Lot 9: ask the name again, drop ShutdownBlockReasonCreate, measure two games

AGENTS.md

@@ -71,7 +71,17 @@ own; propose, with the reasoning, and let the maintainer decide.
 
 **Needs explicit approval, every time:** creating the public repository,
 pushing to it, publishing a release, changing the scheduled tasks or the
-configuration on the maintainer's machine, and any history rewrite.
+configuration on the maintainer's machine, and rewriting the history of
+`main`.
+
+**A branch's history is the agent's to rewrite until it reaches `main`**,
+unpushed or pushed for a pull request, and it should be: before a push,
+reduce a lot's commits to its logical steps. A fix to code written on the
+same branch, a second attempt, a page touched again for the same purpose
+belong in the commit they complete, not after it. A reader of `main` sees
+one commit per thing done; nobody returns to a state corrected an hour
+later, and the repository does not carry it. Force-pushing a pull
+request's branch is part of this; `main` is never rewritten.
 
 **The standing rule:** a change to what the user sees updates
 `docs/getting-started.md` and `docs/how-it-works.md` in the same commit. A
@@ -235,8 +245,6 @@ commands; do not.
 - A process started after `WM_QUERYENDSESSION` dies with
   `STATUS_DLL_INIT_FAILED`. Nothing can run a command at logoff; the session
   marker runs it at the next start instead.
-- The refinement's single timed attempt is a known margin, not a calibration
-  — `docs/design/09-robustness.md`.
 - `.git/HEAD` does not change on commit; `build.rs` watches the ref it names
   and `packed-refs` too, or the stamp goes stale.
 

CHANGELOG.md

@@ -35,6 +35,12 @@ a section is written.
   file: a short-lived helper opens them and takes that cost with it when it
   ends.
 
+### Fixed
+
+- A game still on its loading screen twenty seconds in no longer keeps its
+  launcher's name for the whole session: the watcher asks again every
+  twenty seconds, for two minutes at most.
+
 ### Removed
 
 - The *Writing your own* recipe: the starter configuration's comments and the

config.example.toml

@@ -68,9 +68,11 @@ stop_delay = "2s"
 # A title brings several processes that all match Windows' known game list: a
 # launcher stub, an anti-cheat service, and the game. The satellites usually
 # start first, so the name captured at the very start is often the wrong one.
-# Once, this far into a session, the per-process GPU counters are read and the
-# candidate that is actually rendering wins. Set to "0s" to skip it and keep
-# whichever process matched first.
+# This far into a session, the per-process GPU counters are read and the
+# candidate that is actually rendering wins. While there is no answer yet --
+# nothing rendering, no match -- they are read again at the same interval, six
+# times at most. Set to "0s" to skip it and keep whichever process matched
+# first.
 #
 # This only ever changes the *name*: detection never depends on it, and being
 # unable to read the counters simply means no answer.

docs/design/00-detection.md

@@ -62,6 +62,17 @@ back during the Starfield session did not release the reference, so the writer
 does not blink on focus changes — its lifetime is the session. Starfield was
 launched into the background and the writer still started, with another
 application in the foreground, so detection does not depend on focus.
+*Corrected 2026-09-25:* it can. Battlefield 6, launched and left loading
+behind another window, ran from 15:05:34 without a writer; the writer
+started at 15:08:02, the second the maintainer brought the game to the
+front, and stayed when the focus went away again. Windows' page says the
+writer is told when a game gets focus, and this is Windows starting it on
+that event. Starfield's run stands as measured; what differs between the
+two was not looked at. Either way it is Windows' verdict, and a game
+loading behind another window has no need of its gaming commands yet.
+The same day, Wreckfest 2's *Change settings* window started the writer
+too: it is `Wreckfest2.exe` itself, and Windows took it for the game, a
+14-second session.
 `GameDVR\LastGameActivity` matched the start timestamp to the second in both
 sessions. As a negative control, launching and focusing Notepad spawned nothing.
 
@@ -73,7 +84,16 @@ same afternoon. [Lot 1](01-console-watcher.md) has those figures and the
 decision they led to. *Corrected 2026-09-25:* what it waits for is the
 user's next keyboard or mouse input after the game has exited — some twenty
 seconds after it, and 2 h 22 min once, with nobody at the machine;
-[Lot 18](18-game-gone-nobody-there.md).
+[Lot 18](18-game-gone-nobody-there.md). *And later that day:* not twenty
+seconds. With the maintainer at the keyboard, six titles took 0.6 s to
+85 s. The writer is never released while nobody touches the machine; what
+decides the time otherwise is not known.
+
+Nor is the writer one per title. Two games running together share it:
+American Truck Simulator and Euro Truck Simulator 2, then Wreckfest 2 from
+Steam and Starfield from the Store, on 2026-09-25. Each pair had one writer
+process from the first game's start to after the second's exit
+([Lot 9](09-robustness.md)).
 
 ## The limit found in the field, 2026-09-20
 

docs/design/01-console-watcher.md

@@ -111,6 +111,9 @@ the party chat change nothing. The per-title theory stays dead; what the two
 sessions of each title above differed in is most likely when the maintainer
 next touched the machine, which was not noted then — inferred, not
 measured. [Lot 18](18-game-gone-nobody-there.md) has the details.
+*Corrected the same afternoon:* not twenty seconds. With the maintainer at
+the keyboard, six titles took 0.6 s to 85 s; only "never while nobody
+touches the machine" stands.
 
 The second sessions were measured from independent sources, since the log did
 not yet record the writer's exit: Steam's `gameoverlay_ui.txt` for Battlefield

docs/design/03-game-naming.md

@@ -54,12 +54,15 @@ allowed in exactly one narrow form, and its absence is a normal outcome.
 
 ## The refinement
 
-Once, a little way into the session (`identify_after`, 20 s by default), the
+A little way into the session (`identify_after`, 20 s by default), the
 candidates the Known Game List produced are ranked by their share of the 3D
 GPU engine over a one-second sample. The one rendering wins. The rule never
-promotes a process the list did not match; a single candidate, unreadable
-counters, or a game still on its loading screen all mean "keep the current
-name", and the log says which.
+promotes a process the list did not match; a single candidate that is the
+name in use, or the GPU confirming it, means "keep the current name", and
+the log says which. *Corrected 2026-09-25:* this said "once", and that
+unreadable counters and a loading screen kept the name too. Those, and no
+match at all, are no answer rather than a verdict, and the question is asked
+again at the same interval, six times at most — [Lot 9](09-robustness.md).
 
 **One match left is not nothing to say.** Measured 2026-09-15 on Battlefield 6:
 the EA anti-cheat *launcher* matches the install folder, starts before the
@@ -93,7 +96,9 @@ harmless.
 session, `status` read 0.0 % for every candidate; ten seconds later, 75 %. The
 single timed attempt landed just inside the window that makes it work. That is
 a margin, not a calibration — [Lot 9](09-robustness.md) has the design that
-would replace the timer.
+would replace the timer. *Since 2026-09-25* an attempt that reads nothing
+rendering is followed by another, so a loading screen costs twenty seconds
+rather than the name.
 
 ## Packaged titles are matched wherever they are installed
 

docs/design/05-windowless-watcher.md

@@ -131,4 +131,8 @@ next start — [Lot 9](09-robustness.md).
 
 No icon, no menu, nothing drawn. No `ShutdownBlockReasonCreate` either
 ("Restoring the fan profile…" on Windows' shutdown screen) — worth having,
-belongs with the rest of the shutdown work in Lot 9.
+belongs with the rest of the shutdown work in Lot 9. *Corrected
+2026-09-25:* not worth having. Windows never lists a program without a
+visible window on that screen, registering a reason is what would put it
+there, and nothing is restored at session end to announce; decided against
+in [Lot 9](09-robustness.md).