* feat(scroll): report a panel that cannot reach its refresh cap, and suggest one it can hold
Scroll speeds are solved against display.hardware.limit_refresh_rate_hz,
which is only a ceiling. A panel that cannot reach it still moves whole
pixels per frame, but every scroll runs slow by the shortfall and the
"smooth" ladder is the cap's, not the panel's. A user rig (Pi 4, 2x128x64,
adafruit-hat-pwm, pwm_bits 9, gpio_slowdown 5) measured 107.6-113.1 Hz under
a 120 Hz cap: 60 px/s ran at 55, and nothing said why.
- scroll_config: refresh_shortfall() (more than 3% under the planned rate),
holdable_cap() (a multiple of 10, 5% under the measurement, since the
measurement is the fast end of an uncapped panel's drift), and
describe_refresh_shortfall().
- FrameTimingRecorder.plan_refresh(): once the measured period has held for
three trusted windows, a shortfall is logged once as a warning naming the
cap to use. DisplayManager calls it only for a real panel, not the
emulator or the fallback canvas. The stats file records
planned_refresh_hz (additive).
- GET /api/v3/config/refresh-rate, plus a hint under the Display tab's
Limit Refresh Rate field (js/pages/display.js) with a button that fills in
the suggested cap.
- _panel_refresh_hz (behind the Vegas slider's advice) ignores a measurement
written under a different cap, so a changed cap stops being advised from
the old rate before the display restarts.
Verified on ledpi with a temporary 200 Hz cap: the warning logged about a
minute after the restart ("about 132 Hz ... Set Limit Refresh Rate to
120 Hz"), the endpoint returned the same shortfall, and the Display tab
showed the hint; its button filled in 120. ledpi was restored afterwards.
Rebased onto main after the Display tab became an ES-module page (#771); the
hint moved from inline script into display.js, with a jsdom test.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(scroll): count only agreeing windows toward the refresh shortfall; ignore a null planned rate
Review fixes on #759.
- A window the period estimate rejects (more than MAX_REFRESH_DROP faster
than the adopted period) no longer counts toward the shortfall check, and
it restarts the run. After a loaded start fixed a slow period, later
windows at the real, faster rate were rejected yet still counted, so the
warning could name the slow rate against a cap the panel was meeting. It
now needs REFRESH_CHECK_WINDOWS consecutive windows that agree with the
period.
- _panel_refresh_hz treats a stats file whose planned_refresh_hz key is
present but null as no measurement. The emulator and the fallback canvas
write it that way (DisplayManager never plans a refresh for them), and
their frame rate says nothing about the cap. A file with no such key (an
older display) keeps the old behaviour.
Three new tests fail on the previous code and pass now.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
A collection during interpreter shutdown called GcMonitor after the
module's `time` global was torn down, printing "Exception ignored while
calling GC callback ... 'NoneType' object has no attribute
'perf_counter'" at the end of service and test runs.
- GcMonitor binds its clock and sys.is_finalizing at construction and
does nothing once the interpreter is finalizing.
- install_gc_monitor() unregisters it with atexit; new
uninstall_gc_monitor().
- DisplayManager.cleanup() (reached from SIGTERM via run()'s finally)
unregisters it alongside the frame recorder.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A GcMonitor in src/common/frame_timing.py, installed once per process from gc.callbacks by the display manager (and render_bench), counts collections and seconds per generation, the longest, and those of 20 ms or more. A long one tags the next presented frame 'gc' in record(), so frame_soak shows its late rate under 'after work'; the stats file gains an additive 'gc' block printed as a 'Garbage collection' line; and a Render stall dump says when a long collection ran inside the stall. Diagnostic only: nothing tunes, freezes or disables the collector.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mid-scroll, update_display() no longer checksums every frame: it asks is_currently_scrolling() once per frame and reuses the answer, hashes only when dirty tracking can skip a static frame, and the preview snapshot asks its policy first and hashes only when a write or touch could follow (decide() is monotone, pinned by a property test). With the preview open the snapshot is written at most once a second (VIEWER_INTERVAL 1.0 s, was 0.2 s; the SSE stream re-read it once a second, so four encodes in five went unread); the stream now polls its mtime every 0.25 s (VIEWER_POLL_INTERVAL), so the preview stays about as fresh. The PNG is written at compress_level=1. --preview soaks are not comparable across this change.
Merged with #716: _scan_segments takes the frame's one scrolling answer and #716's static-handover pass-through, as soaked on ledpi (A B B A, 20 min each: main 0.118% / 0.113% late, with #716 and this 0.107% / 0.104%).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A static plugin screen that follows a scroller no longer starts with the ticker's lagging rows on scan-compensated panels, and the 1 Hz loop's second frame is no longer recorded as a ~1 s mid-scroll freeze / Render stall. The display controller calls DisplayManager.end_scroll_for_static_screen() before a static screen's first display() (clears the scan history; _scan_segments passes its frames through in one swap) and set_scrolling_state(False) after it; the scroller's hold stays until then, so late-frame counts are unchanged. A screen's first frame is tagged 'handover': gaps of 250 ms or more before it go to the additive handover_freezes (frame_soak prints 'Handover gaps'), not freezes. The display thread is named display-<plugin id>. The WiFi notice and the schedule-off blank are not covered yet (docs list them as a follow-up).
ledpi A B B A soak (20 min each, --preview): main 0.118% / 0.113% late with 6 / 3 freezes; with this and #717 0.107% / 0.104% late, 0 freezes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* perf(timing): say which render-thread work a late frame followed
The soak already says how often a moving frame reached the panel late, but
not what the render thread was doing just before it. Vegas does two kinds of
work there between frames -- building its strip (compose, extend) and, with
live elements, patching changed pixels into it -- and deciding whether either
is affordable needs their own numbers.
- FrameTimingRecorder.note_op(kind, nbytes) tags the next presented frame.
Totals gain op_frames, late_op_frames, op_freezes and op_bytes per kind;
aggregate() still takes frames without ops. The file schema is unchanged.
- Vegas tags compose and every strip extension (with the bytes it copied).
- frame_soak prints an "after work" table: frames, late %, freezes and MB
moved per kind, only when something tagged its work.
- render_bench gains --strip-screens (Vegas-sized strips), --patch-bytes /
--patch-every / --patch-where (in-place column writes, as a live element
update does) and --extend-every-screens / --extend-width (append + trim on
a fixed cadence that holds the strip's width).
No runtime behaviour changes: this is the measurement gate for live Vegas
elements.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(changelog): note the frame-op attribution and bench modes
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* ci: mypy ratchet -- keep type-clean modules clean
mypy-clean.txt lists the 71 modules under src/ that type-check clean;
scripts/check_types.py runs mypy (--follow-imports=silent) on exactly
those files and fails on any error or a missing/unsorted/duplicate entry.
A new "Type check (mypy ratchet)" CI job runs it with mypy 1.20.2 and
pinned stubs; the manual pre-commit mypy hook now runs the same script
(a local hook, so mypy sees the installed requirements like CI does).
35 modules were made clean with annotation-only fixes: hints, typing.cast,
TYPE_CHECKING imports, implicit-Optional defaults made explicit, and
annotations widened (never guards removed) where mypy called a defensive
isinstance check unreachable. No runtime behaviour change.
mypy.ini: numpy and orjson are treated as Any (follow_imports=skip, also
for stubs). numpy 2.3+ stubs use 3.12 `type` statements that mypy won't
parse at python_version 3.10, and orjson is optional, so seeing its stubs
made the result depend on whether it was installed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore: annotate check_types.py's list-form mypy subprocess
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
src/common/frame_timing.py times every frame the display presents, whoever drew it, and writes cumulative counters to /dev/shm. scripts/frame_soak.py grades a running service (late frames, freezes, where the time goes) and scripts/render_bench.py the hardware and render path alone. A stall watchdog logs the stacks behind any scroll held up for 250 ms or more (LEDMATRIX_STALL_WATCHDOG_MS lowers that). See docs/SCROLL_PERFORMANCE.md, "Soaking a rig".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
