Der Feed lädt zehn Beiträge, und wie viele es sind, entscheidet das Mitglied (#1998)

Eine Ankunft trug vierzig Karten, von denen zehn im Dokument standen und die
restlichen dreißig der Socket unaufgefordert nachschob. Das automatische
Nachladen ist weg: eine Seite ist jetzt genau das, was gefragt wurde, und alles
Weitere holt der Druck auf "Mehr laden".

Wie lang diese Seite ist, sagt eine neue Voreinstellung (`Vutuv.Prefs`,
`feed_page_size`, voreingestellt 10, bis 250) — dieselbe Zahl für die Ankunft,
jedes Nachladen, einen Quellenwechsel und einen geöffneten Kalendertag, statt
der vier verschiedenen, die hier standen. Eine Karte kostet echte Serverzeit
(am 05.09.2026 auf der Produktion gemessen, verschränktes A/B über je zehn
Abrufe: 188 ms für ein Dokument mit zehn Karten gegen 101 ms für dieselbe Seite
ohne), und abwägen kann das nur, wer beides erlebt: die Wartezeit und die
Knopfdrücke. Zu ändern unter der Zeitleiste selbst und auf der neuen Seite
/settings/feed.

Der Kalendertag verliert dabei seine eigene Hundertergrenze. Sie hat nur für
die Tage zwischen der Seitengröße und hundert etwas bewirkt — und das waren
genau die, für die ein `?day=`-Link 300 bis 850 ms und 1,3 MB kostete; wer den
ganzen Tag will, hat den Knopf dafür weiter darunter.

Diesen Text hat ein KI-Agent in meinem Namen geschrieben. Ich weiß, dass das
problematisch ist.

Claude-Session: https://claude.ai/code/session_01NmkMzEZBiSXhjGiYMtVpfw

Author: wintermeyer

.claude/rules/design.md

@@ -341,43 +341,45 @@ whom being seen decides whether they come back. Following somebody here
 the click read as if it had undone something — and only a fresh draw retires
 them.
 
-**Three page sizes, and every difference is deliberate.** A mount loads **40**
-entries (`@first_page_size`), every "Load more" after it **20** (`@page_size`),
-and a source switch **10** (`@filter_page_size`). The arrival page is the one
-page nobody asked for, so it is the one that has to carry the reader past the
-first few scrolls without a round trip; an older page is fetched while they are
-still reading and can afford to be half of it; and a switch is a wait with
-nothing on screen at all, where twenty rendered cards are the bulk of the second
-it takes on a slow line, for a screen that holds three or four.
-
-**The arrival is forty cards, but the HTML document carries ten**
-(`@first_render_size`), and the socket appends the other thirty as soon as it
-connects (`:fill_arrival`, which is `append_older_page/2` — the same act as
-pressing "Load more", so the dedup rule has one owner). Rendering a card costs
-about **10 ms** of server time on production, measured two ways on 2026-08-31 —
-a twelve-day regression across `?day=` volumes and a twenty-run paired A/B
-against an empty day, agreeing on 10.2 and 10.0 — so the forty were roughly
-400 ms of the ~660 ms the browser waited for its first byte, spent drawing
-thirty-seven cards below a fold that shows two or three. Ten is the same number
-`@filter_page_size` uses and for the reason written there: it is what a screen
-holding three or four cards needs in order not to run out. What changed is only
-that an *arrival* is now judged by the same standard as a switch.
-
-This is **not** the lazy discovery rail of #1229, which was reverted the same
-day because content popping in where the reader was looking read as slowness:
-the fill lands below the tenth card, off-screen, and a reader who never scrolls
-never learns it happened. A fill is owed to the page that asked for it
-(`:owes_fill?`) and is dropped if the reader has meanwhile switched sources or
-opened a calendar day — appending an older page to a day would spill posts from
-before it into that day. `more?` comes
-from the same query in all three cases, so a short page still knows there is
-more and the button below picks the rest up. The `/feed.md|txt|json|xml`
-siblings take the arrival page as *their* page size, every page of them, and
-they read it off `Feed.first_page_size/0` rather than a mirrored constant. The
-**size** is all they share: the document is deliberately the whole feed, with no
-source filter applied, because it carries none of the switches the member has
-here and narrowing it by one they cannot see would hand an agent a document with
-no way to ask for the rest.
+**One page size, and the reader owns it.** The arrival, every "Load more", a
+source switch and an opened calendar day all ask for the same number:
+`Vutuv.Prefs.feed_page_size/1`, a member preference defaulting to **10** and
+bounded at 5 to 250, read once in `mount_feed/3` and parked on the socket
+(`page_size/1`), so nothing within one visit can disagree about it. It replaced
+four constants (40 for a mount, 20 for a "Load more", 10 for a switch, a flat
+100 for a calendar day), each with its own argument — and every one of those
+arguments was really an argument about *waiting*. A card costs real server time:
+**188 ms** for a ten-card `/feed` document against **101 ms** for the same page
+carrying none, measured on production on 2026-09-05 in a paired A/B of ten runs
+each (an earlier measurement on 2026-08-31, before the hashtag cache, put a card
+at ~10 ms). The only person who can weigh that wait against pressing a button
+again is the one doing both, so they set it — from the chip row under the
+timeline (`page_size_chips/1`, which stores the choice and reloads the timeline
+at the new length) or on `/settings/feed`.
+
+**The document carries the whole arrival, and nothing arrives behind the
+reader's back.** There is no socket fill: what a page holds is what was asked
+for, and what comes after it is asked for by pressing "Load more". (An earlier
+release split the arrival — forty cards, of which the HTML held ten and the
+socket appended thirty on connect — which is what the one number replaced.)
+`more?` comes from the same query in every case, so a page always knows whether
+there is more and the button below picks the rest up.
+
+**A calendar day is a page like any other.** It used to be fetched whole up to a
+hundred, on the argument that a day is a bounded thing the reader asked to see;
+since the limit is an upper bound, a day smaller than the reader's page still
+arrives whole, so that argument only ever reached the days between their number
+and a hundred — and those are exactly the ones that made a `/feed?day=` link
+cost 300 to 850 ms and 1.3 MB. "Load the whole day" (capped at `@day_all_limit`,
+1,000) is still under a busy one for the reader who wants all of it.
+
+The `/feed.md|txt|json|xml` siblings take that preference as *their* default page
+size, capped at the 100 an explicit `?limit=` is clamped to, so a member's
+browser setting cannot build a bigger document than any caller may request. The
+**size** is all they share with the HTML page: the document is deliberately the
+whole feed, with no source filter applied, because it carries none of the
+switches the member has here and narrowing it by one they cannot see would hand
+an agent a document with no way to ask for the rest.
 
 **A post that arrives while somebody is reading is a fourth page, of one.** It is
 drawn into the timeline at the moment it arrives (`queue/2` — the row on top, the

docs/architecture/posts-and-feed.md

@@ -18,16 +18,21 @@ they cannot drift. If an editable area is not on it, it does not exist.
 | Profile | Basics & photos, Username, Experience, Education, Certificates & licenses, Language skills, Tags, Organizations |
 | Contact details | Email addresses, Phone numbers, Addresses, Websites & links, Social media profiles, Messengers |
 | Notifications & feed | Notifications, Muted words & tags, Search & replace, Feed languages, Tags you follow, Find your contacts, Saved searches |
-| Appearance | Language & display, Data-saving mode |
+| Appearance | Language & display, Posts in your feed, Data-saving mode |
 | Privacy | Visibility, Blocked members, Automatic post deletion, Fediverse |
 | Account | Sign-in & security, Account activity, Import, Export, Organizations, Apps & API, Delete account (red) |
 
 *Appearance* is the newest group and holds what vutuv looks and costs like
-to this member — the interface language, dates, maps and post display on one
-row, data-saving mode on the other — as opposed to what the site holds about
-them. *Language & display* used to sit under Account, which had reached its
-eight rows and left the data-saving switch nowhere to go but the feed group,
-where it was not about the feed.
+to this member — the interface language, dates, maps and post display on the
+first row, how long the feed is on the second, data-saving mode on the third —
+as opposed to what the site holds about them. *Language & display* used to sit
+under Account, which had reached its eight rows and left the data-saving switch
+nowhere to go but the feed group, where it was not about the feed. *Posts in
+your feed* (`Vutuv.Prefs`' `feed_page_size`, 10 by default and up to 250) landed
+here for the same arithmetic — "Notifications & feed" is at its eight rows — and
+it belongs beside data-saving mode anyway: both are the member deciding what a
+page costs them to load. The chip row under the timeline writes the same
+column, so most members never open this page.
 
 Each row is a map with `:key` (the sidebar's active state and the hub's entry
 counts), `:label`, `:path`, a `:hint` line saying what is inside, `:terms`

docs/architecture/settings-and-account.md

@@ -340,6 +340,10 @@ defmodule Vutuv.Accounts.User do
     # `Vutuv.Prefs.get/2`, never straight off the struct.
     field(:feed_tab_ticker?, :boolean)
     field(:feed_tab_ticker_seconds, :integer)
+    # How many posts one feed page holds for this member — the arrival, every
+    # "Load more", a source switch, an opened calendar day. nil = inherit the
+    # installation default; read through `Vutuv.Prefs.feed_page_size/1`.
+    field(:feed_page_size, :integer)
     # The same quote one surface further out (issue #1681): while this member's
     # vutuv tab sits in the background, a new post pages its first line through
     # the browser tab's title. nil = inherit the installation default; read
@@ -580,7 +584,7 @@ defmodule Vutuv.Accounts.User do
   # :email_confirmed? is NOT here either: it flips only via the login-PIN path
   # (Accounts.activate_user/1, its own narrow cast) — castable, it would let a
   # registration self-activate without ever proving control of an email.
-  @optional_fields ~w(noindex? noai? notification_emails? dm_email_each_message? dm_email_delay_minutes email_on_endorsement? email_on_follower? email_on_reference_check? newsletter_emails? saved_search_emails? cv_update_notifications? thread_notifications? browser_notifications? show_online_status? show_mastodon_feed? mastodon_clients? show_code_stats? fediverse_followers? fediverse_reactions? fediverse_replies? also_known_as_input map_google? map_openstreetmap? map_apple? default_map_service post_lines_desktop post_lines_mobile post_hyphenate_desktop post_hyphenate_mobile notification_post_lines like_attribution? headline employment_status employment_status_visibility desired_salary_min desired_salary_currency desired_salary_period desired_salary_visibility desired_workplace_types first_name last_name middle_name nickname honorific_prefix honorific_suffix name_pronunciation gender birthdate birthdate_visibility locale date_region time_zone tag_list auto_post_deletion? auto_post_deletion_after_days auto_post_deletion_keep_photos? auto_post_deletion_keep_answered? auto_post_deletion_keep_bookmarked? auto_post_deletion_delete_replies? auto_post_deletion_min_likes auto_post_deletion_min_bookmarks auto_post_deletion_min_reposts feed_foreign_posts feed_languages feed_tab_ticker? feed_tab_ticker_seconds browser_tab_teaser? low_bandwidth?)a
+  @optional_fields ~w(noindex? noai? notification_emails? dm_email_each_message? dm_email_delay_minutes email_on_endorsement? email_on_follower? email_on_reference_check? newsletter_emails? saved_search_emails? cv_update_notifications? thread_notifications? browser_notifications? show_online_status? show_mastodon_feed? mastodon_clients? show_code_stats? fediverse_followers? fediverse_reactions? fediverse_replies? also_known_as_input map_google? map_openstreetmap? map_apple? default_map_service post_lines_desktop post_lines_mobile post_hyphenate_desktop post_hyphenate_mobile notification_post_lines like_attribution? headline employment_status employment_status_visibility desired_salary_min desired_salary_currency desired_salary_period desired_salary_visibility desired_workplace_types first_name last_name middle_name nickname honorific_prefix honorific_suffix name_pronunciation gender birthdate birthdate_visibility locale date_region time_zone tag_list auto_post_deletion? auto_post_deletion_after_days auto_post_deletion_keep_photos? auto_post_deletion_keep_answered? auto_post_deletion_keep_bookmarked? auto_post_deletion_delete_replies? auto_post_deletion_min_likes auto_post_deletion_min_bookmarks auto_post_deletion_min_reposts feed_foreign_posts feed_languages feed_tab_ticker? feed_tab_ticker_seconds feed_page_size browser_tab_teaser? low_bandwidth?)a
 
   # The ages the automatic post deletion offers (issue #1255), in days. A fixed
   # list rather than a free number field on purpose: this setting deletes
@@ -831,6 +835,13 @@ defmodule Vutuv.Accounts.User do
       greater_than_or_equal_to: notification_post_lines_min(),
       less_than_or_equal_to: notification_post_lines_max()
     )
+    # The feed's page size, bounded by its registry entry so nobody stores a
+    # feed of 4000 cards. A cleared field stays nil = inherit, like the counts
+    # above.
+    |> validate_number(:feed_page_size,
+      greater_than_or_equal_to: Prefs.pref!(:feed_page_size).min,
+      less_than_or_equal_to: Prefs.pref!(:feed_page_size).max
+    )
     |> validate_inclusion(:dm_email_delay_minutes, @dm_email_delay_values)
     |> validate_inclusion(:employment_status, @employment_statuses)
     |> validate_inclusion(:employment_status_visibility, @visibilities)

lib/vutuv/accounts/user.ex

@@ -89,6 +89,29 @@ defmodule Vutuv.Prefs do
       values: ~w(original translate hide),
       group: :feed
     },
+    # How many posts one feed page holds: the arrival, every "Load more", a
+    # source switch and an opened calendar day all ask for this number.
+    #
+    # Ten because that is a screenful three or four times over, and because an
+    # arrival is a wait with nothing on screen yet — a card costs real server
+    # time (measured on production 2026-09-05: a ten-card /feed document ~188 ms
+    # against ~101 ms for the same page with none), so the default is what a
+    # reader needs rather than what they might eventually scroll to. The
+    # ceiling is 250 for the reader who would rather wait once than press a
+    # button twenty times; the floor is 5, below which the button IS the page.
+    #
+    # Its own group rather than `:feed`, because a group is what a reset link
+    # clears: /settings/feed_languages resets `:feed` wholesale, and a member
+    # putting their languages back to the site default must not silently lose
+    # the page size they chose somewhere else.
+    %Pref{
+      key: :feed_page_size,
+      type: :integer,
+      default: 10,
+      min: 5,
+      max: 250,
+      group: :feed_size
+    },
     # The same idea one surface further out (issue #1681): while the vutuv tab
     # sits in the background, a new post pages its first line through the
     # browser tab's title instead of only putting a dot there. Off leaves the
@@ -188,6 +211,9 @@ defmodule Vutuv.Prefs do
   def label(:feed_foreign_posts),
     do: Gettext.gettext(VutuvWeb.Gettext, "Posts in other languages")
 
+  def label(:feed_page_size),
+    do: Gettext.gettext(VutuvWeb.Gettext, "Posts loaded at once")
+
   def label(:low_bandwidth?), do: Gettext.gettext(VutuvWeb.Gettext, "Low-bandwidth mode")
 
   def label(:browser_tab_teaser?),
@@ -209,6 +235,13 @@ defmodule Vutuv.Prefs do
         "What your feed does with posts outside your chosen languages: show them as they are, translate them for you, or hide them. Posts that declare no language always show."
       )
 
+  def hint(:feed_page_size),
+    do:
+      Gettext.gettext(
+        VutuvWeb.Gettext,
+        "How many posts your feed shows before the “Load more” button, and how many that button adds. More posts mean a longer wait for the page."
+      )
+
   def hint(:low_bandwidth?),
     do:
       Gettext.gettext(
@@ -271,6 +304,7 @@ defmodule Vutuv.Prefs do
   @doc "The human label of a pref group."
   def group_label(:post_display), do: Gettext.gettext(VutuvWeb.Gettext, "Posts")
   def group_label(:feed), do: Gettext.gettext(VutuvWeb.Gettext, "Feed")
+  def group_label(:feed_size), do: Gettext.gettext(VutuvWeb.Gettext, "Feed length")
   def group_label(:browser_tab), do: Gettext.gettext(VutuvWeb.Gettext, "Browser tab")
   def group_label(:bandwidth), do: Gettext.gettext(VutuvWeb.Gettext, "Bandwidth")
   def group_label(:privacy), do: Gettext.gettext(VutuvWeb.Gettext, "Privacy")
@@ -493,6 +527,32 @@ defmodule Vutuv.Prefs do
   """
   def low_bandwidth?(user), do: get(user, :low_bandwidth?)
 
+  @doc """
+  How many posts one page of this member's feed holds.
+
+  A named predicate for the same reason as `low_bandwidth?/1`: the feed, its
+  size control and the agent-format siblings all ask, and a page size spelled
+  out at four call sites is a page size that drifts.
+  """
+  def feed_page_size(user), do: get(user, :feed_page_size)
+
+  # The sizes the two controls offer. Round numbers a reader picks from at a
+  # glance rather than a slider over 5..250: the question is "a screenful or the
+  # whole morning", not a precise count.
+  @feed_page_size_steps [10, 25, 50, 100, 250]
+
+  @doc """
+  The offered feed sizes, with `current` folded in when it is not one of them.
+
+  An admin may set any installation default inside the pref's bounds, so the
+  five round steps are not the whole value space. Showing a member on 20 only
+  the five would leave no chip marked and make the nearest press look like a
+  no-op — so the value they actually hold is always one of the choices.
+  """
+  def feed_page_size_steps(current) do
+    [current | @feed_page_size_steps] |> Enum.uniq() |> Enum.sort()
+  end
+
   @doc """
   Drop every boolean-pref param that is not ticked, so the column stays NULL.