc6dcf44b02
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 19:44:58 2026 -0700

        fix(frontend): convert v-dialog em widths to px so dialogs aren't slivers

        Vuetify 4's v-dialog defaults to location 'center center', which activates
        the connected overlay location strategy. That strategy caps width with
        parseFloat(props.maxWidth), so "20em" parses to 20 and the em is dropped,
        collapsing dialogs to a ~20px sliver. Convert every em-based max-width/
        min-width on a v-dialog (or forwarded via $attrs) to unitless px at 16px/em.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit 5918d0b621cd635f3de32f142d4dff7ec9c45242
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 19:41:25 2026 -0700

        move VPullToRefersh out of labs

    commit 57d5cf3888c2dc0aab55e46447b41a459d37003a
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 19:32:39 2026 -0700

        update devenv and deps

commit cd944f0e4cf7044ba8e72cb6f90539ca892adb72
Author: AJ Slater <[email protected]>
Date:   Sun Jun 14 18:25:16 2026 -0700

    Squashed commit of the following:

    commit b5874b67774af29f43603f01690a0ca97b43c46a
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 18:24:54 2026 -0700

        update deps

    commit e0a16dff8e66c357b2d2aa5e4ff3ea7decce4572
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 18:23:51 2026 -0700

        remove pytest-gitignore

    commit 35b84324300352499f3c455aa43201c3f475cd21
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 18:20:46 2026 -0700

        update version to 2.0.0a12

    commit e3e1167851960a42fef9a5ceba3b3b5470f7a56c
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 18:19:54 2026 -0700

        Fix basedpyright error on Folder FK access in moved-parents test

        The Folder model annotates parent_folder as ForeignKey explicitly, so
        django-stubs doesn't synthesize the implicit parent_folder_id attribute.
        Compare the FK object instead (Django compares models by pk). Pre-existing
        since ce1fd81ad; behavior unchanged.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit 0ee9c966f9d1dcfd84d36e6077778fd9354d9574
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 18:13:45 2026 -0700

        update deps

    commit 38f527b57ce30be0f162d7aae2bf51d938e3141b
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 18:13:39 2026 -0700

        ignore benchmark dir

    commit 7007c109accb1654c5b61580336d71cedfb5ce3f
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 17:56:39 2026 -0700

        Add online tagging status table with live progress and session controls

        Admin Tagging tab now shows a live status table for the online tagging
        batch: per-comic state, per-source rate-limit countdowns, batch ETA, and
        match-review status. Backed by a daemon-published session snapshot in the
        tagging cache, exposed via GET /api/v4/admin/tag-sessions/snapshot and
        refreshed over the existing task.progress websocket.

        - Snapshot: session_snapshot.py builds a capped, JSON-safe per-comic
          snapshot from OnlineTagOutcomeStats + pending prompts; published throttled
          from the runner, frozen inactive on finish, cleared at startup / janitor.
        - Session lifecycle: pause / resume / abort / dismiss controls in the table,
          wired to the online-tag store and admin endpoints.
        - Match-review fixes:
          - direct-id fetch for explicit picks (no rate-limit-prone re-search drift)
          - record user_matched / user_skipped outcomes, overlaid on the snapshot
          - client guard against resurrecting just-answered prompts in the dialog
          - optimistic local overlay so resolutions show during rate-limit stalls

        Tests: backend snapshot / resume / session-manager / janitor suites;
        frontend store + status-table specs. ruff/ty/eslint/prettier clean.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit ce1fd81adb4823ac454950ab66905a38e3f6d9e1
    Author: AJ Slater <[email protected]>
    Date:   Sat Jun 13 11:56:39 2026 -0700

        Fix ScribeThread crash on folder move under untracked parent

        _get_move_create_folders_one_layer returned bare path strings, but
        bulk_folders_create expects key tuples (MODEL_REL_MAP[Folder] == (("path",),
        ...)) and iterates each entry to extract the path. Bare strings were iterated
        character by character, producing single-char paths like PosixPath('K') whose
        stat() raised FileNotFoundError and crashed the ScribeThread during import.

        Wrap the path in a tuple to match the contract, and add a regression test
        covering a folder move whose destination parent isn't yet a tracked Folder.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit 265988a4b44ba7ebe9c810bedc11a42796853d90
    Author: AJ Slater <[email protected]>
    Date:   Sat Jun 13 11:30:44 2026 -0700

        Delete duplicate migration 0044 that broke test-DB setup

        The bookmark composite indexes and librarianstatus fields were folded
        into 0043 during the merge-fix commit (c31c61e6), but the orphaned
        0044_bookmark_comic_indexes.py was left behind with identical AddIndex
        operations. Re-running them aborted test-DB creation with
        "index bookmark_comic_user already exists" (415 test errors).

        makemigrations --check reports no changes; full suite 558 passed.

        Co-Authored-By: Claude Opus 4.8 <[email protected]>

    commit c31c61e6175d24d9e1dde1835c7235c367c251c9
    Author: AJ Slater <[email protected]>
    Date:   Sat Jun 13 02:00:11 2026 -0700

        fix some kind of merge issue

    commit 92e526940e96b6ea9679363159395270233e75f2
    Author: AJ Slater <[email protected]>
    Date:   Sat Jun 13 01:54:55 2026 -0700

        Squashed commit of the following:

        commit 27c7647f0ffdcec1a5ca178e47539c4ad060de05
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 01:54:29 2026 -0700

            refactor(complexity): clear radon cc rank-C findings

            bin/lint-complexity.sh reported three rank-C functions; extract helpers
            to drop each below the gate (radon cc --min C now empty).

            - AdminTagByIdView.post: pull primary-identifier parsing and merge
              extra-id resolution into _resolve_primary / _resolve_extra_ids.
            - OnlineTagSessionManager._apply_resolution: extract _handle_unresolved
              and _enqueue_resolved_write.
            - test_update_protagonist: extract repeated protagonist assertions into
              a helper (radon counts each assert as a branch).

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit 55bb659fcc8c02e73c20a34823652a4f81be1d56
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 01:46:04 2026 -0700

            refactor(migrations): squash 0044 into 0043

            Fold the merge_all_sources field into the ComicboxTaggingDefaults
            CreateModel in 0043 and drop the standalone 0044. makemigrations --check
            reports no drift.

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit 7a988737a63286cc85396346579e124e5fff8e5b
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 01:44:59 2026 -0700

            feat(onlinetag): merge metadata from all sources

            Add an opt-in "merge all sources" mode that queries every enabled online
            source per comic and merges the results (comicbox first_wins=False) for
            maximum tag completeness, instead of stopping at the first match.

            - Admin tagging defaults: merge_all_sources boolean (default off),
              settable as the default and overridable per scan.
            - Search dialog: per-scan toggle, gated on >=2 enabled sources; estimate
              + rate-limit warning multiply calls by source count (kept in sync with
              estimate.py) so the doubled API cost is visible.
            - By-ID: chips input accepting multiple URLs/ids; merging fetches two
              explicit ids (one per source) and merges them. Source select shows only
              when a token can't be auto-identified (bare number).

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit 8416c668c041cdbc56a16e56e357ce99aa3d9dae
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 00:34:56 2026 -0700

            update deps

        commit 473221d7a09403b1848f202d55902b71e31319c2
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 00:22:58 2026 -0700

            fix(lint): clear ruff/pyright findings; fix benchmark import path

            - tests: add docstrings, name magic estimate values, use pytest.raises
              and out-of-class exception message
            - onlinetag: collapse implicit f-string concatenation in log call
            - benchmark-import: correct mock_comics import (fixtures.* -> benchmark.*)
              which was a real runtime ImportError in library generation

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit 45a9fde6076a5b0b15be889769e9a55af4b0678a
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 00:13:49 2026 -0700

            refactor(migrations): squash 0044 + 0045 into 0043

            Fold the comic-leading composite Bookmark indexes (former 0044) and the
            LibrarianStatus eta/retry_at fields + JFR status-type choice (former
            0045) into the consolidated 0043 migration; delete both files.

            makemigrations --check reports no changes against the model state.

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit 22da6b9cd85615f3d7b8e7043818babac313d24e
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 00:04:35 2026 -0700

            update deps

        commit 6f27454623e43cccd95253d5dfe66471820c9349
        Author: AJ Slater <[email protected]>
        Date:   Sat Jun 13 00:04:15 2026 -0700

            fix(onlinetag): persist mid-scan prompt answers; live rate-limit countdowns

            Two online-tagging bugs plus a status-UX overhaul.

            Prompt answers given while a scan was running were lost on refresh: the
            response task sat in the queue behind the busy OnlineTagThread, so the
            cache kept the prompt. Now the drain loop removes answered prompts from
            the cache inline (race-free, single-threaded) and marks the fingerprint
            so the running scan stops re-persisting it; the network re-fetch + write
            is deferred until the scan releases the thread.

            A rate-limited lookup looked hung: collect_results could raise mid-pass
            (budget exhausted, network error) without ever calling finish(), leaving
            the status row frozen forever. finish() now runs in a finally.

            The rate-limit status was a static "rate limited ~57s" — a second
            subtitle writer clobbered the parseable one, and nothing refreshed it
            during comicbox's blocking sleep. Replaced with two absolute-timestamp
            countdowns the admin UI ticks down to live:
            - retry_at: time until the next online request, re-anchored per attempt.
            - eta: total time remaining, carrying forward the launcher dialog's
              estimate (ported verbatim to estimate.py), re-estimated on each comic
              completion and pushed out by rate-limit waits.

            Adds eta/retry_at to LibrarianStatus (migration 0045, which also folds in
            the pending status_type choices drift), plumbed through Status and
            StatusController.

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit eccf41bc603b118bdbd63f4e4918a30da87140ae
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 22:39:15 2026 -0700

            feat(tagging): admin-orderable online source priority

            defaultSources order is now run priority end to end. comicbox runs
            sources in the given order under first-wins, so the admin can choose
            which source is primary and which is fallback:

            - Tagging tab renders the source rows in defaultSources order with
              per-row up/down arrows; enabling a source appends it at lowest
              priority. A hint explains that the first matching source wins.
            - The launcher dialog normalizes its checkbox selection to the
              admin's priority order before starting a session (checkbox v-model
              records click order, not priority).
            - Serializers validate defaultSources / sources: known names only,
              deduped, order preserved (unknown source → 400).
            - explicit_id / session settings pass ordered tuples to comicbox's
              retyped OnlineLookupSettings.sources.

            Co-Authored-By: Claude Fable 5 <[email protected]>

        commit 02869ddf4d027c325637dd41081199c59f1141ad
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 17:06:25 2026 -0700

            add METRON_INFO to default formats to write

        commit a214c9c93cf679194beeb5c42eb66deea933a3da
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 14:58:36 2026 -0700

            update deps

        commit 65a294362171bded71547430dd51b322ea83dcef
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 14:58:08 2026 -0700

            don't make abort button red

        commit de49b5dd9529446ab34a7468bda694fae39c7a00
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 14:52:14 2026 -0700

            fix(frontend): remove duplicated Tag Write Errors sidebar item

            admin-menu.vue contained the same tag-write-errors CodexListItem twice
            (re-added in a later squash while the original stayed), so any error
            rendered two identical sidebar entries. One item per error kind; the
            counts live in the admin tables.

            Co-Authored-By: Claude Fable 5 <[email protected]>

        commit e0bd853d9ed1991859849e11d85d487409dc1194
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 13:43:19 2026 -0700

            fix(onlinetag): stop spurious rewrites and prompt loss in online tagging

            Three interacting bugs broke a single-comic tagging run:

            - Pass-1 and prompt-resolution writes now require result.matched
              (new comicbox flag): unmatched results carry the comic's merged
              existing metadata, so gating on truthy tags re-wrote files with
              no new information — which also triggered a re-import mid-session.

            - Pending prompts and tag-write errors move to a dedicated 'tagging'
              cache (codex.cache.tagging_cache). The session_cache docstring
              claimed key namespacing protected them from cache.clear(), but
              Django's file-based clear deletes everything: every import that
              changed anything (importer finish), Library/Group CRUD, and startup
              wiped prompts mid-answer, causing 'resolve_prompt: unknown prompt'.

            - Prompt resolution builds its replay session with defer_prompts=True
              so the preloaded resolution is actually consulted (comicbox only
              installs its cache-reading selector with defer or a handler;
              without it an ambiguous re-search fell through to comicbox's
              interactive CLI prompt inside the daemon). A fingerprint miss —
              the re-search returned different candidates — now re-queues the
              fresh deferred prompt instead of dying silently.

            Co-Authored-By: Claude Fable 5 <[email protected]>

        commit 9df7837bd3e3480e4fefd382f09db6d326cfadf4
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 11:45:22 2026 -0700

            feat(admin): auto-enable a tagging source when its credentials are first saved

            Saving credentials for a source that previously had none now adds it to
            default_sources via the auto-saving draft, so a freshly configured source
            works without a second checkbox click. Re-saving credentials for an
            already-configured source leaves the enable checkbox untouched.

            Co-Authored-By: Claude Fable 5 <[email protected]>

        commit fd5b9c0b6fb5fa0da079bae268e9e7553098b853
        Author: AJ Slater <[email protected]>
        Date:   Fri Jun 12 11:38:11 2026 -0700

            Squashed commit of the following:

            commit 9543008f24441b6b2992f0c818d019f42135f7df
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 10:21:24 2026 -0700

                dockerignore benchmark dir

            commit 00134a40fe1c37d96e64066735cb42cc03f463c8
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 10:20:24 2026 -0700

                mv fixutres to benchmark dir

            commit 451c5b7c1df5bffdbed68bdcfeaa96cdb6b47550
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 10:13:54 2026 -0700

                change import path for mock comics

            commit 21d9af8e54f4fb1bfe015d1cf96e8ad0c415d434
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 10:13:05 2026 -0700

                ignore fixtures dir

            commit 5c589f69d2a146ff1e839b095c1b563f73ed54ba
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 10:11:25 2026 -0700

                common subdir for mock & benchmark

            commit cd768965a544069fb623f9798a2ef7f472016586
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 10:01:01 2026 -0700

                move benchmark import out of ci

            commit d1a3a213d514317acfcb0dcf19b3a9163e2880b8
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 09:52:32 2026 -0700

                v2.0.0a10

            commit 43b6042bca9c546d49268a52002efa4db69651c4
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 09:51:47 2026 -0700

                update deps

            commit 02709b3d43119451ff177ecc6d0d3b9d76cfc634
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 02:39:57 2026 -0700

                fix typecheck and ty warnings

                Merge implicit f-string concatenations, add missing @override on
                LibrarianStatusCursorPagination.get_ordering, and replace a stale
                ty: ignore in SeeOtherRedirectError._get_query_params with a real
                Mapping isinstance narrowing.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 68c1d9fcbf2a5a2d7b4b11b94148151cb74a4555
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 02:28:58 2026 -0700

                remove bad news line

            commit 4b92f77b022b7f8392edb695bf4b5eb1871529ac
            Author: AJ Slater <[email protected]>
            Date:   Fri Jun 12 02:18:54 2026 -0700

                adapt to comicbox 4.0.0a3 API

                Per tasks/comicbox-handoff.md:

                - Bump comicbox pin to ~=4.0.0a3 (constructor no longer accepts or
                  needs logger=; comicbox logs through the host's loguru sinks).
                - Drop the logger= kwarg from every Comicbox() call site.
                - Catch comicbox.exceptions.ComicboxError instead of broad Exception
                  where the intent is "comicbox failed on this file": the reader page
                  endpoint 404s only on comicbox errors now, and tag-by-id fetch
                  failures surface in the admin Tagging-tab error panel instead of a
                  generic thread-crash log.
                - No workarounds to remove: TagWriter already drains bulk_write with
                  cancel=abort_event (which now actually works), the explicit-id
                  root-unwrap remains valid, and tests already build events with
                  keyword args.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 027fa1d8b2f79866dd0e37e00dfa45617db7b251
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 21:55:25 2026 -0700

                update deps

            commit 919fad63cf5790b82275e5f20207e38f65934a6c
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 21:48:25 2026 -0700

                update deps

            commit 147510b43c0d860f1d35e3b4d3420b48b62f7a3e
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 14:00:48 2026 -0700

                poller: apply discarded .only() projection in manual poll path

                _handle_pending_polls called qs.only(*_LIBRARY_ONLY) without reassigning,
                so full Library rows were fetched. With the projection applied, the
                library.save() after setting last_poll writes only loaded fields, which
                also avoids clobbering update_in_progress set concurrently by the importer.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit cc629e5baa28fc3e19da9ae40676a8df4b6248f4
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 07:52:01 2026 -0700

                admin: fix stale Last Poll column after websocket table refetch

                DateTimeColumn snapshotted its Date in data(), which runs once per
                component instance. When LIBRARY_CHANGED refetched the libraries
                table, Vue reused the row's component and only the dttm prop changed,
                so a new library's Last Poll stayed at the epoch (new Date(null))
                until a full page reload. Derive the Date as a computed instead.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 188dd22775d6beb284b6ddafb3629dcb0affbccb
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 00:18:16 2026 -0700

                edit news for brevity

            commit abad1fe2f68c2ba6899c056e9075d6d8e6555b1a
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 00:17:05 2026 -0700

                NEWS: browser performance entry for 2.0.0

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 994d6a586c249376827c1d72494071b4281528fa
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 00:15:27 2026 -0700

                news: v2.0.0 performance section for the import speedups

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit cbb07c980bd5bb317e7a506bec81ba8f66415163
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 00:06:03 2026 -0700

                cards: Max bookmark timestamp aggregate instead of JSON array + parse loop

                bookmark_updated_ats built a sorted, deduped JSON_GROUP_ARRAY of every
                matching bookmark timestamp per card, which the serializer parsed
                string-by-string with fromisoformat just to take the max (~50 cards x
                up to ~50 timestamps per browse response for heavy readers). A Max
                aggregate produces the same value with no JSON construction, per-group
                sort, transfer, or Python loop.

                The alias is distinct from bookmark_updated_at: when the primary sort
                is bookmark_updated_at ascending that alias already holds a Min
                aggregate, and the table-view collection branch annotates it
                independently. Responses verified byte-identical, incl. OPDS and
                bookmark-filtered listings.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 01a9394d8ecaaa22e72885b4257545a82220b3d4
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 00:06:03 2026 -0700

                breadcrumbs: batch the folder ancestor chain in one query

                The lazy parent_folder walk issued one SELECT per ancestor level — a
                depth-12 folder paid 11 round-trips per browse with cold cachalot, and
                every import invalidates the Folder table. Fetch all ancestors at once
                via the indexed materialized path (PurePath.parents prefixes scoped by
                library_id; verified equal to the FK chain on the real corpus, 5.6ms
                -> 1.7ms at depth 12).

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit d0c2f64c9dbb9cc796fe17a9408e0aaff4788733
            Author: AJ Slater <[email protected]>
            Date:   Thu Jun 11 00:06:03 2026 -0700

                mtime probe: plain aggregate over a pk-subquery; stop fragmenting the TTL key

                Two fixes to the collection mtime probe paid on every browse/head:

                - The probe annotated a per-row Greatest, GROUP BYed every column,
                  DISTINCTed and sorted all rows just to read one global MAX. Keep the
                  filtered queryset (demoted joins + FTS MATCH intact) as a
                  pk-subquery and aggregate over a fresh queryset (23.8ms -> 7.9ms at
                  18k, identical value). A plain .aggregate() on the filtered qs is
                  NOT equivalent — join demotion doesn't survive Django's aggregation
                  rewrite and FTS5 raises 'unable to use function MATCH'.

                - The 5s TTL cache key included the page number and order params,
                  which never affect the probed value — every page flip and order
                  toggle re-paid the probe (~20% of a bookmark-filtered page-flip
                  request). Key on user/model/collection/pks/filters only.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 578807ab0294391095e2f6b6fe3d1b1540f643a4
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 23:55:07 2026 -0700

                search: materialize the FTS match set once per browse request

                The raw comicfts MATCH re-executed the FTS5 scan in every statement
                carrying the search filter — counts, mtime probe, pagination pk query,
                main query, intersections: 5-6 scans per request, measured ~870ms of a
                1.3s search request at 17k matches.

                Browse/mtime requests now materialize the match pks once per request
                (cached_property) and bind a pk IN-list instead; cold-cachalot browse
                drops to exactly one MATCH execution. Scoped deliberately:

                - Rank-ordered requests (order_by=search_score) keep the raw MATCH —
                  ComicFTSRank needs it active in the scored statement.
                - choices/metadata keep the constant-size MATCH/subquery forms: their
                  statements repeat the filter per probe arm, and an inlined list
                  binds one variable per pk (SQLite caps statements at 32,766).
                - Match sets over 15,000 pks keep raw MATCH for the same reason — the
                  list may appear twice per statement (main query + cover subquery).

                The cover subquery consumes the swapped set directly instead of
                re-wrapping it in a membership sub-SELECT.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit e56b7eb2dd54cab402041de418832f865269540c
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 23:46:46 2026 -0700

                lint: format follow-ups for the order/cover perf commits

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit ca67645e4bf757568021e19781d9179ae599bda2
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 23:46:46 2026 -0700

                browser: move the last-route write off the read path; settings row fetched once

                Every browse GET re-fetched the settings row (two uncachable
                django_session probes) and synchronously rewrote
                SettingsBrowserLastRoute on navigation — a full-row UPDATE (rewriting
                created_at) that evicted cachalot's settings cache for everyone and
                stalled 1.85s behind the WAL writer lock during imports.

                - Navigation now queues a LastRouteUpdateTask to the librarian
                  bookmark thread (same machinery as page-turn bookmarks): the
                  aggregator keys on the settings pk, so rapid navigation collapses
                  to one filter().update() per flood window. Measured: navigation
                  under a held write transaction 1,850ms -> 67ms; same-route repeats
                  are fully write-free at 2 queries.
                - The settings instance is memoized per request keyed on
                  (model, client) — the save path reuses the loaded row, dropping the
                  second fetch and the second session probe.
                - The remaining synchronous saves (real settings changes via params,
                  PATCH) use update_fields so created_at is never rewritten.

                The persisted route now lags navigation by the writer's flood window
                (<=5s) — the same freshness contract page-turn bookmarks already have.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 94ef67e41554869cfe4c19494cea7e422ff5d30b
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 21:10:00 2026 -0700

                covers: batch cover mtimes post-pagination instead of a second subquery

                cover_mtime was a .values('updated_at') Subquery over the identical
                correlated cover queryset that already computes cover_pk — two full
                executions of the most expensive expression per card (measured 40-49%
                of the card query across publisher/folder/search pages).

                Drop the annotation; resolve the representative comics' updated_at
                after pagination with one indexed pk__in batch over the page's
                cover_pks and attach it to the cached instances (the serializer
                iterates the same objects). Custom-cover mtimes keep their cheap rowid
                subquery and still win the coalesce. _CoverMtimeCoalesce is gone with
                the annotation that needed it. Responses verified byte-identical.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 0a0546e2c92de8dd96773a37790980ea2bcb16b2
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 21:06:50 2026 -0700

                browser: zero-pad from page rows; skip dead card serialization in table mode

                - _get_zero_pad re-ran the whole ordered/annotated book query as a MAX
                  subquery (~20% of warm-request SQL) to derive one digit count.
                  Materialize the final page once (priming the result cache the
                  serializer reuses) and take the max issue_number off the page rows
                  in Python. The OPDS path keeps the aggregate variant.

                - Table mode serialized ~100 cards through BrowserCardSerializer and
                  then popped them from the result. Pop the card fields from
                  self.fields before to_representation instead (and symmetrically the
                  rows field in card mode) — same wire output, none of the per-card
                  work. Fixed the stale 'mobile fallback' comment claiming cards stay
                  populated.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit d1789eda2ac5a4351c3039ee42d4677db1c669c6
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 21:06:50 2026 -0700

                table mode: evaluate the collection page query once, not twice

                compute_collection_intersections read the page pks via
                values_list('pk'), a clone that re-executed the entire annotated/
                grouped/sorted collection query; the serializer then ran the original
                queryset again. Iterating the queryset itself primes its result cache,
                which the serializer reuses — the heaviest table-mode query now runs
                once per request.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 820bbc5617c1ac0e1b1a50d5d7c54a461b3b7c28
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 21:06:50 2026 -0700

                comic listings: stop aggregating order aliases; GROUP BY pk only

                Two GROUP BY-width fixes for Comic-model querysets:

                - _alias_filename aggregated Min/Max over the comic's own single path,
                  which turned the alias into an aggregate and dragged every selected
                  column into GROUP BY inside the listing and both cover subqueries.
                  A comic is its own file; use the bare expression (as
                  annotate_comic_extra_specials already did). A filename-ordered
                  folder browse measured 135s before, 0.6s after, identical rows.

                - The ids JsonGroupArray made Django compute GROUP BY over all ~50
                  selected Comic columns including TEXT blobs. Force GROUP BY id —
                  every other column is functionally dependent on the pk, so groups
                  are identical and SQLite sorts a one-column key (25-34% off the
                  books query). Not applied to the cover path, where the forced
                  literal column doesn't survive nested-subquery aliasing.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 14ecafa0781ea291822a0be894cf0a1cc390d4bd
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:56:40 2026 -0700

                browser: UNREAD filter must not hide comics other users finished

                The UNREAD arm was Q(bookmark=None) | (mine & unfinished):
                bookmark=None matches only comics with no bookmark from ANY user or
                session, so a comic finished by someone else (or a stale anonymous
                session) matched neither arm and silently vanished from my unread
                listing (reproduced: 50 foreign finished bookmarks shrank my unread
                count by 50).

                UNREAD now probes 'I have no finished bookmark on this comic' with a
                correlated NOT EXISTS scoped to the requesting user/session. It
                correlates on the queryset's comic pk, so on collection querysets it
                binds to the same joined comic row as the other filters in the
                combined .filter() call — group rows still require ONE comic that
                satisfies every condition (pinned by test).

                New comic-leading composite indexes (comic,user)/(comic,session) keep
                the probe off the user index: under stale sqlite_stat1 the planner
                flipped a user-scoped probe onto it — a measured 17s vs 27ms plan.
                EXPLAIN confirms 'SEARCH ... USING INDEX bookmark_comic_user
                (comic_id=? AND user_id=?)'. READ/IN_PROGRESS semantics unchanged;
                single-user responses verified byte-identical.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit cad52d79e8242e3b69e4cbd604757f2c7faab728
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:56:40 2026 -0700

                metadata: stop GET writing m2m through tables and corrupting tags

                _copy_m2m_intersections called ManyRelatedManager.set(qs, clear=True)
                for every real m2m field on the Comic path — obj is a live saved
                instance (the 'values dicts' comment was stale), so a GET /metadata
                DELETE+INSERTed up to 12 through tables, and a multi-comic selection
                permanently rewrote the first comic's tags to the selection's
                intersection (reproduced: 339 characters silently became 43).

                Serve the intersection queryset through the instance's prefetch cache
                instead: RelatedManager.get_queryset consults it keyed by field name,
                so the serializer reads the already-optimized queryset query-free.
                Response verified byte-identical; the 38-query identifier N+1 (the
                post-write re-read discarded select_related) disappears with it.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 71b52749a3b24a8493c6803f0c3cda04466280d4
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:19:11 2026 -0700

                browse: bind favorite collection code as str so cachalot can cache

                Binding the Collection StrEnum member itself made cachalot's
                parameter-type check raise UncachableQuery for every query the
                favorite subquery landed in — silently uncaching all browse queries.
                Repeat table-mode browses re-paid ~10s of SQL every time; with .value
                they serve from cache (16-22ms, 2-3 queries). The bound SQL value is
                identical either way.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit a200e1ba3b261508cbe2c22cce6712ad19a6e0ea
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:19:11 2026 -0700

                metadata: pk subquery instead of inlined IN lists; count the filtered relation

                Two fixes to the intersection queries behind GET /metadata:

                1. _get_comic_pks returned a materialized frozenset, inlining one bound
                   SQL variable per comic per union arm. With 12 m2m arms, any
                   collection over ~2,730 comics exceeded SQLite's 32,766-variable
                   limit and the endpoint returned 500 (reproduced at 4.2k, 8.5k and
                   17.8k comics). Passing the distinct .values(pk) queryset keeps the
                   statement constant-size (36 params); the same union measured 101ms
                   at 8,576 comics. num_comics is computed once via COUNT and threaded
                   to both consumers.

                2. _get_fk_intersection_query counted the hardcoded 'comic' reverse
                   m2m even when the filter traversed the main_*_in_comics FK
                   back-relations — a semantically wrong HAVING over the unfiltered
                   relation that also forced a separate unindexed join (main_team
                   measured 2,069ms -> 12ms counting the filtered relation).

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 5beb474835c7375f4b726eca63aa05f2ef8b9966
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:18:54 2026 -0700

                arc browse: demote story-arc chain joins to INNER for StoryArc queries

                With only codex_comic demoted, SQLite planned arc search browses from
                the arc side and built a cartesian product against the FTS match set:
                arcs root with a common search term measured 156s (main query 150.8s).
                Demoting codex_storyarcnumber and the M2M through table lets the plan
                start from the FTS/comic side (36ms). Scoped to StoryArc-model
                querysets: there the chain IS the collection relation, so NULL-extended
                rows are exactly the empty collections force_inner_joins exists to
                drop. On other models these tables appear as table-view annotation
                LEFT JOINs, where demotion would drop rows whose comics have no arcs.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 1434a5a771fc7445472cb52726f4e86c645df999
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:18:54 2026 -0700

                cover subqueries: join FTS rank only when ordering by search_score

                The direct fts_q join exists solely to populate codex_comicfts.rank
                for the cover subquery's search_score ORDER BY, but it was applied on
                every search regardless of order key — and the MATCH re-executed once
                per correlated cover evaluation. The pre-materialized fts_sq membership
                test already filters correctly. A name-sorted search browse measured
                135s before, 6.5s after (byte-identical rows); rank-ordered searches
                keep the join and pick the same top-ranked cover as before.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 5bdc927cc0db62775913f0e9b3ca5a6d86d18192
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:18:40 2026 -0700

                table mode: keep intersection-sort subqueries out of GROUP BY

                RawSQL.get_group_by_cols() returns [self], so the correlated
                intersection-sort subqueries landed in the GROUP BY of aggregated
                collection queries. Folder is the one collection model without a
                forced-GROUP-BY entry, so SQLite evaluated the subquery once per
                pre-aggregation joined row — at folders root the ancestor M2M fans out
                to every descendant comic (366s request; each of the two queries
                measured ~256s alone, 85ms after). Same mechanism as the
                _CoverMtimeCoalesce fix, recurring through a different expression type.

                _IntersectionSortRawSQL overrides get_group_by_cols() to return []:
                the subquery's only outer reference is the collection pk, which is
                always grouped, so grouping is unchanged (verified row-identical).

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 9f61f262b69fadc4cac463af5aafe88a9d248c9c
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 20:14:18 2026 -0700

                fix librarian status feed order: cursor pagination discarded the order_by

                The sidebar progress feed showed import phases scrambled — read next
                to the search statii, query below create. The viewset ordered active
                rows by (preactive, active, pk), but DRF CursorPagination re-orders
                the queryset with its own ordering, so the v4 admin refactor's plain
                pk default (545fc2aa9) silently replaced the registration-stamp order
                with row-creation order.

                LibrarianStatusCursorPagination orders the active poll by
                (preactive, active, pk) — matching start_many's padded registration
                stamps, with NULL preactive sorting first for rows started without
                pre-registration (the cover thread case). The Jobs tab history view
                keeps the pk default. The active feed is ~20 rows against a 200-row
                page, so the cursor filter never engages the nullable fields.

                Regression test creates status rows with pk order deliberately
                opposite the registration stamps and asserts both routes.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 396312c6fe9964b36f895dd6a2c8bb84757191c3
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 19:55:39 2026 -0700

                import timing: don't double-count sub-steps in the phase share total

                _run_phases reuses the timed_step helper instead of an inline
                perf_counter copy, and _log_phase_times sums only top-level phases —
                dotted "phase.step" entries already accrue inside their parents, so
                including them inflated the total and shrank every phase's share.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 693d9dbb84efd6846c0a3941b1f3ca35de650afc
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 19:47:39 2026 -0700

                metadata edit: full-size Protagonist label, live menu, stale-pick clear

                The Protagonist row inherited the old Main sub-row styling (smaller
                label, dimmed row) — it is a first-class row now, styled like its tag
                siblings. The choice menu already tracked the entered character/team
                names live (it is name-based, no pks involved); what was missing is
                that removing the picked name left the selection dangling — a watcher
                now clears it. A protagonist seeded from the DB but absent from the
                entered lists still survives load untouched.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 042c42db2fa0f1cce5deddb7ad3a3c2e8eea3c0a
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 19:29:28 2026 -0700

                metadata: unify main_character/main_team as a single Protagonist field

                Display: the starred main-chip machinery (MAIN_TAGS/markTagMain) is
                replaced by one "Protagonist" row above the tag table showing the
                mainCharacter and/or mainTeam chips — both only in the
                should-never-happen case where both are set. Each chip carries its own
                browser filter key (characters vs teams) via a per-item filter
                override in MetadataTags. En route this fixes mapTag's sticky filter:
                the loop reassigned the filter param, so every row after the first
                reused the first row's filter key.

                Edit: the two Main Character / Main Team sub-selects collapse into one
                Protagonist select offering the entered characters and teams; the pick
                is sent as the comicbox "protagonist" patch field on write, and the
                importer resolves it back to main_character or main_team (nulling the
                other) on re-import.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 632a238590f188ee249c4a8164a6936bcdfc779c
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 19:29:16 2026 -0700

                importer: stop clearing an unchanged protagonist on update imports

                The fk link phase set main_character/main_team to None whenever a
                comic's remaining LINK_FKS dict lacked a "protagonist" key. The query
                prune pops that key when the protagonist already matches the DB, so
                re-importing a comic with any *other* FK change silently NULLed its
                protagonist. Updates now only touch the protagonist fields when the
                metadata explicitly carried one (set-or-clear), matching the
                absent-means-untouched semantics of every other simple FK tag; creates
                keep both fields defaulted. Also: a name that is both a Character and
                a Team now links only main_character — both fields filled is invalid.

                Regression test drives query → prune → link with only scan_info
                changed and asserts the protagonist survives (it NULLed before).

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit b2e6423badc0d49e4f0d2cca8f7c59bbbc634ead
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 16:16:30 2026 -0700

                import perf: drop the timestamp updater's count aggregate (1.5s -> 0.1s)

                TimestampUpdater filtered every collection model through a
                Count(DISTINCT comic) alias to exclude empty collections — a per-row
                GROUP BY join aggregate costing ~1.5s per import even with zero
                deletions, in every import scenario.

                The count was redundant for two of the three OR branches: a
                comic__updated_at join match implies the collection has children, and
                force-updated pks were exempted from the child filter anyway. Only
                the custom-cover branch can match an empty collection, so it now
                carries its own join-presence test (comic__isnull=False) instead.
                Also replaces the fetch-instances + CASE-WHEN bulk_update round trip
                with a batched .update(updated_at=Now()) over the distinct pks.

                Also corrects the stale IMPORTER_FILTER_BATCH_SIZE comment: the
                binding constraint on OR chains is SQLITE_LIMIT_EXPR_DEPTH (1000),
                not the old 999-variable cap (modern SQLite allows 32766), so 900
                remains correct for the remaining chain users.

                Benchmark, 1000 mock comics: delete phase 1.46s -> 0.10s on fresh and
                below 0.1s on reimport (5.9s total). Re-stamp semantics covered by
                tests/importer/test_move_timestamps.py.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit bac8ee8af974bec78781f5c764e543f8429540a1
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 15:51:29 2026 -0700

                import perf: bulk_update only changed comic fields (update step 8x)

                update_comics passed all ~40 BULK_UPDATE_COMIC_FIELDS to bulk_update
                regardless of what changed — CASE-WHEN compilation and parse scale
                with rows x fields, so a force-reimport whose comics needed nothing
                but a stat refresh paid the full-width rewrite: 4.4s of the 7s
                remaining reimport time, almost all inside the two bulk_update
                manager calls.

                Collect the union of fields actually applied per batch (proposed md
                diffs + FK link fields) and update that union plus the
                presave-derived/timestamp set (stat, size, date, decade,
                age_rating_metron_index, updated_at). Fields outside
                BULK_UPDATE_COMIC_FIELDS are dropped as before.

                Benchmark, 1000 mock comics: forced reimport 10.0s -> 7.0s,
                update_comics sub-step 4.4s -> 0.55s. Fresh and noop unchanged.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 55401e6d372da3a9bf1c8b1cedeecbfd6fa06a27
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 15:46:01 2026 -0700

                import perf: narrow existing-row queries with a selector IN, not Q-OR chains

                Multi-key existing-row lookups (query phase query_existing_mds, link
                and create pk-map builders) built per-key AND/OR Q chains — one
                Django filter resolution per leaf and a giant OR for SQLite to parse,
                per 500-900-key batch. Profiling a fresh import showed ~40% of its
                time in this filter machinery.

                Both consumers already match exact key tuples in Python from the
                fetched values, so the chain only ever bounded the fetch. Replace it
                with a single indexed IN on the model's most selective non-null key
                rel (MODEL_SELECTOR_REL_MAP: imprint/series name, volume series name,
                credit person, identifier key, story-arc name, folder path). Superset
                rows — the same selector value under a different parent — are
                harmless extra map entries. Keys with a None selector value fall back
                to the old chain (selector columns are non-null, so normally none).

                Also: read phase sub-instrumented (extract vs aggregate) and the
                benchmark gained --profile (cProfile per scenario).

                Benchmark, 1000 mock comics: fresh 12.1s -> 9.4s (query phase 2.6s ->
                1.0s, link 2.5s -> 1.7s); reimport 11.8s -> 10.0s (query 3.4s ->
                1.6s). Tag/link counts byte-identical.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 610c51b67580006567b168e0e260c6c6097433d4
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 14:11:28 2026 -0700

                import perf: skip archive opens when the fs stat snapshot is unchanged

                The fs prefilter compared disk mtime against the stored *embedded*
                metadata mtime, which is older than the file's mtime for any archive
                copied into the library after tagging — so the common at-rest comic
                never skipped and every modified-event re-poll opened its archive
                (an unrar spawn per CBR) just for comicbox to no-op.

                Reuse the stat call to also compare disk (mtime, size) against the
                stored Comic.stat snapshot — the poller's own modified test — and
                skip the worker entirely on a match. Paths with no stored embedded
                mtime still always survive, so a library imported with the
                import-metadata flag off picks up tags after the flag turns on.

                Benchmark, 1000 mock comics: noop poll 1.0s -> 0.2s (7.6s at
                baseline); zero archive opens. Reimport unaffected (force path never
                populates the prefilter's mtime map).

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 06bec468cd68b423efb8f2b9da01be0f250c9d34
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 14:06:59 2026 -0700

                import perf: skip true no-op stat-only updates (noop poll 7.6s -> 1s)

                _envelope_deltas included metadata_mtime unconditionally, so every
                file whose tags comicbox skipped still routed through the stat-only
                update: the comic row rewritten, its covers purged and re-queued for
                generation, and the import counted as changed -> cache.clear() +
                LIBRARY_CHANGED broadcast. A modified-event burst over an unchanged
                library rewrote every row and regenerated every cover.

                Compare metadata_mtime against the stored value, and skip the
                stat-only routing entirely when the envelope is unchanged AND the
                on-disk (mtime, size) still matches Comic.stat — the same equality
                the poller's SnapshotDiff uses, so a skipped file is exactly one the
                poller would not re-emit (no stale-stat re-poll loop). Touched files
                (stat moved, content same) keep today's full stat-only path, which
                deliberately refreshes Comic.stat via presave.

                Benchmark, 1000 mock comics: noop poll 7.6s -> 1.0s with "No updates
                necessary" (the remaining ~0.9s is archive opens for the comicbox
                mtime check). Fresh and reimport unchanged.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 488af595b62d643f130098f38eb2d0d8760230cd
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 13:40:39 2026 -0700

                import perf: batch-resolve comic fk links (fresh import 3x)

                Comic creation resolved every FK link with objects.get() per
                (comic, field) — folder, two protagonist lookups and ~one per FK
                field, ~14 single-row SELECTs per created comic. create_comics was
                59% of a fresh import.

                prepare_fk_link_instance_maps now runs once per chunk after the FK
                create/update steps: one batched query per referenced field over the
                distinct key tuples (reusing the link phase's pk-map builders), plus
                in_bulk for instances. get_comic_fk_links stitches from the maps.
                Instances, not pks, so Comic(**md), the update path's
                setattr/default handling and presave's age_rating dereference are
                unchanged. Missing rows still raise DoesNotExist and skip the comic.
                The parent-folder map is also library-scoped now; the old
                Folder.objects.get(path=...) could MultipleObjectsReturned when two
                libraries shared a directory path.

                Also fixes a latent crash exposed by the new path:
                _build_pk_map_multi_key sorted key tuples that mix None with values
                (volume year/number_to, story-arc numbers) — TypeError when two keys
                first differ at a None position. Sort key now orders None last.

                Benchmark, 1000 mock comics: fresh 35.1s -> 12.1s; create_comics
                sub-step 21.7s -> 0.25s (map building 0.5s). Cumulative since
                baseline: fresh 2.9x, forced reimport 131.6s -> 11.8s (11x).

                Sub-step timing for create_and_update phases included
                (timed_step helper; dotted names excluded from share totals).

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 94dd32a4ab9a74e2fd250aea22340c41ff35b5e5
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 08:59:44 2026 -0700

                import perf: compare fk prune links as values tuples (query 12s -> 3.7s)

                The fk prune select_related'd only the directly referenced FK fields,
                then walked each key rel via getattr chains — every hop past depth one
                (volume -> series -> imprint -> publisher) lazy-loaded a single-row
                SELECT per comic, and the protagonist check lazy-loaded main_character
                and main_team likewise.

                Query every referenced field's key rels in one batched values_list and
                compare row slices against the proposed key tuples. Also makes the
                protagonist prunable when it's the only referenced field (it was
                skipped before because it isn't a concrete FK for select_related).

                Benchmark, 1000 mock comics: reimport 23.9s -> 13.2s, query phase
                12.2s -> 3.7s. With the m2m prune rewrite, forced reimport is 10x
                faster than the 131.6s baseline. Fresh and noop unchanged.

                Co-Authored-By: Claude Fable 5 <[email protected]>

            commit 55ceae11efd87c421f59b616ec65bf64882d4a8e
            Author: AJ Slater <[email protected]>
            Date:   Wed Jun 10 08:53:01 2026 -0700

                import perf: scan m2m prune off the through tables (10x)

                The m2m prune walked prefetched related objects per comic and read
                key attrs through instance getattr chains — for credits, identifiers
                and story-arc-numbers each FK in the key tuple lazy-loaded with a
                single-row SELECT per through row. A forced reimport of 1000 comics
                spent 121s (92%) in this phase.

                Scan the through tables directly instead: one values_list query per
                (field, batch) with the key attrs joined in SQL, pruned via set ops
                against the proposed links. No comic or related-model instances, no
                lazy FK loads. get_through_model moves from the link phase to const
                so the query phase can share it.

                Same pruning semantics (kept/deleted/FTS-existing/moved…
ajslater added a commit that referenced this pull request Jun 16, 2026
@ajslater
Squashed commit of the following:
07e695d 
commit 15a6b8c58bd31e2d3ce558db9a35c3bf3d4be623
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 21:15:28 2026 -0700

    version 2.0.0a14

commit 3ed935b45d978fe81118996285111a57fc340067
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 21:14:52 2026 -0700

    update deps

commit 4201d0747a3f43ee7aa5298ce229c07ab5b02eac
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 20:03:44 2026 -0700

    fix(lazy-import): stop the tag-icon hover loop on no-metadata comics

    Hovering a browser card's tag icon triggers a lazy metadata import gated
    on whether the comic "has metadata". Both the frontend hover gate and the
    lazy-import worker keyed off Comic.metadata_mtime — but that field is
    comicbox's *embedded* metadata mtime, which stays NULL forever for an
    archive with no metadata file. So a no-metadata comic always reported
    hasMetadata=false; after each import the browser refreshed, the card
    remounted, the cursor was still over the icon, and the import re-fired
    endlessly.

    Add a dedicated Comic.metadata_imported_at marker, stamped by a single
    path-scoped bulk UPDATE at import finish for every comic a forced/lazy
    pass touched (covers the SKIPPED no-metadata comics that never reach the
    per-comic write path), gated on force_import_metadata so ordinary watcher
    imports don't rewrite it. The worker gate and the has_metadata annotations
    (browser + reader) now read the marker, with metadata_mtime kept as a
    fallback. The field + backfill (COALESCE(metadata_mtime, updated_at), so
    existing comics don't regress to un-imported) is folded into the
    consolidated v2 migration 0043.

    metadata_mtime is left untouched — it is fed back to comicbox as the
    skip-re-extraction baseline, so overloading it would corrupt that protocol.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit a665bd183640a9abe33c6b083a784c3dcdfa4533
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 19:45:54 2026 -0700

    rename batch to session

commit 00f29ab33681432f6d3bb25e4bea46a5328dbd03
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 19:44:28 2026 -0700

    feat(onlinetag): refine match review dialog for pause/resume workflow

    - Expand the first match panel by default so the admin can act at once.
    - Replace the dead "Abort Session" button (abortSession never existed)
      with a "Pause" button wired to pauseSession, matching the Tagging tab.
    - Show "Cancel" instead of "Dismiss" while the session is unfinished
      (active or resumable), derived from the store snapshot.
    - Add a tooltip to the sidebar "Matches to Review" item.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit c1beab7ecbef3a5ce6fa765551d06141e6b03d39
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 02:06:47 2026 -0700

    feat(onlinetag): default delete_original to True

    Flip the ComicboxTaggingDefaults.delete_original field default and the
    0043 seed so newly tagged comics replace their originals by default.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit 910fa88989256d8ef6f01b2a501c9e4f81e25cfa
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 02:00:01 2026 -0700

    fix(tagging): scroll long comic names instead of truncating

    The Online Tagging Status table truncated long comic names with an
    ellipsis, putting them out of reach. Replace the path cell's
    overflow/text-overflow ellipsis with overflow-x: auto so a long name
    scrolls horizontally within its existing width.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit 05d44ba3b0f808cdb8a780da05f147333a486625
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 01:59:23 2026 -0700

    fix(onlinetag): enable tag-by-id from reader metadata screen

    The reader builds its metadata book from books.current, which carries a
    pk but no ids array. The online-tag dialog's idTaggable gate requires
    book.ids.length === 1, so the tag-by-id tab never appeared for a single
    comic opened from the reader, while browser card items (which include an
    ids array) worked. Derive ids from the pk to match the card shape.

    Co-Authored-By: Claude Opus 4.8 <[email protected]>

commit 8afde179d0dbc5a9e35d3523afb88e1c4daf9ad0
Author: AJ Slater <[email protected]>
Date:   Mon Jun 15 01:45:28 2026 -0700

    Squashed commit of the following:

    commit a6c315e224b2b6170d437099736afcbbf97e5479
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 23:31:51 2026 -0700

        update deps

    commit 98f55c7bf96be3d4c56c3fb12b10e95270a39cbc
    Merge: 5ad35f1cd 1274bad66
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 23:31:20 2026 -0700

        Merge branch 'develop' into pre-release

    commit 5ad35f1cd4d7b40935e4ca19c018f312bdef887b
    Author: AJ Slater <[email protected]>
    Date:   Sun Jun 14 23:26:35 2026 -0700

        Squashed commit of the following:

        commit 5e6f4dfc55ee4ad3f8fab48b7b33ffe02f17d8a4
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:26:21 2026 -0700

            update version 2.0.0a13

        commit 59d607d6c886cccb093fbc07dea5a2ef97ab1dc0
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:25:51 2026 -0700

            test(onlinetag): exercise replay path in unmatched-write test

            The candidate carried issue_id 123, routing the resolve through the
            explicit-id fetch (real fetch_tags_by_explicit_id) instead of the
            replay session the test sets up and claims to verify. That crashed
            on the empty touched .cbz before any assertion ran. Drop the issue_id
            so the candidate falls back to the replay path guarded by _first_tags.

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit dfa2349e8b821fbaa93b53fb4639bd2d66db8522
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:19:14 2026 -0700

            fix(mail): ASCII hyphen in noqa comments to dodge complexipy panic

            complexipy's extract_comment_marker slices noqa comment text by a
            fixed byte offset (16); a multi-byte em-dash straddling that offset
            lands mid-char and panics the Rust core, aborting the whole run.
            Swap the two em-dashes in the noqa comments for ASCII hyphens.

            Co-Authored-By: Claude Opus 4.8 <[email protected]>

        commit 1799ce937559d62a667a0bf66f819f1593c93c5e
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:16:34 2026 -0700

            decomplexify test for radon

        commit 6213c0afe641942e9e1d28f090970c6da524ceb4
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:14:19 2026 -0700

            remove unused ty ignore

        commit dc59f1dc9560e177f9e0c36ea4552208014746d6
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:13:38 2026 -0700

            remove redundant definition

        commit 5d7a204d5069783c8638bc8f76f84bf15840239a
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 23:10:36 2026 -0700

            Squashed commit of the following:

            commit 1274bad663bf382d76453e65a6fc935b8b3d4559
            Merge: a8aca7b33 72fad6861
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 22:57:33 2026 -0700

                Merge branch 'main' into develop

            commit 72fad6861b0af14731847b5444f097fdec699fda
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 22:06:27 2026 -0700

                v1.12.8 (#781)

            commit a8aca7b3341e7afdee8e99829be88d77e4d988c5
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:50:21 2026 -0700

                fix(test): call super().tearDownClass() in BaseTestImporter

                Skipping super() left the class-level atomic from setUpTestData
                uncommitted-but-unrolled-back, leaking the codex_init() "admin"
                superuser into later test classes and causing UNIQUE constraint
                failures (e.g. test_bookmark_filter_isolation) in full-suite runs.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit d7570ecf99e1a15449dec2f7a0550122b9d405df
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:16:59 2026 -0700

                remove gitignore

            commit be211c5724d5d7cc25ca367813fa34244097f372
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:12:03 2026 -0700

                fix(types): narrow OPDS redirect settings to Mapping for ty

                Replace a stale `ty: ignore[no-matching-overload]` with an isinstance
                guard so `settings` is a Mapping before `_copy_params_into`, clearing
                the `invalid-argument-type` diagnostic from `make ty`.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit 5dd3a0dd9feb18aa4b13c15b000c048befdeceef
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:09:05 2026 -0700

                update devenv

            commit e5a2a41881d35e2d8716bd6dfefdd64fa3792b5c
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:07:34 2026 -0700

                shfmt

            commit 1f867c1497e4edc14315d63c7eeb8c3e2b0f07bd
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:05:38 2026 -0700

                shfmt

            commit c5366787b56212e12a69c5aaab1776474962a31b
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 21:02:56 2026 -0700

                fix(bookmark): scope read/unread filters per user (v1.12.8)

                The UNREAD filter used Q(bookmark=None), a multi-valued relationship
                test for "comic has no bookmark from anyone" rather than "no bookmark
                for me". Once any user finished a comic it dropped out of every other
                user's (and anonymous visitors') unread view. UNREAD now negates the
                per-user finished predicate (~(my_filter & finished=True)), compiling
                to a per-user NOT-EXISTS subquery.

                Also harden get_my_bookmark_filter: an anonymous visitor with no
                established session_key resolved to `session_id IS NULL`, which matches
                every authenticated user's bookmarks. Return a match-nothing filter
                instead; a session key is minted only when a bookmark is written.

                Adds tests/test_bookmark_filter_isolation.py covering cross-user and
                anonymous isolation for READ/UNREAD/IN_PROGRESS.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit 541907b6708dfbbde197432960c6adc6dcf44b02
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 19:44:58 2026 -0700

                fix(frontend): convert v-dialog em widths to px so dialogs aren't slivers

                Vuetify 4's v-dialog defaults to location 'center center', which activates
                the connected overlay location strategy. That strategy caps width with
                parseFloat(props.maxWidth), so "20em" parses to 20 and the em is dropped,
                collapsing dialogs to a ~20px sliver. Convert every em-based max-width/
                min-width on a v-dialog (or forwarded via $attrs) to unitless px at 16px/em.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit 5918d0b621cd635f3de32f142d4dff7ec9c45242
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 19:41:25 2026 -0700

                move VPullToRefersh out of labs

            commit 57d5cf3888c2dc0aab55e46447b41a459d37003a
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 19:32:39 2026 -0700

                update devenv and deps

        commit cd944f0e4cf7044ba8e72cb6f90539ca892adb72
        Author: AJ Slater <[email protected]>
        Date:   Sun Jun 14 18:25:16 2026 -0700

            Squashed commit of the following:

            commit b5874b67774af29f43603f01690a0ca97b43c46a
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 18:24:54 2026 -0700

                update deps

            commit e0a16dff8e66c357b2d2aa5e4ff3ea7decce4572
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 18:23:51 2026 -0700

                remove pytest-gitignore

            commit 35b84324300352499f3c455aa43201c3f475cd21
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 18:20:46 2026 -0700

                update version to 2.0.0a12

            commit e3e1167851960a42fef9a5ceba3b3b5470f7a56c
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 18:19:54 2026 -0700

                Fix basedpyright error on Folder FK access in moved-parents test

                The Folder model annotates parent_folder as ForeignKey explicitly, so
                django-stubs doesn't synthesize the implicit parent_folder_id attribute.
                Compare the FK object instead (Django compares models by pk). Pre-existing
                since ce1fd81ad; behavior unchanged.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit 0ee9c966f9d1dcfd84d36e6077778fd9354d9574
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 18:13:45 2026 -0700

                update deps

            commit 38f527b57ce30be0f162d7aae2bf51d938e3141b
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 18:13:39 2026 -0700

                ignore benchmark dir

            commit 7007c109accb1654c5b61580336d71cedfb5ce3f
            Author: AJ Slater <[email protected]>
            Date:   Sun Jun 14 17:56:39 2026 -0700

                Add online tagging status table with live progress and session controls

                Admin Tagging tab now shows a live status table for the online tagging
                batch: per-comic state, per-source rate-limit countdowns, batch ETA, and
                match-review status. Backed by a daemon-published session snapshot in the
                tagging cache, exposed via GET /api/v4/admin/tag-sessions/snapshot and
                refreshed over the existing task.progress websocket.

                - Snapshot: session_snapshot.py builds a capped, JSON-safe per-comic
                  snapshot from OnlineTagOutcomeStats + pending prompts; published throttled
                  from the runner, frozen inactive on finish, cleared at startup / janitor.
                - Session lifecycle: pause / resume / abort / dismiss controls in the table,
                  wired to the online-tag store and admin endpoints.
                - Match-review fixes:
                  - direct-id fetch for explicit picks (no rate-limit-prone re-search drift)
                  - record user_matched / user_skipped outcomes, overlaid on the snapshot
                  - client guard against resurrecting just-answered prompts in the dialog
                  - optimistic local overlay so resolutions show during rate-limit stalls

                Tests: backend snapshot / resume / session-manager / janitor suites;
                frontend store + status-table specs. ruff/ty/eslint/prettier clean.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit ce1fd81adb4823ac454950ab66905a38e3f6d9e1
            Author: AJ Slater <[email protected]>
            Date:   Sat Jun 13 11:56:39 2026 -0700

                Fix ScribeThread crash on folder move under untracked parent

                _get_move_create_folders_one_layer returned bare path strings, but
                bulk_folders_create expects key tuples (MODEL_REL_MAP[Folder] == (("path",),
                ...)) and iterates each entry to extract the path. Bare strings were iterated
                character by character, producing single-char paths like PosixPath('K') whose
                stat() raised FileNotFoundError and crashed the ScribeThread during import.

                Wrap the path in a tuple to match the contract, and add a regression test
                covering a folder move whose destination parent isn't yet a tracked Folder.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit 265988a4b44ba7ebe9c810bedc11a42796853d90
            Author: AJ Slater <[email protected]>
            Date:   Sat Jun 13 11:30:44 2026 -0700

                Delete duplicate migration 0044 that broke test-DB setup

                The bookmark composite indexes and librarianstatus fields were folded
                into 0043 during the merge-fix commit (c31c61e6), but the orphaned
                0044_bookmark_comic_indexes.py was left behind with identical AddIndex
                operations. Re-running them aborted test-DB creation with
                "index bookmark_comic_user already exists" (415 test errors).

                makemigrations --check reports no changes; full suite 558 passed.

                Co-Authored-By: Claude Opus 4.8 <[email protected]>

            commit c31c61e6175d24d9e1dde1835c7235c367c251c9
            Author: AJ Slater <[email protected]>
            Date:   Sat Jun 13 02:00:11 2026 -0700

                fix some kind of merge issue

            commit 92e526940e96b6ea9679363159395270233e75f2
            Author: AJ Slater <[email protected]>
            Date:   Sat Jun 13 01:54:55 2026 -0700

                Squashed commit of the following:

                commit 27c7647f0ffdcec1a5ca178e47539c4ad060de05
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 01:54:29 2026 -0700

                    refactor(complexity): clear radon cc rank-C findings

                    bin/lint-complexity.sh reported three rank-C functions; extract helpers
                    to drop each below the gate (radon cc --min C now empty).

                    - AdminTagByIdView.post: pull primary-identifier parsing and merge
                      extra-id resolution into _resolve_primary / _resolve_extra_ids.
                    - OnlineTagSessionManager._apply_resolution: extract _handle_unresolved
                      and _enqueue_resolved_write.
                    - test_update_protagonist: extract repeated protagonist assertions into
                      a helper (radon counts each assert as a branch).

                    Co-Authored-By: Claude Opus 4.8 <[email protected]>

                commit 55bb659fcc8c02e73c20a34823652a4f81be1d56
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 01:46:04 2026 -0700

                    refactor(migrations): squash 0044 into 0043

                    Fold the merge_all_sources field into the ComicboxTaggingDefaults
                    CreateModel in 0043 and drop the standalone 0044. makemigrations --check
                    reports no drift.

                    Co-Authored-By: Claude Opus 4.8 <[email protected]>

                commit 7a988737a63286cc85396346579e124e5fff8e5b
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 01:44:59 2026 -0700

                    feat(onlinetag): merge metadata from all sources

                    Add an opt-in "merge all sources" mode that queries every enabled online
                    source per comic and merges the results (comicbox first_wins=False) for
                    maximum tag completeness, instead of stopping at the first match.

                    - Admin tagging defaults: merge_all_sources boolean (default off),
                      settable as the default and overridable per scan.
                    - Search dialog: per-scan toggle, gated on >=2 enabled sources; estimate
                      + rate-limit warning multiply calls by source count (kept in sync with
                      estimate.py) so the doubled API cost is visible.
                    - By-ID: chips input accepting multiple URLs/ids; merging fetches two
                      explicit ids (one per source) and merges them. Source select shows only
                      when a token can't be auto-identified (bare number).

                    Co-Authored-By: Claude Opus 4.8 <[email protected]>

                commit 8416c668c041cdbc56a16e56e357ce99aa3d9dae
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 00:34:56 2026 -0700

                    update deps

                commit 473221d7a09403b1848f202d55902b71e31319c2
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 00:22:58 2026 -0700

                    fix(lint): clear ruff/pyright findings; fix benchmark import path

                    - tests: add docstrings, name magic estimate values, use pytest.raises
                      and out-of-class exception message
                    - onlinetag: collapse implicit f-string concatenation in log call
                    - benchmark-import: correct mock_comics import (fixtures.* -> benchmark.*)
                      which was a real runtime ImportError in library generation

                    Co-Authored-By: Claude Opus 4.8 <[email protected]>

                commit 45a9fde6076a5b0b15be889769e9a55af4b0678a
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 00:13:49 2026 -0700

                    refactor(migrations): squash 0044 + 0045 into 0043

                    Fold the comic-leading composite Bookmark indexes (former 0044) and the
                    LibrarianStatus eta/retry_at fields + JFR status-type choice (former
                    0045) into the consolidated 0043 migration; delete both files.

                    makemigrations --check reports no changes against the model state.

                    Co-Authored-By: Claude Opus 4.8 <[email protected]>

                commit 22da6b9cd85615f3d7b8e7043818babac313d24e
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 00:04:35 2026 -0700

                    update deps

                commit 6f27454623e43cccd95253d5dfe66471820c9349
                Author: AJ Slater <[email protected]>
                Date:   Sat Jun 13 00:04:15 2026 -0700

                    fix(onlinetag): persist mid-scan prompt answers; live rate-limit countdowns

                    Two online-tagging bugs plus a status-UX overhaul.

                    Prompt answers given while a scan was running were lost on refresh: the
                    response task sat in the queue behind the busy OnlineTagThread, so the
                    cache kept the prompt. Now the drain loop removes answered prompts from
                    the cache inline (race-free, single-threaded) and marks the fingerprint
                    so the running scan stops re-persisting it; the network re-fetch + write
                    is deferred until the scan releases the thread.

                    A rate-limited lookup looked hung: collect_results could raise mid-pass
                    (budget exhausted, network error) without ever calling finish(), leaving
                    the status row frozen forever. finish() now runs in a finally.

                    The rate-limit status was a static "rate limited ~57s" — a second
                    subtitle writer clobbered the parseable one, and nothing refreshed it
                    during comicbox's blocking sleep. Replaced with two absolute-timestamp
                    countdowns the admin UI ticks down to live:
                    - retry_at: time until the next online request, re-anchored per attempt.
                    - eta: total time remaining, carrying forward the launcher dialog's
                      estimate (ported verbatim to estimate.py), re-estimated on each comic
                      completion and pushed out by rate-limit waits.

                    Adds eta/retry_at to LibrarianStatus (migration 0045, which also folds in
                    the pending status_type choices drift), plumbed through Status and
                    StatusController.

                    Co-Authored-By: Claude Opus 4.8 <[email protected]>

                commit eccf41bc603b118bdbd63f4e4918a30da87140ae
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 22:39:15 2026 -0700

                    feat(tagging): admin-orderable online source priority

                    defaultSources order is now run priority end to end. comicbox runs
                    sources in the given order under first-wins, so the admin can choose
                    which source is primary and which is fallback:

                    - Tagging tab renders the source rows in defaultSources order with
                      per-row up/down arrows; enabling a source appends it at lowest
                      priority. A hint explains that the first matching source wins.
                    - The launcher dialog normalizes its checkbox selection to the
                      admin's priority order before starting a session (checkbox v-model
                      records click order, not priority).
                    - Serializers validate defaultSources / sources: known names only,
                      deduped, order preserved (unknown source → 400).
                    - explicit_id / session settings pass ordered tuples to comicbox's
                      retyped OnlineLookupSettings.sources.

                    Co-Authored-By: Claude Fable 5 <[email protected]>

                commit 02869ddf4d027c325637dd41081199c59f1141ad
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 17:06:25 2026 -0700

                    add METRON_INFO to default formats to write

                commit a214c9c93cf679194beeb5c42eb66deea933a3da
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 14:58:36 2026 -0700

                    update deps

                commit 65a294362171bded71547430dd51b322ea83dcef
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 14:58:08 2026 -0700

                    don't make abort button red

                commit de49b5dd9529446ab34a7468bda694fae39c7a00
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 14:52:14 2026 -0700

                    fix(frontend): remove duplicated Tag Write Errors sidebar item

                    admin-menu.vue contained the same tag-write-errors CodexListItem twice
                    (re-added in a later squash while the original stayed), so any error
                    rendered two identical sidebar entries. One item per error kind; the
                    counts live in the admin tables.

                    Co-Authored-By: Claude Fable 5 <[email protected]>

                commit e0bd853d9ed1991859849e11d85d487409dc1194
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 13:43:19 2026 -0700

                    fix(onlinetag): stop spurious rewrites and prompt loss in online tagging

                    Three interacting bugs broke a single-comic tagging run:

                    - Pass-1 and prompt-resolution writes now require result.matched
                      (new comicbox flag): unmatched results carry the comic's merged
                      existing metadata, so gating on truthy tags re-wrote files with
                      no new information — which also triggered a re-import mid-session.

                    - Pending prompts and tag-write errors move to a dedicated 'tagging'
                      cache (codex.cache.tagging_cache). The session_cache docstring
                      claimed key namespacing protected them from cache.clear(), but
                      Django's file-based clear deletes everything: every import that
                      changed anything (importer finish), Library/Group CRUD, and startup
                      wiped prompts mid-answer, causing 'resolve_prompt: unknown prompt'.

                    - Prompt resolution builds its replay session with defer_prompts=True
                      so the preloaded resolution is actually consulted (comicbox only
                      installs its cache-reading selector with defer or a handler;
                      without it an ambiguous re-search fell through to comicbox's
                      interactive CLI prompt inside the daemon). A fingerprint miss —
                      the re-search returned different candidates — now re-queues the
                      fresh deferred prompt instead of dying silently.

                    Co-Authored-By: Claude Fable 5 <[email protected]>

                commit 9df7837bd3e3480e4fefd382f09db6d326cfadf4
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 11:45:22 2026 -0700

                    feat(admin): auto-enable a tagging source when its credentials are first saved

                    Saving credentials for a source that previously had none now adds it to
                    default_sources via the auto-saving draft, so a freshly configured source
                    works without a second checkbox click. Re-saving credentials for an
                    already-configured source leaves the enable checkbox untouched.

                    Co-Authored-By: Claude Fable 5 <[email protected]>

                commit fd5b9c0b6fb5fa0da079bae268e9e7553098b853
                Author: AJ Slater <[email protected]>
                Date:   Fri Jun 12 11:38:11 2026 -0700

                    Squashed commit of the following:

                    commit 9543008f24441b6b2992f0c818d019f42135f7df
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 10:21:24 2026 -0700

                        dockerignore benchmark dir

                    commit 00134a40fe1c37d96e64066735cb42cc03f463c8
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 10:20:24 2026 -0700

                        mv fixutres to benchmark dir

                    commit 451c5b7c1df5bffdbed68bdcfeaa96cdb6b47550
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 10:13:54 2026 -0700

                        change import path for mock comics

                    commit 21d9af8e54f4fb1bfe015d1cf96e8ad0c415d434
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 10:13:05 2026 -0700

                        ignore fixtures dir

                    commit 5c589f69d2a146ff1e839b095c1b563f73ed54ba
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 10:11:25 2026 -0700

                        common subdir for mock & benchmark

                    commit cd768965a544069fb623f9798a2ef7f472016586
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 10:01:01 2026 -0700

                        move benchmark import out of ci

                    commit d1a3a213d514317acfcb0dcf19b3a9163e2880b8
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 09:52:32 2026 -0700

                        v2.0.0a10

                    commit 43b6042bca9c546d49268a52002efa4db69651c4
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 09:51:47 2026 -0700

                        update deps

                    commit 02709b3d43119451ff177ecc6d0d3b9d76cfc634
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 02:39:57 2026 -0700

                        fix typecheck and ty warnings

                        Merge implicit f-string concatenations, add missing @override on
                        LibrarianStatusCursorPagination.get_ordering, and replace a stale
                        ty: ignore in SeeOtherRedirectError._get_query_params with a real
                        Mapping isinstance narrowing.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 68c1d9fcbf2a5a2d7b4b11b94148151cb74a4555
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 02:28:58 2026 -0700

                        remove bad news line

                    commit 4b92f77b022b7f8392edb695bf4b5eb1871529ac
                    Author: AJ Slater <[email protected]>
                    Date:   Fri Jun 12 02:18:54 2026 -0700

                        adapt to comicbox 4.0.0a3 API

                        Per tasks/comicbox-handoff.md:

                        - Bump comicbox pin to ~=4.0.0a3 (constructor no longer accepts or
                          needs logger=; comicbox logs through the host's loguru sinks).
                        - Drop the logger= kwarg from every Comicbox() call site.
                        - Catch comicbox.exceptions.ComicboxError instead of broad Exception
                          where the intent is "comicbox failed on this file": the reader page
                          endpoint 404s only on comicbox errors now, and tag-by-id fetch
                          failures surface in the admin Tagging-tab error panel instead of a
                          generic thread-crash log.
                        - No workarounds to remove: TagWriter already drains bulk_write with
                          cancel=abort_event (which now actually works), the explicit-id
                          root-unwrap remains valid, and tests already build events with
                          keyword args.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 027fa1d8b2f79866dd0e37e00dfa45617db7b251
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 21:55:25 2026 -0700

                        update deps

                    commit 919fad63cf5790b82275e5f20207e38f65934a6c
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 21:48:25 2026 -0700

                        update deps

                    commit 147510b43c0d860f1d35e3b4d3420b48b62f7a3e
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 14:00:48 2026 -0700

                        poller: apply discarded .only() projection in manual poll path

                        _handle_pending_polls called qs.only(*_LIBRARY_ONLY) without reassigning,
                        so full Library rows were fetched. With the projection applied, the
                        library.save() after setting last_poll writes only loaded fields, which
                        also avoids clobbering update_in_progress set concurrently by the importer.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit cc629e5baa28fc3e19da9ae40676a8df4b6248f4
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 07:52:01 2026 -0700

                        admin: fix stale Last Poll column after websocket table refetch

                        DateTimeColumn snapshotted its Date in data(), which runs once per
                        component instance. When LIBRARY_CHANGED refetched the libraries
                        table, Vue reused the row's component and only the dttm prop changed,
                        so a new library's Last Poll stayed at the epoch (new Date(null))
                        until a full page reload. Derive the Date as a computed instead.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 188dd22775d6beb284b6ddafb3629dcb0affbccb
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 00:18:16 2026 -0700

                        edit news for brevity

                    commit abad1fe2f68c2ba6899c056e9075d6d8e6555b1a
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 00:17:05 2026 -0700

                        NEWS: browser performance entry for 2.0.0

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 994d6a586c249376827c1d72494071b4281528fa
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 00:15:27 2026 -0700

                        news: v2.0.0 performance section for the import speedups

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit cbb07c980bd5bb317e7a506bec81ba8f66415163
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 00:06:03 2026 -0700

                        cards: Max bookmark timestamp aggregate instead of JSON array + parse loop

                        bookmark_updated_ats built a sorted, deduped JSON_GROUP_ARRAY of every
                        matching bookmark timestamp per card, which the serializer parsed
                        string-by-string with fromisoformat just to take the max (~50 cards x
                        up to ~50 timestamps per browse response for heavy readers). A Max
                        aggregate produces the same value with no JSON construction, per-group
                        sort, transfer, or Python loop.

                        The alias is distinct from bookmark_updated_at: when the primary sort
                        is bookmark_updated_at ascending that alias already holds a Min
                        aggregate, and the table-view collection branch annotates it
                        independently. Responses verified byte-identical, incl. OPDS and
                        bookmark-filtered listings.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 01a9394d8ecaaa22e72885b4257545a82220b3d4
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 00:06:03 2026 -0700

                        breadcrumbs: batch the folder ancestor chain in one query

                        The lazy parent_folder walk issued one SELECT per ancestor level — a
                        depth-12 folder paid 11 round-trips per browse with cold cachalot, and
                        every import invalidates the Folder table. Fetch all ancestors at once
                        via the indexed materialized path (PurePath.parents prefixes scoped by
                        library_id; verified equal to the FK chain on the real corpus, 5.6ms
                        -> 1.7ms at depth 12).

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit d0c2f64c9dbb9cc796fe17a9408e0aaff4788733
                    Author: AJ Slater <[email protected]>
                    Date:   Thu Jun 11 00:06:03 2026 -0700

                        mtime probe: plain aggregate over a pk-subquery; stop fragmenting the TTL key

                        Two fixes to the collection mtime probe paid on every browse/head:

                        - The probe annotated a per-row Greatest, GROUP BYed every column,
                          DISTINCTed and sorted all rows just to read one global MAX. Keep the
                          filtered queryset (demoted joins + FTS MATCH intact) as a
                          pk-subquery and aggregate over a fresh queryset (23.8ms -> 7.9ms at
                          18k, identical value). A plain .aggregate() on the filtered qs is
                          NOT equivalent — join demotion doesn't survive Django's aggregation
                          rewrite and FTS5 raises 'unable to use function MATCH'.

                        - The 5s TTL cache key included the page number and order params,
                          which never affect the probed value — every page flip and order
                          toggle re-paid the probe (~20% of a bookmark-filtered page-flip
                          request). Key on user/model/collection/pks/filters only.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 578807ab0294391095e2f6b6fe3d1b1540f643a4
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 23:55:07 2026 -0700

                        search: materialize the FTS match set once per browse request

                        The raw comicfts MATCH re-executed the FTS5 scan in every statement
                        carrying the search filter — counts, mtime probe, pagination pk query,
                        main query, intersections: 5-6 scans per request, measured ~870ms of a
                        1.3s search request at 17k matches.

                        Browse/mtime requests now materialize the match pks once per request
                        (cached_property) and bind a pk IN-list instead; cold-cachalot browse
                        drops to exactly one MATCH execution. Scoped deliberately:

                        - Rank-ordered requests (order_by=search_score) keep the raw MATCH —
                          ComicFTSRank needs it active in the scored statement.
                        - choices/metadata keep the constant-size MATCH/subquery forms: their
                          statements repeat the filter per probe arm, and an inlined list
                          binds one variable per pk (SQLite caps statements at 32,766).
                        - Match sets over 15,000 pks keep raw MATCH for the same reason — the
                          list may appear twice per statement (main query + cover subquery).

                        The cover subquery consumes the swapped set directly instead of
                        re-wrapping it in a membership sub-SELECT.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit e56b7eb2dd54cab402041de418832f865269540c
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 23:46:46 2026 -0700

                        lint: format follow-ups for the order/cover perf commits

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit ca67645e4bf757568021e19781d9179ae599bda2
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 23:46:46 2026 -0700

                        browser: move the last-route write off the read path; settings row fetched once

                        Every browse GET re-fetched the settings row (two uncachable
                        django_session probes) and synchronously rewrote
                        SettingsBrowserLastRoute on navigation — a full-row UPDATE (rewriting
                        created_at) that evicted cachalot's settings cache for everyone and
                        stalled 1.85s behind the WAL writer lock during imports.

                        - Navigation now queues a LastRouteUpdateTask to the librarian
                          bookmark thread (same machinery as page-turn bookmarks): the
                          aggregator keys on the settings pk, so rapid navigation collapses
                          to one filter().update() per flood window. Measured: navigation
                          under a held write transaction 1,850ms -> 67ms; same-route repeats
                          are fully write-free at 2 queries.
                        - The settings instance is memoized per request keyed on
                          (model, client) — the save path reuses the loaded row, dropping the
                          second fetch and the second session probe.
                        - The remaining synchronous saves (real settings changes via params,
                          PATCH) use update_fields so created_at is never rewritten.

                        The persisted route now lags navigation by the writer's flood window
                        (<=5s) — the same freshness contract page-turn bookmarks already have.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 94ef67e41554869cfe4c19494cea7e422ff5d30b
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 21:10:00 2026 -0700

                        covers: batch cover mtimes post-pagination instead of a second subquery

                        cover_mtime was a .values('updated_at') Subquery over the identical
                        correlated cover queryset that already computes cover_pk — two full
                        executions of the most expensive expression per card (measured 40-49%
                        of the card query across publisher/folder/search pages).

                        Drop the annotation; resolve the representative comics' updated_at
                        after pagination with one indexed pk__in batch over the page's
                        cover_pks and attach it to the cached instances (the serializer
                        iterates the same objects). Custom-cover mtimes keep their cheap rowid
                        subquery and still win the coalesce. _CoverMtimeCoalesce is gone with
                        the annotation that needed it. Responses verified byte-identical.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 0a0546e2c92de8dd96773a37790980ea2bcb16b2
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 21:06:50 2026 -0700

                        browser: zero-pad from page rows; skip dead card serialization in table mode

                        - _get_zero_pad re-ran the whole ordered/annotated book query as a MAX
                          subquery (~20% of warm-request SQL) to derive one digit count.
                          Materialize the final page once (priming the result cache the
                          serializer reuses) and take the max issue_number off the page rows
                          in Python. The OPDS path keeps the aggregate variant.

                        - Table mode serialized ~100 cards through BrowserCardSerializer and
                          then popped them from the result. Pop the card fields from
                          self.fields before to_representation instead (and symmetrically the
                          rows field in card mode) — same wire output, none of the per-card
                          work. Fixed the stale 'mobile fallback' comment claiming cards stay
                          populated.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit d1789eda2ac5a4351c3039ee42d4677db1c669c6
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 21:06:50 2026 -0700

                        table mode: evaluate the collection page query once, not twice

                        compute_collection_intersections read the page pks via
                        values_list('pk'), a clone that re-executed the entire annotated/
                        grouped/sorted collection query; the serializer then ran the original
                        queryset again. Iterating the queryset itself primes its result cache,
                        which the serializer reuses — the heaviest table-mode query now runs
                        once per request.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 820bbc5617c1ac0e1b1a50d5d7c54a461b3b7c28
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 21:06:50 2026 -0700

                        comic listings: stop aggregating order aliases; GROUP BY pk only

                        Two GROUP BY-width fixes for Comic-model querysets:

                        - _alias_filename aggregated Min/Max over the comic's own single path,
                          which turned the alias into an aggregate and dragged every selected
                          column into GROUP BY inside the listing and both cover subqueries.
                          A comic is its own file; use the bare expression (as
                          annotate_comic_extra_specials already did). A filename-ordered
                          folder browse measured 135s before, 0.6s after, identical rows.

                        - The ids JsonGroupArray made Django compute GROUP BY over all ~50
                          selected Comic columns including TEXT blobs. Force GROUP BY id —
                          every other column is functionally dependent on the pk, so groups
                          are identical and SQLite sorts a one-column key (25-34% off the
                          books query). Not applied to the cover path, where the forced
                          literal column doesn't survive nested-subquery aliasing.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 14ecafa0781ea291822a0be894cf0a1cc390d4bd
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:56:40 2026 -0700

                        browser: UNREAD filter must not hide comics other users finished

                        The UNREAD arm was Q(bookmark=None) | (mine & unfinished):
                        bookmark=None matches only comics with no bookmark from ANY user or
                        session, so a comic finished by someone else (or a stale anonymous
                        session) matched neither arm and silently vanished from my unread
                        listing (reproduced: 50 foreign finished bookmarks shrank my unread
                        count by 50).

                        UNREAD now probes 'I have no finished bookmark on this comic' with a
                        correlated NOT EXISTS scoped to the requesting user/session. It
                        correlates on the queryset's comic pk, so on collection querysets it
                        binds to the same joined comic row as the other filters in the
                        combined .filter() call — group rows still require ONE comic that
                        satisfies every condition (pinned by test).

                        New comic-leading composite indexes (comic,user)/(comic,session) keep
                        the probe off the user index: under stale sqlite_stat1 the planner
                        flipped a user-scoped probe onto it — a measured 17s vs 27ms plan.
                        EXPLAIN confirms 'SEARCH ... USING INDEX bookmark_comic_user
                        (comic_id=? AND user_id=?)'. READ/IN_PROGRESS semantics unchanged;
                        single-user responses verified byte-identical.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit cad52d79e8242e3b69e4cbd604757f2c7faab728
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:56:40 2026 -0700

                        metadata: stop GET writing m2m through tables and corrupting tags

                        _copy_m2m_intersections called ManyRelatedManager.set(qs, clear=True)
                        for every real m2m field on the Comic path — obj is a live saved
                        instance (the 'values dicts' comment was stale), so a GET /metadata
                        DELETE+INSERTed up to 12 through tables, and a multi-comic selection
                        permanently rewrote the first comic's tags to the selection's
                        intersection (reproduced: 339 characters silently became 43).

                        Serve the intersection queryset through the instance's prefetch cache
                        instead: RelatedManager.get_queryset consults it keyed by field name,
                        so the serializer reads the already-optimized queryset query-free.
                        Response verified byte-identical; the 38-query identifier N+1 (the
                        post-write re-read discarded select_related) disappears with it.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 71b52749a3b24a8493c6803f0c3cda04466280d4
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:19:11 2026 -0700

                        browse: bind favorite collection code as str so cachalot can cache

                        Binding the Collection StrEnum member itself made cachalot's
                        parameter-type check raise UncachableQuery for every query the
                        favorite subquery landed in — silently uncaching all browse queries.
                        Repeat table-mode browses re-paid ~10s of SQL every time; with .value
                        they serve from cache (16-22ms, 2-3 queries). The bound SQL value is
                        identical either way.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit a200e1ba3b261508cbe2c22cce6712ad19a6e0ea
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:19:11 2026 -0700

                        metadata: pk subquery instead of inlined IN lists; count the filtered relation

                        Two fixes to the intersection queries behind GET /metadata:

                        1. _get_comic_pks returned a materialized frozenset, inlining one bound
                           SQL variable per comic per union arm. With 12 m2m arms, any
                           collection over ~2,730 comics exceeded SQLite's 32,766-variable
                           limit and the endpoint returned 500 (reproduced at 4.2k, 8.5k and
                           17.8k comics). Passing the distinct .values(pk) queryset keeps the
                           statement constant-size (36 params); the same union measured 101ms
                           at 8,576 comics. num_comics is computed once via COUNT and threaded
                           to both consumers.

                        2. _get_fk_intersection_query counted the hardcoded 'comic' reverse
                           m2m even when the filter traversed the main_*_in_comics FK
                           back-relations — a semantically wrong HAVING over the unfiltered
                           relation that also forced a separate unindexed join (main_team
                           measured 2,069ms -> 12ms counting the filtered relation).

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 5beb474835c7375f4b726eca63aa05f2ef8b9966
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:18:54 2026 -0700

                        arc browse: demote story-arc chain joins to INNER for StoryArc queries

                        With only codex_comic demoted, SQLite planned arc search browses from
                        the arc side and built a cartesian product against the FTS match set:
                        arcs root with a common search term measured 156s (main query 150.8s).
                        Demoting codex_storyarcnumber and the M2M through table lets the plan
                        start from the FTS/comic side (36ms). Scoped to StoryArc-model
                        querysets: there the chain IS the collection relation, so NULL-extended
                        rows are exactly the empty collections force_inner_joins exists to
                        drop. On other models these tables appear as table-view annotation
                        LEFT JOINs, where demotion would drop rows whose comics have no arcs.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 1434a5a771fc7445472cb52726f4e86c645df999
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:18:54 2026 -0700

                        cover subqueries: join FTS rank only when ordering by search_score

                        The direct fts_q join exists solely to populate codex_comicfts.rank
                        for the cover subquery's search_score ORDER BY, but it was applied on
                        every search regardless of order key — and the MATCH re-executed once
                        per correlated cover evaluation. The pre-materialized fts_sq membership
                        test already filters correctly. A name-sorted search browse measured
                        135s before, 6.5s after (byte-identical rows); rank-ordered searches
                        keep the join and pick the same top-ranked cover as before.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 5bdc927cc0db62775913f0e9b3ca5a6d86d18192
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:18:40 2026 -0700

                        table mode: keep intersection-sort subqueries out of GROUP BY

                        RawSQL.get_group_by_cols() returns [self], so the correlated
                        intersection-sort subqueries landed in the GROUP BY of aggregated
                        collection queries. Folder is the one collection model without a
                        forced-GROUP-BY entry, so SQLite evaluated the subquery once per
                        pre-aggregation joined row — at folders root the ancestor M2M fans out
                        to every descendant comic (366s request; each of the two queries
                        measured ~256s alone, 85ms after). Same mechanism as the
                        _CoverMtimeCoalesce fix, recurring through a different expression type.

                        _IntersectionSortRawSQL overrides get_group_by_cols() to return []:
                        the subquery's only outer reference is the collection pk, which is
                        always grouped, so grouping is unchanged (verified row-identical).

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 9f61f262b69fadc4cac463af5aafe88a9d248c9c
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 20:14:18 2026 -0700

                        fix librarian status feed order: cursor pagination discarded the order_by

                        The sidebar progress feed showed import phases scrambled — read next
                        to the search statii, query below create. The viewset ordered active
                        rows by (preactive, active, pk), but DRF CursorPagination re-orders
                        the queryset with its own ordering, so the v4 admin refactor's plain
                        pk default (545fc2aa9) silently replaced the registration-stamp order
                        with row-creation order.

                        LibrarianStatusCursorPagination orders the active poll by
                        (preactive, active, pk) — matching start_many's padded registration
                        stamps, with NULL preactive sorting first for rows started without
                        pre-registration (the cover thread case). The Jobs tab history view
                        keeps the pk default. The active feed is ~20 rows against a 200-row
                        page, so the cursor filter never engages the nullable fields.

                        Regression test creates status rows with pk order deliberately
                        opposite the registration stamps and asserts both routes.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 396312c6fe9964b36f895dd6a2c8bb84757191c3
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 19:55:39 2026 -0700

                        import timing: don't double-count sub-steps in the phase share total

                        _run_phases reuses the timed_step helper instead of an inline
                        perf_counter copy, and _log_phase_times sums only top-level phases —
                        dotted "phase.step" entries already accrue inside their parents, so
                        including them inflated the total and shrank every phase's share.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 693d9dbb84efd6846c0a3941b1f3ca35de650afc
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 19:47:39 2026 -0700

                        metadata edit: full-size Protagonist label, live menu, stale-pick clear

                        The Protagonist row inherited the old Main sub-row styling (smaller
                        label, dimmed row) — it is a first-class row now, styled like its tag
                        siblings. The choice menu already tracked the entered character/team
                        names live (it is name-based, no pks involved); what was missing is
                        that removing the picked name left the selection dangling — a watcher
                        now clears it. A protagonist seeded from the DB but absent from the
                        entered lists still survives load untouched.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 042c42db2fa0f1cce5deddb7ad3a3c2e8eea3c0a
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 19:29:28 2026 -0700

                        metadata: unify main_character/main_team as a single Protagonist field

                        Display: the starred main-chip machinery (MAIN_TAGS/markTagMain) is
                        replaced by one "Protagonist" row above the tag table showing the
                        mainCharacter and/or mainTeam chips — both only in the
                        should-never-happen case where both are set. Each chip carries its own
                        browser filter key (characters vs teams) via a per-item filter
                        override in MetadataTags. En route this fixes mapTag's sticky filter:
                        the loop reassigned the filter param, so every row after the first
                        reused the first row's filter key.

                        Edit: the two Main Character / Main Team sub-selects collapse into one
                        Protagonist select offering the entered characters and teams; the pick
                        is sent as the comicbox "protagonist" patch field on write, and the
                        importer resolves it back to main_character or main_team (nulling the
                        other) on re-import.

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit 632a238590f188ee249c4a8164a6936bcdfc779c
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 19:29:16 2026 -0700

                        importer: stop clearing an unchanged protagonist on update imports

                        The fk link phase set main_character/main_team to None whenever a
                        comic's remaining LINK_FKS dict lacked a "protagonist" key. The query
                        prune pops that key when the protagonist already matches the DB, so
                        re-importing a comic with any *other* FK change silently NULLed its
                        protagonist. Updates now only touch the protagonist fields when the
                        metadata explicitly carried one (set-or-clear), matching the
                        absent-means-untouched semantics of every other simple FK tag; creates
                        keep both fields defaulted. Also: a name that is both a Character and
                        a Team now links only main_character — both fields filled is invalid.

                        Regression test drives query → prune → link with only scan_info
                        changed and asserts the protagonist survives (it NULLed before).

                        Co-Authored-By: Claude Fable 5 <[email protected]>

                    commit b2e6423badc0d49e4f0d2cca8f7c59bbbc634ead
                    Author: AJ Slater <[email protected]>
                    Date:   Wed Jun 10 16:16:30 2026 -0700

                        import perf: drop the timestamp updater's count aggregate (1.5s -> 0.1s)

                        TimestampUpdater filtered every collection model through a
                        Count(DISTINCT comic) alias to exclude empty collections — a per-row
                        GROUP BY join aggregate costing ~1.5s per import even with zero
                        deletions, in every import scenario.

                        The count was redundant for two of the three OR branches: a
                        comic__u…
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@ajslater