tract runs every model on one core on every platform. Measured against
ONNX Runtime's providers on the MagicPad 2 and the reference desktop:
ORT CPU alone is 3-10x, the Hexagon at int8 runs the detectors in
1-3 ms, TensorRT is ~2x the CUDA provider. NNAPI, XNNPACK, WebGPU and
CUDA int8 were tried and excluded with the numbers that excluded them.
The spec keeps the build C-free: ort::set_api takes a table from a
dlopened runtime or from ort-tract, chosen once per process. Rungs
are chosen by building a real session, cached until an input changes,
and compiled engines are built in the background after the first
frame. The embedder stays f32 everywhere; int8 detectors are a
distinct model_id and are gated on a recall measurement.
Read and measured, not built. The bare 512 generator exports at a fixed
shape and loads under tract with nothing unsupported; at f32 on the
desktop CPU it takes 7.4 s per 512×512 tile, which puts a full-resolution
fill of the fixture's border at ten minutes. The three routes that would
make it viable are recorded, with the quarter-resolution fill the cheapest
and Hexagon int8 the one the model was designed for.
Picking a chip stored the choice for the merge and changed nothing on
screen — the chip did not even highlight, since the selected property
was never written back. Now the pick is reflected, and the job, waiting
for its decision, takes a Preview request, draws the alignment on the
chosen surface at proxy cost and reports again; the drain puts the new
picture and its size up. Auto is the surface the field of view suggests.
Also:
The largest rectangle inside the frames' coverage is found a row at a
time — a histogram of consecutive covered rows and a stack pass per row —
so the composite is never held to be measured (FR-MRG-11). It is written
as DefaultCropOrigin/DefaultCropSize (FR-MRG-4): the file opens on the
picture, the border is still in it, and resetting the crop shows it.
rawler reports the crop as the picture, which the test checks.
FR-MRG-4 records the question raised the same day — fill the border
rather than crop it — as open: a non-generative fill through the heal,
or a generative inpainter with its licence and weights. Neither decided.
derived_from and merge are top-level sidecar fields (FR-MRG-6): one line
per source in order, and how the composite was made. A build that
predates them keeps the lines as unknown and writes them back. The job
writes the sidecar beside the composite and stages it with its own
record when the composite goes through the outbox.
DARKROOM_START_MERGE=a.CR2,b.CR2 lands on the merge page at startup with
the job running on local files, on the model of DARKROOM_START_IDENTITY,
for looking at the page where synthetic clicks do not reach it. The fetch
and the start are shared with the grid's button.
panorama.md §11 records what exists, the fixture's figures, and the six
things still open, auto-crop first.
dr_ui::merge is the orchestration with no interface in it: decode each
frame to sensor data and build its graph as a session would (orientation,
lens profile); render each through the camera-space tap at proxy size and
detect keypoints there, so the alignment is measured in the undistorted
frame the tiles are rendered in; align; solve one gain per frame from the
proxies' overlaps; draw the aligned set in colour for the page; then wait.
Nothing is written until a Decision arrives (FR-MRG-1). The merge writes
a linear DNG through the outbox with a destination record, so the drain
puts it beside its sources on a folder library and a server alike, and
the library rescans (FR-MRG-3).
merge.slint is the page, on the import page's model: the alignment
table with a failed frame named on its row and the button held off
(FR-MRG-5), the preview, the projection choice, Stop and Back. A
"Merge to panorama" button joins the grid's selection bar at two frames.
Headless, the example produces the fixture's 22 993 x 5 980 DNG in 45 s
on the reference desktop, exposures balanced across the stop of drift.
compose_camera_linear composes the fused pass with an empty operation
list, the file's orientation as the baseline, a view rect for the tile,
and a store of rgba32float. On the GPU, render_camera_linear is the only
entry that accepts it: it fills the profile uniforms neutral — unit white
balance, identity matrix, curve off — so what lands in the texture is the
sensor's numbers after the lens warp and nothing else (FR-MRG-2). A third
bind-group layout carries the format, as the linear one does, and the
readback is generalised to any pixel width for the f32 copy.
Thirty-two bits because the composite is written back at the sensor's
scale: a 14-bit sensor has 16 384 steps to white and f16 keeps 2 048 of
them in the top octave.
Twelve real frames from the fixture set now align in 4.5 s — 4.4 s of
matching, 118 ms of bundle adjustment — where the first run took 51 s and
left the first two frames out.
The matcher computes each pair's similarity matrix once, across the
cores, with a dot product written to vectorise; both nearest-neighbour
directions read it. The frames that failed were portrait: fitted into the
landscape input they used 512 of 1024 px, and their thin overlap did not
survive at half resolution. The same weights are now exported at 768×1024
as well and the detector picks the shape by aspect. The example aligns
from embedded previews and draws the set on a cylinder; on the fixture the
sweep is 152° at a fitted 47.9 mm against the EXIF's 50, RMS 1.5 px, and
the overlaps show no ghosting.
A new crate holding the CPU half of a merge (FR-MRG-10): the grayscale
proxy with orientation, the XFeat decoder ported step for step from the
reference detectAndCompute, mutual-nearest-neighbour matching, a robust
pairwise homography with the focal length read off it, a hand-rolled
Levenberg–Marquardt bundle adjustment over every rotation and the focal,
the three output projections, and align(), which chains it all and names
the frames it could not place rather than guessing (FR-MRG-5).
Dependency-free without the xfeat feature — linalg.rs says why the dense
algebra is hand-rolled — and tested on synthetic sweeps whose answer is
known exactly. The noise test records the single-row degeneracy: one
pixel of noise is a tenth of a percent of focal, which is a uniform
stretch of the sweep, not a misalignment.
tools/onnx-probe-on-device.sh cross-builds dr-segment's onnx_probe
without the embedded segmentation model, pushes it with a model to the
attached device and times two runs. The 768×1024 XFeat export takes
~400 ms on the reference tablet's NEON cores against ~300 ms on the
desktop, with identical output ranges — inside NFR-MRG-1's 1 s per frame.
The blend half of S15.4 waits for a chunked blend to exist.
Camera-linear u16 samples on the first source's black-subtracted scale
with its white level, never rescaled to fill 16 bits, with its body,
matrices, illuminants and as-shot neutral carried — so the panorama is
developed afterwards as one photograph from the sensor's own numbers.
The only thing a warp cannot preserve is the colour filter array, and
the clause says so.
The fused chain, as operation.rs's tests fix it, is warp → as-shot white
balance → operations → base curve → camera matrix → store. LinearWorking
stores after the matrix, so the existing linear tap carries the body's
base curve, and a composite stitched from it and developed as an
unprofiled body would render that curve twice.
FR-MRG-2 therefore stitches camera-linear RGB — after the warp, before
white balance, curve and matrix — and the composite carries the first
source's body, matrices and as-shot neutral so its own develop applies
the profile once. The composer already makes this a uniform question:
white balance, matrix and the curve flag are reserved uniforms, so the
tap is a compose entry with no operations and a render entry that fills
them neutral. panorama.md §5.1 states the shape and asks for f32 buffers.
tools/export-xfeat.sh exports the convolutional network alone at 768×1024
grayscale, on the pattern of export-seg-model.sh: thirteen standard
operator types, no dynamic axes, the keypoint decoding left to Rust.
examples/onnx_probe loads it through the ort-over-tract backend the app
ships with nothing unsupported and runs it in ~300 ms on the desktop CPU.
The weights are Apache-2.0, read from the repository's LICENSE, with no
grant on the checkpoint — recorded in models/LICENCE.md before they land,
as FR-MRG-8 asks. The probe stays: the next model will need the same
check.
A hand-rolled 64×48 LinearRaw DNG — one IFD, 16-bit RGB, DNGVersion,
ColorMatrix1, AsShotNeutral — comes back through rawler 0.7 with cpp 3,
the samples in the order written and the matrix parsed into the camera
definition; CameraProfile::extract builds a profile from it. ImageMagick
reads the same bytes.
dr_decode::decode currently accepts the file as CFA and passes three
times the samples on, so the cpp == 3 branch is the decode work FR-MRG-3
needs, and the only decode work. panorama.md §8 records the result.
A merge writes a new source file beside its sources (D18) rather than a
multi-source Version, which answers the schema question §7 had been holding
open for panorama, HDR merge and focus stacking together. The panorama is
undeferred as FR-MRG-1 … 11; the other two stay in §7 with their data model
decided.
FR-MRG-10 and 11 fix where the work runs — every per-pixel stage on the GPU,
the composite never held as one texture — because the output exceeds
max_texture_dimension_2d before it exceeds memory. panorama.md carries the
stage table, the chunked output driver, the model licences and the porting
sources. S15 gates all of it.
Coverage falls from 83.0% to 77.2%: thirteen requirements entered with no
code, and outstanding.md §11 says so.
The 106 points the eye boxes were cut from, stored beside the reading as
16-bit fixed point over the frame: 424 bytes a face, a seventh of a pixel
on a 6000-pixel frame, where f16 at the same size would have been six.
Derived data like the embedding, kept for the same reason — it cost a
fetch and a model run, and the next per-face pass should run from the
catalog. Shards carry it; a peer's shard from before it is still read.
The register grew both clauses the same day this was built: FR-CULL-8a is
the per-face state the reading is, and FR-CULL-13 is the rule that a
signal is shown and filtered and never writes a judgement. The tags,
faces.md §17 and catalog.md now say which is which; FR-CULL-8a records
what of it is built, and that its third model is under the InsightFace
grant by the same decision as the pair.
The people filter was served from faces_image without touching a row;
reading the eye columns in the same subquery touched every one, and
ALTER TABLE had put those seven floats after the embedding and the crop
blob. One count took 24 seconds on the reference library, thirteen of
them system time. faces_eyes covers the subquery again: five
milliseconds.
FR-CULL-13, with §3.9.1's exclusion of blink detection re-read as the
exclusion of blink selection it always was: the stored fact and the chip
are built, a pass that picks the frame where everyone's eyes are open is
not. faces.md §17 has the models, the crop measurements, the four-state
rule and its floors, the native and proxy sheets read face by face, and
what remains to measure.
An "Eyes open" chip beside the people chips, offered only while someone
is chosen and dropped when the last person goes, so no term narrows the
grid with nothing on the bar to say so. It compiles the rule in
dr_face::eyes into the person's face subquery — Anna, eyes open, whoever
else is blinking beside her — and drops a frame only on a closed eye that
could be read: sunglasses, eyes too small or soft to read, and faces never
read all pass, so an old library shows everything under the chip until
the measuring pass has run. A test drives the same readings through the
SQL and through the rule and requires them to agree.
The People screen badges a face "Eyes closed", "Sunglasses" or "Eyes
unclear" so the reason a frame is or is not in the grid can be read off
the face; the sweep loads the three models when they are beside the pair
and reads eyes on the indexing and measuring passes from the native
render; the coverage line counts unread faces as work to measure so an
already-indexed library keeps its Index button. The term travels with the
place.
2d106det for the eye contours, OCEC for open or closed, SGC for
sunglasses — all three pinned to a batch of one by the same script as
the pair, and installed by every packager so the eyes-open filter works
out of the box. The two classifiers are MIT, code and weights; the
README records their provenance, SGC's undocumented training set, and
the hashes as fetched and as shipped.
SCRFD's eye point places a face, not an eye: on turned and smiling heads
the classifier's window had the eye in a corner, and two model-free ways
of re-centring it — the darkest blob, the most contrasty window — both
lost open eyes (19 → 15 and 19 → 9 of 25). Three landmark models were
then run over the same faces; Face Mesh V2 and InsightFace's 2d106det
tied at 22 of 25 and 2d106det ships, being the cheapest by far and under
the grant the detector and embedder already carry. The eye box is the
tight bounding box of its ten lid points, cut upright from the native
render, which is what the classifier was trained on.
The larger change is that the reading now carries, per eye, the source
pixels across the box and the sharpness of the patch — because the
commonest wrong answer on the reference library was a soft eye read as
closed, and a classifier shown a smear will always say something. An eye
under either floor, or narrower than six tenths of its partner (the far
eye of a turned head, whose contour collapses), is not asked; a face with
no readable eye is a fourth state, Unreadable, that no filter drops. On
twenty native renders the one real blink is caught, the laughing faces
are closed, the profiles are judged on the near eye, and the one thing
left beyond any floor is a face with a pot held over it.
Three nullable columns beside quality — P(open) for each eye and
P(sunglasses) — because the verdict is a rule with thresholds in it and a
rule belongs in code, not in rows that would have to be re-measured. NULL
is "never read": a face from before the models, or from a device without
them, and every reader treats it as unknown rather than as closed.
The measuring pass V14 built for the embedding's length is what fills
them, so the sweep's work list now also names faces with no eye reading
— but only on a device that has the models, or it would fetch every
original to do nothing to it. A peer's shard without the reading is still
adopted, unlike one without the quality: the pass finds this work by the
NULL rather than by the run marker, so adoption costs it nothing.
Two MIT classifiers from the same author as the reference pipeline's
whole-body detector: OCEC answers P(open) for one 40×24 eye, SGC
P(sunglasses) for a 48×48 head. Both load in tract once their batch
dimension is pinned by tools/fix-face-model-shapes.sh, like the embedder.
The crops come through the same fitted similarity the aligned face does,
so an eye window is a constant in template units rather than a second
warp, and a tilted head yields an upright eye. Measured on 60 proxies
from the reference library: the eye window plateaus at 22×11, the S
variant beats M and L (which overfit their own domain), and for
sunglasses the aligned face beats a head framing but the higher of the
two catches 11 of 12 pairs against 9 for either alone.
The reading keeps both eyes and the sunglasses number apart, because a
wink averages to the least informative value and a lens of dark glass
draws a confident answer from the eye classifier — over a woman in
sunglasses it read the right eye 0.97 open. Sunglasses take precedence,
and a face behind them is neither open nor a blink.
Its plugin section still asked for the contradiction to be resolved, its
render-path section still asked whether FR-DSP-2 was a requirement and
said NFR-RES-2 had no answer, and its closing section still called D12
open. Each now records what was decided and keeps the argument that was
weighed, so the document reads as the history it says it is rather than
as a plan the register has moved past.
Eight citations named §5.1, §5.2 and a §5 selector language that
requirements.md's §5 has not contained since it became a pointer at
architecture.md; two named §9 for the golden images and the benchmark
suite, which are §8; and the three pointers into architecture.md were
each one section off. All now name the section that holds the thing.
architecture.md §12's subsections are numbered 6.1–6.13, colliding with
its real §6. That numbering is what every ARCH §6.n citation in the tree
uses, so it stays, and a note at the head of §12 says so instead of
leaving the next reader to work it out.
FR-DEV-3f's open question about persisting the film stock was answered
in sidecar.rs; the clause now says so.
R5 says in its own note that zoom_resolution.rs establishes it as a
pixel equality; that file was tagged FR-DSP-5 alone. FR-DEV-19's three
sub-clauses carry eighty-three tags between them while the parent had
none; MaskLayer, which is the thing they edit, now carries it. And
NFR-R3 — a crash in decode does not take down the application, the
image is marked failed — is exactly what the decoder's panic guard and
the face sweep's unreadable mark do, tagged FR-RAW-4 and NFR-SEC-1 and
not the clause that asked for them.
NFR-COMPAT-1 and NFR-COMPAT-2 were instructions to write a requirement,
not requirements: "state the API level", "state the channels". Both
are now stated from what the build enforces and what exists.
The baseline is minSdk 28 / targetSdk 36 from the Android Dockerfile,
a Vulkan adapter at wgpu's default limits because compute needs storage
textures — device_from already called that the floor and is tagged for
it — with no optional feature required, since the f16 in FR-DEV-2 is a
texture format and not shader arithmetic. The reference device is the
HONOR ROD2-W09 the figures are taken on, and the second-vendor clause is
recorded as unmet rather than quietly dropped: there is no Mali or
PowerVR device, so an Android figure here is an Adreno figure.
The channels are all self-distribution — Arch package, local Flatpak,
sideloaded APK, NSIS installer — because D13's face weights rule out
every store, and the two consequences are written down: SAF stays
although a sideloaded build need not have it, and S11 becomes a
pre-publication step.
§7 still listed "AI subject masking — deferred per D11" while
MaskSource::Subject and MaskSource::Category, backed by dr-segment's
instance and semantic models, had been the primary way a local
adjustment is made for weeks. The code was tagged FR-DEV-3, which
names gradients and brushes and says nothing about a model.
FR-DEV-3i now states what exists: a subject or a category found by a
local model, stored as identity with the run's signature so that it
merges per field and reads as stale rather than wrong, then treated as
any other layer by the edge, stroke, composition and reveal clauses.
The one place it departs from FR-DEV-19 — coverage written run-length
coded beside the layer, so a stored subject renders without a model —
is recorded in the clause instead of left for the next audit to find.
The segmentation crate and the UI's selection module are tagged to it.
R5's note said FR-DSP-2 "was rewritten rather than implemented". It was
not: the clause still demanded viewport tiling, the matrix listed it
unbuilt, and frame-budget.md's rewrite had been proposed and never
applied. Decided 2026-09-19 to keep it as written until S6 runs on a
mid-range Android device, because the measurement that argues against
tiling was taken on a discrete desktop GPU and the clause exists for the
device whose memory the image exceeds. Both notes now say that.
The licensing half of D13 had been open since 2026-08-09, while the
InsightFace detectors and embedder shipped in the tree and indexed real
libraries. models/face/README.md already stated the position the
project was actually taking; the register did not.
Now it does: this is non-commercial software, self-installed, and it
uses the weights under their research grant as such. The risks are
written where the decision is — the grant binds every user, it is not
GPL-compatible, it rules out every public channel, and publishing is
what reopens the decision. S14's licence search is what would close it.
NFR-R8 carried the words "decide explicitly" for six weeks, asking
whether v1 has a full CPU render path or whether "CPU fallback" means
staging only. The viewer had already answered it: with no adapter it
opens the library on embedded previews and cached proxies, keeps every
catalog edit available, and withholds develop and export. That is the
degraded mode, it is now the requirement, and NFR-RES-2 no longer
promises a fallback render path that was never going to be built.
Three statements disagreed. §1.2's platform table said "phone
supported"; §3.5 said phones were out of scope and cited §1.3, which
does not mention them; D15 said "no phone" and gave the reason. D15 is
the decision, so the other two now point at it and say the same thing:
the build runs on a phone, and nothing is designed for one.
The register said two things about plugins. §7 had listed "Plugin API"
as deferred since the first draft, in a bare row; §3.10 then specified
it in 23 clauses that counted against coverage. Twenty-one of them had
no implementation of any kind, and could not have: no crate loads
anything at runtime. The coverage figure was measuring the contradiction.
Decided 2026-09-19: §7 is right. §3.10 stays as the design of record,
each of its clauses is marked "(post-v1)" on its defining line, and
NFR-SEC-6 — which exists only for plugins — goes with them, as does D16.
The traceability tool learns the marker. A deferred requirement is still
defined, so a tag naming it is not an orphan, but it leaves the
denominator and is listed in its own table rather than under "not yet
tagged". The marker must sit on the definition line; a mention of
"post-v1" in prose changes nothing, and where an ID is defined twice the
deferral on either line wins. Both are tested. Coverage moves from 72.2%
of 194 to 80.6% of 170 without a line of application code changing,
which is the honest figure: it now measures what v1 owes.
D12 had been OPEN since the 2026-08-08 calibration, and D3 said it
depended on D12. In the meantime the milestone D3 named was delivered
and closed on 2026-08-30 and the application reached 0.12.2 with every
cluster the calibration selected at least begun. The decision the
register was waiting for had been made by building, so the register
now says so: full scope stands, v1 has no date, and "post-v1" in §7 is
the one way a clause leaves the count.
The status line also stops calling this a draft from August; it has
carried eleven dated amendments since.
The sweep fetches the whole original before it can learn anything
about it, and the one file in the reference library the decoder
refuses on sight is a 521 MB stitched panorama — so every pass on the
tablet spent half a gigabyte of Wi-Fi to find that out again. The
catalog already knows the byte count, and that is enough to decide
before the fetch: originals over 256 MB are marked examined with
nothing found and a zero edge, counted as failed, and named in the
log. Below the line is every camera RAW the library holds; above it,
four files, all panoramas.
A budget and not a verdict on panoramas. The right treatment for one
is a tiled pass — read it in strips, detect in each, stitch the boxes
back — and the zero edge is what that pass would select on. Until it
exists, this is what keeps a background sweep on a phone from paying
for the decision the decoder cannot make.
Rating and flagging were reachable from the grid alone, so a photograph
opened in develop could not be judged without leaving it; FR-UI-5 said
"rating" without qualifying the view and was built as though it had.
And FR-UI-1's expanded row has said "filmstrip" since it was written
while the roll stayed on demand in both classes. Both are amended to
say what they meant: judgement follows the photograph, without
auto-advance outside the culling mode, and the roll is open by default
where there is room for it.
The larger change is a rule. Per-face signals — eye state from a
classifier, head pose from the five landmarks the detector already
yields — are worth having for culling, and §3.9.1 excluded detecting a
blink outright. The exclusion was always of judgement, not of knowing:
a blink is a fact about a frame of the same kind as a clipped
highlight. FR-CULL-8a specifies the two signals; FR-CULL-13 says what
any signal may do (be shown, filtered, sorted, propose a burst
representative) and what none may (write a rating or flag without a
user action between). R7 states the same thing as a user need.
Licensing was read before either was written. OCEC's eye-state weights
are MIT with a clean data chain; every open gaze model is trained on
Gaze360 or its peers, whose licences restrict derived models by name,
so gaze is deferred in §7 and head pose stands in for it. D13 records
both so they are not re-searched.
Replacing a closed-eyed face from a neighbouring frame was raised and
is written down as D17 rather than built: it is the multi-source schema
question §7 already defers for panorama and HDR, with its non-goals —
never automatic, provenance declared — fixed now.
Traceability regenerated: three new IDs, none yet tagged.
Choosing "Thorough" made the library look empty. The detector setting
writes under its own faces.model_id, and every reader of "the faces"
keyed on that exact id: the clustering pass, the coverage figure, the
sweep's work list, the shard export and import, and the sync merge's
face matching. On the reference library that restarted coverage at
1,834 of 19,140, drew a People rail of 36 faces for a person with 520,
queued a ~400 GB re-fetch on each device, and stranded the desktop's
3,583 confirmations under the old id: the tablet held the same faces
under the new one and the merge refused to match them. Same photograph,
same box, same embedder, two ids — that is one face, not two libraries.
The embedder half of the id is now the key. embedder_of and embedder_sql
give it to every query; writes keep the full id, so which detector drew
a box stays on record. record_detections is unchanged and is where the
generations meet: an image holds one pipeline's faces at a time, and a
re-detection carries confirmations across by box overlap. The merge's
match_faces applies the same rule within an embedder. The calibration
is keyed on the embedder too, since the similarity space did not change.
Shards travel every generation, each under its own id, and a peer adopts
whichever it is sent — including a stronger detector's pass over an
image it indexed itself with a weaker one, which is the re-detection its
own sweep would otherwise queue, already done. Never downwards: a tablet
on Fast keeps the desktop's Thorough faces. The sweep gains the same
tail — images a weaker detector indexed, after the ones nothing has —
driven by FaceDetector::supersedes, so choosing a stronger detector still
improves the library over time without first making it disappear.
A decode failure in the face sweep was counted, logged at debug where
nobody saw it, and left unmarked — so the next pass fetched the same
file and failed the same way. For the 521 MB panorama behind rawler's
panic that was half a gigabyte per sweep, on a tablet. It is now marked
examined with nothing found and a zero edge, which is what a later "try
again with a better decoder" pass would select on, and the warning
names the file. The failure count is unchanged: it did fail.
rawler panics on some input rather than returning Err — a DNG whose IFD
claims a >50000 px image, which the reference library has: a 521 MB
stitched panorama, IMG_4181-Pano.dng. On a worker thread a panic is the
end of the thread, so the face sweep that met it stopped thirteen
seconds in, three sweeps running on the tablet and three on the
desktop, with "17301 image(s) to index" as the last word. FR-RAW-4
says a malformed file must not abort a batch, and that is this crate's
promise whatever the library beneath it does: every entry point that
calls into rawler now runs under catch_unwind, and a file that panics
the decoder is one failed file with the panic's message in the error.
Verified on the panorama itself: metadata reads, decode returns the
error, the thread survives. The crash hook still records the panic,
which is right — it is a defect in a dependency and the record is how
it gets reported.
Walking the photo roll was one download per frame: every step showed
"Downloading…" over an empty canvas while tens of megabytes came down,
and moving between a pair of near-identical frames paid that a dozen
times. Now, once the opened photograph has landed, the ones around it
are fetched into the originals cache while it is being looked at, so
the next step is a disk read.
A single worker serves the latest wish only, closest first and working
outwards — next, previous, next-but-one, previous-but-one… — one file
at a time. Each open replaces the wish, so a fast walk never leaves a
trail of stale downloads competing with the one being waited on. A
process-wide in-flight registry makes a click on a photograph that is
still being fetched ahead wait for that transfer and read it from disk,
rather than start a second download of the same file.
How far each side is a setting under STORAGE — Off, 2, 5, 10 or 20,
defaulting to 5 — and it is moot while "keep originals after opening"
is off, since a fetch the cache would discard on arrival is transfer
for nothing. Nothing is fetched ahead while offline. The transfers show
in the activity list while they run and are removed when they end.
The launcher and the task bar have shown a generic tile for a working
window since the call was written. set_xdg_app_id sat at the top of
run(), on the reasoning that the app id is read when the surface is
created — true, and beside the point: the call goes through Slint's
global context, and there is no global context until something installs
a platform. That is BackendSelector inside shared_gpu, or AppWindow::new
falling back to the default, and both happen further down. Called before
either, it returned NoPlatform and did nothing at all.
It moves to just after the window is constructed, which is not the same
as shown — run() is far below — so there is a platform to talk to and
the surface does not exist yet.
The failure was logged at debug, which is why a year of grey squares went
unremarked: the whole symptom is invisible from inside the application.
It is a warning now, naming the consequence.
SQLite's busy timeout defaults to zero, and nothing ever set one: the
loser of a write race got SQLITE_BUSY at the moment it asked. WAL does
not cover this — it makes one writer and many readers free, and this
application constantly has two writers, the face sweep committing a
batch while the derived sync imports shards or reclustering reads.
The cost was not a retry but lost work. A sweep that had already paid
for the detection and the embedding — seconds per image, the expensive
part — discarded the result on "storing faces for 214: database is
locked" and moved on to the next image. Both the desktop and the tablet
logged runs of those on consecutive images, which is a face sweep
quietly failing to store the faces it had just computed.
Ten seconds, on every connection, set in configure() so that nothing
can open the catalog without it — the figure the job runner's own tests
have used for this reason since they were written. It is far longer
than any transaction here, so it bounds pathology rather than making
anyone wait.
fix/corrupt-remote-catalog-deadlock. The catalog on the server had been
malformed since 7 September and every device declined to overwrite it,
so collections and people stopped syncing everywhere at once; the
damage came from two devices assembling chunks in one upload directory.
Each chunked upload now has its own directory, the server keeps three
generations of the catalog behind the current one, every push is
verified before and after, a damaged copy that arrived whole is merged
from the generation before it rather than pinned in place, and the
catalog is backed up daily as NFR-R2 always asked.
NFR-R2 asks for the catalog to be backed up on a schedule and before
schema migrations. Only the second half existed: every backup on disk
was a pre-migration copy, and a library that never migrated was never
backed up at all.
A backup is now also taken at the end of a library sweep when the newest
one is more than a day old — the moment the catalog is quiet and a day's
collection and people edits have just been folded in — on its own
thread and its own connection, so the copy of a 130 MB file is not spent
on the UI. Whether one is due is read from the backup directory, not
the catalog, so the ordinary case costs nothing. An empty catalog is
skipped: there is nothing in it a rescan would not rebuild. Pruning to
KEEP_BACKUPS applies as before.
Two checks around the upload, both cheap next to what they prevent.
Before: the snapshot is quick_checked before it leaves. It is the copy
every other device merges from, and a damaged one costs each of them a
download, a failed merge and a refusal to push.
After: the staged upload's size on the server is compared to the bytes
sent before it is rotated into place. A chunked upload is assembled
server-side, and an assembly that goes wrong is a file of plausible
size no device can open — caught here, on the device that caused it,
for one listing; otherwise on every other device, after the fact. A
mismatch, or a size the server will not confirm, discards the upload
and leaves the current copy and its generations untouched.
The server held one copy of the catalog, overwritten in place on every
push. When that copy was damaged there were two answers, both bad:
refuse to touch it for ever, which is what every device did for a week,
or overwrite it with ours, which loses whatever another device had
added since — the escape hatch of the previous commit.
A push now uploads to catalog.upload.sqlite, rotates catalog.sqlite to
.1 (and .1 to .2, .2 to .3, dropping the oldest), and moves the upload
into place. Rotation is server-side renames, oldest first so that every
destination is empty when it is written to — move_to refuses to
overwrite, by design — and a failure at any step leaves a gap in the
generations and never a missing current copy. The only transfer is the
upload itself.
A damaged current copy that arrived whole now merges from the newest
readable generation before ours goes over it, which loses nothing, and
is kept as .1 by the ordinary rotation rather than by a separate 40 MB
upload. NFR-R2 asks for the catalog to be backed up; this is the half
of it that lives with the copy other devices read.
The upload directory was named from the destination path alone, on the
reasoning that two files could then not collide. Two devices uploading
the same file could, and did: both wrote 00001…00009 into one directory,
and whichever MOVEd first assembled a mix of the two — a catalog of
exactly the right size whose pages came from two different databases.
SQLite called it malformed, every client declined to overwrite it, and
collections stopped syncing on all of them for a week. A transfer that
died on a phone's link left its chunks there for the next device to
assemble in, by the same mechanism.
The name now carries a nonce as well, so no two uploads share a
directory, and a failed transfer deletes its own directory on the way
out rather than leaving 5 MB chunks for the server to sweep eventually.
"The chosen detector is not installed" was the whole of what a user
saw, on a machine where the files were three directories away from
where the lookup went. The search order was in a doc comment and
nowhere a user could read it. Now a missing pair logs the detector
file it wanted and every directory it tried, which is what the first
Windows install needed and what the next misplaced download will.
The catalog sync refuses to upload when it cannot read the server's copy,
because the upload is a read-modify-write and writing blind would discard
another device's collections. That is the right rule for a timeout, a
dropped connection or a newer schema — the remote is fine, only our view
of it failed.
A file SQLite calls malformed is not that. No device will ever read it
again, so refusing to write over it preserves nothing — and every client
declines in turn, pinning the damaged file in place for good. Collections
and people then stop crossing between devices on all of them at once,
each logging "catalog not pushed" on every pass. This library did exactly
that from 2026-09-07, on the desktop and on a freshly installed phone
alike, while 32 collections sat undelivered.
Now a copy that arrived whole and still will not open is set aside under
a dated name and replaced by ours. Whole is checked against the size the
server advertises: a truncated download will not open either, and on a
phone that is the far likelier story, so anything short — or any size the
listing cannot confirm — is treated as the transport failure it is and
the server's copy is left alone. A placeholder's size is not trusted for
the comparison, since it means nothing.
The report says when this happened, and the log line calls it "pushed
over a damaged copy" rather than folding it into an ordinary push: it is
the one push that discarded something.
dtourolle