505 Commits

Include renames

Author	SHA1	Message	Date
dtourolle	81b1ae8c42	Measure the haze from the picture, and divide it back out ... Four files named dehaze as a member of the compositional detail family — `detail.rs` twice, `dr-gpu`'s detail module, `ops/README.md` and `capture_sharpen.rs` — and no such node existed. Every one of them was describing the family by listing clarity, texture and a control the photographer could not reach. Haze is the one degradation the controls already in the chain cannot remove, and the reason is spatial rather than tonal. Scattering composites an airlight over the scene in proportion to distance, so the lift is per-pixel: a black point that clears the mountains crushes the foreground, and a contrast curve that clears the mountains does the same. So the node has to estimate the transmission at every pixel, which is the dark-channel prior — the local minimum over the channels and over a patch is the airlight that has been added there — and then invert the scattering model with it. The airlight is taken as neutral and as unit, which removes the one part of the published method this stage cannot perform. Estimating it properly is a whole-frame reduction, and the detail chain has none: it hands each pass the pass before it. It is also unnecessary, because white balance is the first node in the chain and has already driven the illuminant to grey, so only the magnitude is unknown — and an unknown magnitude on the veil is a scale factor on the amount slider, which the photographer is setting by eye regardless. The patch is a fraction of the frame's shorter edge, through `RenderScale::frame_fraction`, and never a count of pixels. It has to be wide enough to contain something dark and narrow enough that what it measures is still local, and both of those are statements about how much of the composition it covers — so it must cover the same proportion of the picture on a proxy as in the export, or the file is sharpened for a patch three times narrower than the one that was tuned on screen. Affording it needs an identity a Gaussian does not have. Erosions compose by adding their structuring elements, so the minimum over a run of d followed by the minimum over k points spaced d apart is the exact minimum over the whole kd window. At the square root that is 16 taps rather than 61 at 4K, and it is the same filter rather than an approximation of one — which is the difference from the strided kernel `local_contrast` refuses, where sampling an image that is not band-limited aliases into the base and comes back as mottling. It runs first among the compositional detail nodes, at order 125: after noise reduction, because dividing by a transmission below one amplifies the noise in the veiled distance by exactly the factor it recovers the contrast by, and before clarity and texture, coarse before fine, so that their base is computed on the picture the veil has left rather than on a modelling about to be divided out. What it cannot honour is the placement dehaze most wants. It shifts colour — it subtracts a grey term and rescales, so saturation changes wherever the veil is thick — and the colour work would ideally be correcting the picture that leaves here. The detail stage runs as a group after every point operation, because a neighbourhood pass is a separate dispatch over a texture the fused pass has finished writing, so an order placing this node ahead of `vibrance` would be a lie the chain cannot tell. Interleaving would mean splitting the fused pass in half around it, at the cost of a second full-frame dispatch and intermediate for every edit in the catalogue whether it dehazes or not. The declaration records that rather than leaving it to be rediscovered. FR-DEV-18 is added to the requirements register alongside it. The tag had nowhere to point, and an orphan tag fails the traceability gate rather than quietly counting for nothing.	2026-09-06 19:01:47 +02:00
dtourolle	7c3e1d2c54	Let the shadows and the highlights carry a colour the picture never had ... The colour mixer is the only chromatic control in the chain, and it can only turn a hue that is already in the frame. Ask it for cool shadows against warm highlights and it has nothing to take hold of: the shadows of a correctly balanced photograph are near enough neutral that there is no band there to turn, and a monochrome conversion hands it a picture with no hue in it at all. Split toning is the oldest look in the book and every developer worth comparing against ships it; there was no way to reach it from here. So colour_grading, declared like any other node — a hue and a strength for the shadows, the midtones and the highlights, and a global cast over the frame. It targets a tonal range rather than a hue, which is the whole difference between the two controls: it puts colour where none was rather than turning what it finds. It sits at 105, after the mixer has had the last word on the colours that are in the picture and before the detail stage. The mechanism is one helper. Three cosines 120 degrees apart are the hue wheel written directly as an RGB direction, and their sum is zero at every angle, so exp2 turns them into three gains whose product is exactly one — a cast tilts the balance without moving the level. A grade that doubled as an exposure change is the failure that has the photographer chasing brightness with a colour slider, and it is corrected with a control that cannot reach it. The three tonal weights partition the scale rather than overlapping, the midtones being whatever the two ends leave, so setting all three to one hue is exactly the global cast and a split tone does not colour its own midtones as a side effect of its halves meeting. Full strength is half a stop on the leading channel, the ceiling white balance already holds itself to. Neutral is declared rather than inferred, which is what `active:` is for. A hue with no strength behind it is a direction with no distance, so under the default rule nudging one would have put the node into every fused shader for a change nobody can see. Summing the strengths is zero exactly when all four are, and they cannot go negative to cancel each other. The opposite reading — neutral as "nothing has been touched" — fails the other way round: red is hue zero, so a grade toward red never moves a hue off its default and would never have been applied at all. It asks for a colour wheel, the widget the descriptor vocabulary has been carrying with no operation behind it. Nothing draws one yet, and that is fine by construction: the panel takes the first widget it implements and falls through to sliders otherwise, so this arrives as eight ordinary controls that work. Each parameter is named for its own range for exactly that reason — in a flat list, four sliders called "Hue" are four controls nobody can tell apart. FR-DEV-12 is written into requirements.md beside it. A TRACES tag naming a requirement that is not defined there is an orphan, and the traceability gate fails on those rather than quietly counting them. The label catalogue gets one line for the operation's display name; the eight parameters derive correctly and are left to.	2026-09-06 18:51:40 +02:00
dtourolleandClaude Opus 5	59917c5183	Call the tool Compose, since that is what its panel says ... The rail entry read "Crop" while the panel it opens is headed COMPOSE and the button leaving it said "Done Cropping". One mode, three names, and the odd one out was named after a single control rather than after the decision — which is what made cropping look like a category of its own in the first place. Straightening, the quarter turns and the flips are already in that panel, and perspective will be. `ViewMode.crop` keeps its name: it identifies a canvas interaction, which is exactly what it still is. Found by looking at the running application rather than by reading, which is also how the two halves of this were noticed to disagree at all. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 18:29:08 +02:00
dtourolleandClaude Opus 5	e235e99cce	Move the film to Effect in the descriptor that is actually read ... An earlier commit claimed to move `film_sim` from `[tone, colour]` to `[effect]` and did not. It edited `ops/film_sim.yaml`, where `attributes:` is read, validated against the vocabulary, and then dropped: a `rust:` node publishes its own descriptor, and the type still said tone and colour. The stock went on appearing in the Light group beside exposure and again in Colour beside white balance, exactly as before, and every test passed. Nothing caught it because nothing could. The declaration parsed, the parity tests compare ids rather than attributes, and an operation filed under the wrong groups renders perfectly. It surfaced only on screen, as a missing Effects tab — which is indistinguishable from a category that genuinely has nothing in it, and is precisely how `Optics` looked for as long as it was empty. So three changes rather than one: `FilmSim`'s descriptor declares `Attribute::Effect`, which is the move the earlier commit described. `attributes:` joins the keys a `rust:` node may not carry, beside `params`, `uniforms`, `wgsl`, `helpers`, `define` and `label`. The rule was already written — "its descriptor comes from the type" — and attributes were the one field that slipped past it. A key that is silently ignored is worse than one that is rejected, because it reads as though it worked; the eight hand-written declarations lose a line that never did anything. And a test asserts that every attribute the chain carries reaches the tab strip. That is the property that was actually broken, and its failure mode is invisible from every direction: the controls exist, they are in the shader, and there is no way to filter to them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 18:28:59 +02:00
dtourolleandClaude Opus 5	6a97fdf6f9	Put the adjustment groups in the rail where a finger is driving ... Reported from the tablet: the tool rail is very useful there, and the same interface under a mouse and keyboard is not. That is `ui-navigation.md` D-N2's central assumption failing in use, and the interesting part is which half of it failed. D-N2 was right that platform is the wrong axis and width is the wrong axis: a tablet in landscape wants what a desktop wants, and a desktop window dragged narrow wants what a small screen wants. `apply_layout_class` still decides the layout class from the window and nothing here changes that. What D-N2 got wrong is the sentence "touch changes hit regions, not layout" — it identified input as the real difference between the targets and then assumed that difference could never reach the layout. Two controls answer one question — which group of adjustments am I looking at — and neither is better in general. A horizontal strip above the column is one gesture to a target the eye has already found, and it pans when the operation set is rich, so a group can sit off the end with nothing saying so: a pointer user tolerates that, a finger user never discovers it. The same list down the rail is every entry visible at once, each finger-sized, on the edge of the screen the hand is already holding, and it costs no width because the rail is already there. So `ToolRail` grows a second section, and `GroupStrip` stands down when it does. The two are never both on screen, which is why they can share `adjust-tab-picked`: Rust is not told which was pressed and has no reason to want to. Mode and group stay independent axes as N1 requires — one entry lit in each section, and choosing a group while a tool is held still filters without putting the tool down. They stay drawn differently, which N1 also required. The tools fill with `active-dim` and invert their ink; the groups take a bar down the leading edge — the strip's underline turned ninety degrees — so a lit entry says which kind of state it is without the reader having to remember which section it was in. The rule between the sections is the second signal. The rail scrolls now. Its own note argued against a Flickable because "this list is four entries written in this file"; with the groups in it the list comes from the operation set, which is exactly the "something the user's data decides" that note excluded this control from. The axis is input, and it is a preference because the automatic answer is a guess that cannot be made reliable. Neither platform can be asked what the user is holding: an Android tablet in a keyboard case is being driven like a desktop, and a touchscreen laptop is whichever its owner says. `dr_plat::is_touch_first` reports the usual case per platform, and `GroupNavigation` lets it be overridden. Settings names what Automatic resolves to on this device rather than leaving it to be found by pressing. D-N6 records the reversal beside the decision it reverses, including the half that still stands and the question it opens: whether Local is a mode at all, or a scope that would collapse the two sections into one list. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 18:14:03 +02:00
dtourolleandClaude Opus 5	1ad35e2b87	Read the library's sidecars, so a cull done elsewhere arrives ... Judgements only ever travelled outward. A rating went to the catalog and to the photograph's sidecar, the sidecar reached the server, and there it stopped: the scan indexes files, `derived_sync` exchanges thumbnails, face shards and collections, `dr_catalog::merge` reconciles everything in a catalog except `versions.rating` and `versions.flag`, and the one sidecar reader that existed ran when a single photograph was opened in develop and handed its answer to the develop graph. `JobKind::ReadSidecar` was declared for exactly this when the job queue was written and was never enqueued or handled anywhere. The grid draws `versions.rating`. So a day of culling on the tablet could not reach the laptop by any path the application had, and the laptop's catalog says so plainly: 23,568 images, one of them judged. `pull_sidecars` closes it, off the back of work the scan already does. `dr_sync::scan` reports the `.drsc` files it meets in listings it was making anyway — no extra request, and a directory whose ETag is unchanged is still pruned before it is listed at all. A new `sidecars` table records the ETag of each one this device has taken in, so the fetch is one GET per sidecar that genuinely changed rather than one per photograph. A library nobody has edited costs nothing. The judgement is taken rather than maximised. The sidecar is the authoritative store and the fuse has already settled any contest between devices on `revision`, so lowering a rating from four to one on the tablet lowers it here — taking the larger would have refused every demotion the photographer ever made, which is most of what a second pass over a shoot is. A zero is the exception: it means *never judged*, not "judged zero", so a sidecar carrying none cannot erase a star this device holds. That is `merge_judgement`'s asymmetry and it carries the same known cost — clearing a rating does not propagate. A sidecar names a stem, so both halves of a RAW-and-JPEG pair are judged: they are one photograph (FR-CAT-11) sharing one document, and judging only one of them would leave the grid disagreeing with itself over which it drew. The `LIKE` that finds them is a filter, not the decision — `sidecar_path` is applied to every candidate, because a folder is entitled to contain a `%` and a rating landing on the wrong frame would be silent and permanent. Failing to read one is not a failure to scan: the ETag goes unrecorded, the ratings already here stay where they are, and the next scan tries again. The count is reported to the status line as well as the log, because a grid that silently gains three hundred stars is indistinguishable from one that has gone wrong — and because while this number was structurally zero there was nothing to tell the photographer their cull had not arrived. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 16:14:03 +02:00
dtourolleandClaude Opus 5	601c984894	Fold a photograph's rival default versions back into one ... Picking the newer of two default versions stopped the wrong edit being shown, but it did not close the split: the losing version stayed in the file, and a device holding disjoint work — a crop made here, an exposure change made there — still contributed only one of the two. Worse, the next write made it larger. `amend` looks its version up by uuid, neither of the two was ours, so the miss minted a *third* `default = 1` block and the file grew one rival per device per photograph. `Sidecar::fuse_default_versions` folds them down. The version with the highest `(revision, modified)` is the accumulator and every other default is merged into it as the remote, which is what makes the fold order-independent — `Version::merge` raises its own revision to `max + 1` as it goes, so merging a chain in ascending order stops being ascending after the first step and a third device would be dropped. Contested values resolve to the winner, disjoint keys survive from both sides because the merge is key-wise, and ratings come across under `merge_judgement`, so a device that never judged the frame cannot erase one that did. The result is a function of the file's bytes alone, so two devices that fuse independently reach the same document and converge instead of overwriting each other. Called wherever a sidecar is parsed: - `amend`, with the write's own uuid, so the fold lands on the identity this device is about to use and the lookup below it hits instead of missing. - `spawn_sidecar_fetch`, so opening a photograph shows everything done to it rather than whichever half won. - `drain_one`, because `merge_into` reconciles by uuid and would otherwise publish the split rather than resolve it. - `presets::load_local` and `save_local` — a local sidecar's folder may be synced by something else entirely, and gets the same split. A file with one default under the expected uuid comes back byte-identical, so this costs nothing on the ordinary write and no sidecar is uploaded merely for having been read. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 16:13:53 +02:00
dtourolleandClaude Opus 5	c4ddcbe0f7	Look the lens up and say plainly whether one was found ... `dr-lens` has held a complete Lensfun lookup — distortion, TCA and vignetting coefficients from a lens name, a focal length and an aperture — with no dependents anywhere in the workspace. The three corrections it feeds now exist in the graph, so this connects the two and finishes the chain. The coefficient structs stay duplicated. `dr-pipeline` is organised around having no dependencies so its codegen is testable without a device or a database (ARCH §6.5a), and `dr-lens` carries an XML parser and 5.5 MB of profile data. Neither crate can convert to the other, so the conversion goes above both, in `develop.rs`, which is the only place that sees them together. Both traits grow the same defaulted door. The optical corrections do not sit on the same side of the fetch — distortion and CA rewrite coordinates and are `Warp`s, vignetting applies a gain to the pixel already there and is an ordinary node — and fanning a profile out by which trait each happens to implement would make the caller reason about that distinction. Each correction takes its own share of the whole profile instead, and `set_lens_profile` walks both lists identically. The lookup happens in `set_source_metadata` rather than in its caller, because that is the one place a session is told which file it came from. Doing it there makes it unforgettable, in the shape `FilmRebake` already uses for the other derived thing — and, more to the point, makes *clearing* unforgettable: a session that opened a second photograph while still holding the first one's profile would correct it for the wrong optics, invisibly, in a way that looks exactly like the lens. It needs the whole shot and not just a name. Distortion is interpolated across a zoom's focal range and vignetting depends strongly on aperture — a fast prime can be two stops down in the corners wide open and clean by f/8 — so a lookup missing either returns coefficients measured for a shot nobody took. Missing any of the three refuses rather than guesses. A profile is derived, not persisted: it comes from the file's EXIF and a database, so it is not a parameter, not in the sidecar and not undoable. What is an edit is the manual trim beside it, which each correction composes with the measurement — so a photographer can lean on it, override it, or work without one. `InfoPanel` gains a lens line, and it distinguishes three cases rather than two. `dr-lens` states the rule it exists for: an automatic correction that silently did nothing is worse than one the user can see is unavailable. A session with no header draws nothing, a header naming no lens reads "Lens not recorded", and a lens the database has never heard of reads "· no profile". Collapsing the last two would send somebody hunting for a profile that was never missing — which, for third-party and adapted glass, is the ordinary case. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 15:11:32 +02:00
dtourolleandClaude Opus 5	a1165ef182	Put the coordinate-domain lens corrections into the graph ... `lens.rs` has held a `Warp` trait, a composer and two implementations — distortion and lateral chromatic aberration — since they were written, and `compose_warps` was called by nothing outside its own tests. The corrections existed, were correct, and never touched a photograph. `EditGraph` now holds them, and `compose_full` emits them between the framing prologue and the fetch. Distortion first, then CA: each warp receives the position the previous one produced, and lateral CA is a magnification about the optical axis of the *undistorted* frame, so measured on a barrel-distorted one it would be fitted to a radius no profile describes. They reach the panel the way framing already does — through `capabilities`. That was the one open question and existing practice answered it: framing is also not an `Operation`, also has parameters a photographer sets, and also arrives through that list. Because `Preset::capture` walks the same list, the sidecar, the clipboard and the undo stack carry a warp's parameters with nothing registered anywhere, and no file under `ui/` names one (FR-DEV-3a). `state()` destructures `EditGraph` field by field precisely so that a new field cannot be forgotten, and it was not. Chromatic aberration is the only thing that samples per channel, and `splits_channels` is what keeps everything else from paying for it. Red and blue are fetched from positions green is not — green is the reference and never moves, so a wrong correction still leaves one channel sharp rather than softening all three. With no CA in the chain the single-fetch path is emitted instead. The interpolating sampler is now chosen by framing *or* an active warp. Asking framing alone would have nearest-neighboured a distortion correction on an unstraightened frame, and that aliasing reads as a bad profile rather than as a missing filter. The warps go in the geometry invalidation key rather than the colour one: they decide which source pixel a colour is read from, so a tile cached across a distortion change would keep drawing the previous correction. The pipeline cache needs nothing new — `hash_source` already covers the generated body, and uniform values never enter it, so arming a warp recompiles and dragging it does not. Both are asserted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 15:03:16 +02:00
dtourolleandClaude Opus 5	e17b909d41	Connect the lens vignetting correction to the pipeline ... `ops/vignetting.rs` has carried a complete descriptor, polynomial, helper and test suite without an entry in `ops/`, so it was never in `chain()`. It reached no photograph and no panel, and `Attribute::Optics` was an empty category in consequence — filtered out of the tab strip for having no rows, by a chain that had never been given its only member. Declaring it needs the one thing the operation was written against and which did not exist. `wgsl_body` reads `radius`, and the module claimed "the composer publishes `radius` in the shader prologue for exactly this reason". It did not. `sample_source` now does, in both sampling branches, beside the `source_px` it already published for the same class of caller. It is corner-normalised there, which is the part that is easy to leave out. `p` spans ±0.5·aspect, so its length at the corner is 0.5·length(aspect) — about 0.901 on a 3:2 frame, not 1. Lensfun's polynomials are fitted against a corner radius of 1, so passing `length(p)` straight in evaluates every one of them short of where it was measured, by a factor that changes with the aspect ratio. It would have read as a correction that is simply too weak, which is indistinguishable from a bad profile. Both `lens.rs` and `framing.rs` asserted the normalisation `p` does not have; corrected. `order: 5` puts the correction ahead of the tonal stages, and the ordering is load-bearing rather than tidy. Recovering a corner means dividing by an attenuation below one — about two stops for a fast prime wide open — so run after the highlights have been rolled off and clipped, the lift has nowhere to go and the corners posterise instead of brightening. `layer_chain` now drops `Optics` as well as the neighbourhood operations. A local vignetting slider would have worked, which is what makes it worth excluding: `radius` measures from the centre of the whole photograph and a mask cannot move the optical axis, so it would lay a frame-centred radial ramp across the picture and multiply it by the mask. The existing exclusion covers operations that move and do nothing; this one covers an operation that moves and does something its name does not promise. The rule both share is now written down. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 14:34:24 +02:00
dtourolleandClaude Opus 5	8e7b1350bf	Name the frame's category for the decision, not the maths ... `Attribute::Geometry` becomes `Attribute::Compose`, and `film_sim` moves from `[tone, colour]` to `[effect]`. Two categories were doing the wrong job. "Geometry" describes what crop, straighten and the quarter turns do to coordinates — but it describes lens distortion correction exactly as well, and that is not a compositional choice at all. Naming the attribute for the photographer's decision is what separates it from `Optics`: one is what the lens did, the other is what they chose. The maths the two have in common is not the thing worth filing them under. A film stock declared both `tone` and `colour`, so "Kodachrome" appeared in the Light group beside exposure and again in Colour beside white balance — two places, neither of which is where anyone looks for it. It is neither: `Effect` is defined in this same file as "applied rather than corrected — a look, not a fix", which is what a stock is. That it moves tone and colour is true of every look, and is not what the attribute is for. `from_name` still accepts "geometry" on the way in. That string is persisted in `develop.copy_attributes`, and an entry it fails to parse is not an error — `presets::scope_for` logs it and drops it — so without the alias an existing settings file would have quietly narrowed what a paste carries. `name` writes the current spelling, so the file migrates itself the first time it is saved. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-09-05 14:34:03 +02:00
dtourolleandClaude Opus 5	f6c9343bcc	Ask the pixels where the edge is, not just what belongs ... The colour gate decided *what* was in a category and had no way to decide *where* its edge fell. A colour test has no notion of an edge. So a refined sky lost its flag and kept the model's twenty-pixel-blocky outline, and no setting of the control could move that outline onto the horizon. This adds the second half: a **marker-based watershed**. The mask is eroded to give two markers, and the flood runs in the ribbon left between them, meeting along the most expensive line it can find. The cost is a sum of terms exactly as docs/segmentation.md §2 specifies — the photograph's own edges, and the colour model's disagreement. ## Why this is not the watershed §15 threw away That path failed because the merge *ladder* collapsed: 45,808 basins reduced to one region plus specks. There is no ladder here. Markers prevent over-segmentation by seeding rather than by merging afterwards, so the one component that broke is the one component this does not have. The markers are also better than the textbook's. scikit-image derives them by thresholding the gradient — guessing where objects are — where these come from a model that knows what sky is. Marker selection is what normally goes wrong with this method, and it was already solved. ## The gate still runs, and it runs first A flood cannot replace the colour gate. It only refines contours that already exist, and there is no contour around a flag precisely because the model never noticed one — the flag in the tests sits forty-five pixels from the boundary against a ribbon of six. The tests caught this; the first version of this commit had the flood standing in for the gate and the flag stayed. So the gate goes first and *creates* the contour, and the flood then puts every contour — the horizon and the new hole alike — onto a real edge. ## Erosion that does not delete flagpoles Eroding by a cell and a half destroys anything thinner than three cells: a mast, a bare branch, and equally a strip of sky between two of them. Those would be left unseeded and the flood would fill them from whichever side surrounds them, so a flagpole would come back — and come back *confident*. Erosion therefore stops at the ridge of the distance transform. Whatever would otherwise vanish keeps a one-pixel seed down its centre, floored at `min_thickness` so a hot pixel does not qualify. That floor also moves the signal-versus-noise decision out of colour space, where it was a share of a fitted distribution nobody can picture, and into image space, where it is a width in pixels a photographer can see. ## Two modelling errors the outward test found Both were invisible while the refinement could only subtract, because the gate was multiplied by weights that were already zero outside the mask. The moment the boundary could move outward they decided the answer. **A diagonal covariance is wrong along a gradient.** Sky moves along all three opponent features together — luminance up, red-green drifting, blue-yellow down — so treating them as independent charges a colour two deviations along that gradient three times over. Measured: sky fifteen rows past the sample scored 11.6 against a threshold of 11.34, so the model refused the very thing it was refining. The fit now carries a full 3x3 covariance, inverted by cofactors rather than by a dependency (D13, the NDK). **Eroded seeds understate the spread, always, in a known direction.** The sample is drawn from the middle of a category and never from its edge, so for anything with a gradient the colours nearest the boundary are exactly the ones left out. The broad mode is therefore fitted wider than its sample by `SHOULDER`. Same pixel: Mahalanobis 5.9 uncorrected, 1.5 corrected — the difference between refusing the horizon and reaching it. Only the broad mode is widened; the tight ones are what discriminate. ## What was given up Strict subtractivity. It bounded the damage and kept `scene.rs`'s partition true for free, and it had to go: a mask that may only shrink can sharpen a horizon inward but never outward, so wherever the coarse contour sat inside the true edge, the error survived every setting of the control. The travel bound replaces it. Everything beyond the ribbon is already a marker, so the flood never reaches it — not "can only remove" but "can only move this far", and the distance is the model's own uncertainty. That single bound also retires the connectivity test, the reachability radius and the separate additive path that an outward-growing rule would have needed. A blue car below the horizon cannot be gained, not because a rule forbids it, but because the flood is never there. `the_colour_gate_only_removes` keeps the older property where it still holds; `the_flood_cannot_travel_further_than_the_ribbon` holds the new one across the whole travel of the control. ## Cost The flood visits only unlabelled pixels, so confining it to the ribbon is not an optimisation added on top — it is what a seeded flood does. A ribbon of a few tens of pixels around one contour is a small part of a proxy. The distance transform is no longer cached, because it has to be measured from the mask as the gate leaves it and the gate moves with the control. That is one transform plus one flood per change of the control, against a precompute that runs the model once. Verified: fmt clean, clippy --workspace -D warnings clean, 63 dr-segment tests. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 21:29:12 +02:00
dtourolleandClaude Opus 5	710fcbc1bd	Format the face work with the workspace's own rustfmt ... Not authored in this session. `cargo fmt --all` reformats every crate, so running it while working on `dr-segment` picked up five files from the recent face and library work that had been committed unformatted. Committed on its own rather than swept into the change that happened to produce it: the diff is pure whitespace, and mixed into a commit that alters an algorithm it would be noise in exactly the place someone is trying to read carefully. `cargo fmt --all -- --check` is a CI gate (tools/ci-local.sh), so this had to land somewhere regardless. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 21:29:12 +02:00
dtourolle	89c4ff1820	Store what the model found, so a reopened photograph keeps its masks ... A subject or category layer was written to the sidecar as identity alone — which run, which instance, which category — on the reasoning that the pixels are reproducible by running the same model over the same image. They are, but only by *running the model*, and nothing runs one except a photographer pressing "find subjects". So on every path that did not already have a run in memory the layer resolved to no coverage, `MaskPass::render` logged "has no distance field; skipping", and the adjustment was silently absent: - reopening an edited photograph rendered it without its local adjustments, and then saved that state back on the way out; - a batch export from the grid could not have them at any point, because `render_from_library` opens a session, applies a version and renders, and there is no model anywhere on that path. Three hundred files written without the edits their photographer made, over a log warning. Neither failure announced itself. The generated shader still emits the layer's block and the empty placeholder multiplies it by zero, so the result is a well-formed frame that is simply missing an edit — `mask_is_stale` already named the state and called it "not stale, just unrenderable". The coverage now travels in the file, as one `coverage = w h levels payload` line at the end of the layer's block. Two levels, and that is not a compromise. The model hands out a byte per pixel but `Shaped::build` measures its distance field from `coverage >= 128` and throws the shoulder away on the first line; everything soft about the rendered edge comes afterwards from the layer's feather and falloff, which are read off the distance. So one bit per pixel is not an approximation of what the model said — it is exactly the part of it that reaches a pixel, and the stored mask renders the identical frame. Storing all 256 levels would have stored 1.7 MB of bilinear interpolation to reconstruct a predicate, and would not even have compressed: a model mask is a bilinear upsample of a coarse grid, so almost no two adjacent bytes are alike. Measured on a simulated sky and a simulated figure at 1600x1067, against 1.71 MB raw: 4.0 kB and 6.5 kB at two levels, 46 kB and 76 kB at sixteen, 835 kB and 1.43 MB at all 256. The level count is still written into the line, so a later build that finds a use for the shoulder can write sixteen and this one will read them rather than misreading a stream of lengths as pairs. The coder is hand-rolled — run-length pairs in a base-64 varint — because `dr-pipeline` links nothing, which is the property that lets the descriptor and codegen logic be tested without a device. `flate2` would have been fewer lines and a dependency in the one crate that has none. Where it lives matters more than how it is coded. The raster sits on `MaskLayer` beside the source, not inside `MaskSource::Subject`: the source is *identity*, which is what makes it diff as a handful of numbers and merge per field under FR-NC-9, and a raster in there would have given the merge a binary blob to arbitrate. It takes no part in `MaskLayer`'s equality for the same reason — a device that has run the model and one that has not hold the same edit, and counting the difference would raise a conflict over a cache and let `remote_wins` answer it by discarding the only copy of the pixels. Encoding happens in `masks_for_storage`, on the save path, rather than in `ensure_subject_fields` where every coverage already funnels through. `ensure_subject_fields` runs on a drag — dilating a mask with a compound morphology rebuilds the field every frame — and encoding a megapixel raster per frame is the kind of work NFR-P5 exists to keep off a gesture. Saving happens once, when the photograph stops being the open one, and already costs a network round trip. Version skew holds both ways. A file with no `coverage` line reads exactly as it did before, which is a layer that needs the model run; an unreadable one costs the pixels and not the layer, because the layer is the edit and the raster is a cache of it. An old build reading a new file drops the key it does not understand, which costs a model run and no work. And a payload that will not compress is refused rather than truncated: a checkerboard would encode to twice the raster it came from, so past 64 kB nothing is stored and the behaviour falls back to what it was — half a mask would render as a mask that is confidently wrong, which is the failure that tells nobody.	2026-08-30 21:28:54 +02:00
dtourolle	6acc98baad	Carry the photograph's header into an export made from develop ... The same file exported from the library grid kept its camera, its lens, its capture date and its rights statement. Exported from the develop button it kept none of them, and `{date}` in a filename template resolved to nothing at all. Two buttons, one photograph, two different files -- and the develop one was the version the photographer had just finished working on. A session now remembers the header it was opened from, and `open_session` takes that header rather than the orientation read out of it, so a photograph cannot be opened for editing without saying which file it came from. `render_open_frame` clones it onto `Source::Rendered`; both arms of `export_one` -- the worker's own decode and the frame handed over already rendered -- turn a header into a `{date}` and a `SourceMetadata` through the same function, so the two paths cannot come to different readings of one file. What of it actually reaches the exported bytes is still decided inside `dr-export` from the settings, which is what keeps the location-stripping option working here rather than giving it a second implementation to disagree with. The alternative was to hang the metadata on `Source::Rendered` alone and keep it beside the session in the interface. That touches less, but it makes the header and the pixels two cells to hold in step across the six places an image is opened, replaced or fails to open, and the failure mode of getting that pairing wrong is not a missing tag: it is one photograph exported under another's byline and coordinates, silently. Kept on the session, the two travel together or not at all. The header is stored decoded rather than transcribed at open time, deliberately. `dr-export` argues that source metadata is a parameter and not a field on `Frame`, because two exports of one frame may legitimately disclose different amounts; by the same reasoning a session may remember where its pixels came from without that being a decision about what to publish, and the allowlist that decides remains the single function in `export.rs`. A file with no header is left with none -- an empty `{date}` and nothing for the encoder to copy -- rather than today's date standing in for a capture time nobody recorded.	2026-08-30 21:28:31 +02:00
dtourolleandClaude Opus 5	353382c07f	Hand the photographer's place between devices ... A place recorded on the tablet should be where the desktop opens. Exchanged through `.darkroom-derived/place.json`, beside the thumbnail shards and the catalog snapshot. Newest timestamp wins outright: unlike the catalog this is replaced rather than merged, because two devices cannot both be where the photographer is and so there is nothing of theirs inside ours to preserve. It still refuses to upload over a copy it could not read, for a smaller version of the reason `sync_catalog` does: a record we have not compared against may be the newer one, and overwriting it would move the other device's photographer without ever having seen where they were. Last in the pass, and its failures are logged rather than reported. Everything else in that folder is *derived* -- a faster way to learn what the device could work out for itself -- so losing it costs time. A place is a fact only the other device knew, and losing it costs a scroll. A sync that ran out of connectivity should spend what it had on the shards. The full pass runs after a thumbnail sweep or when Sync is pressed, neither of which happens on an ordinary launch -- so a handover would arrive one launch late, which is one too many for a feature whose whole claim is picking up where you stopped. `spawn_place_fetch` is the small half: one GET of a few hundred bytes, started beside the scan. And it can still be refused. A handover is welcome on the way in and unwelcome once the photographer has started: a grid that jumped elsewhere mid-scroll because a round trip finally landed would have lost their place to the feature meant to keep it. Any scroll, scrub, scope change, filter or opened photograph closes the latch, and a record arriving after that is written to disk and takes effect next launch. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 20:40:22 +02:00
dtourolleandClaude Opus 5	d44bffa4a8	Remember where the photographer was ... Opening the application was always a fresh arrival at the beginning of the library, whatever you had been doing when you closed it. What is written down is the view, the scope, the rating filter and the photograph on screen -- the open one in develop, the first visible one in the grid. Not just a scroll position: a position without the filter that produced it names a row of a list that no longer exists. Restoring them has an order for the same reason -- scope, then filter, then position, then the view -- because each step changes what an ordinal *means*. Addressed by remote path and collection UUID, never by an ordinal or a row id. `images.id` and `collections.id` are local to one catalog, and a grid ordinal is local to one ordering; a record naming either would land somewhere arbitrary on a second device and after any filter change on this one. Where the ordinal is needed, `library::ordinal_of_path` computes it through the grid's own `ORDER BY`, taken verbatim by a window function rather than spelled a second time as an inequality -- which is the mistake `grid_order_for` already warns about, and which a manually ordered collection would make unreadable. Every failure degrades rather than reports. A collection this device has not merged leaves the scope at the whole library; a photograph that has since been deleted falls back to when it was taken, which puts the grid in the right week; a torn file yields no place and the library opens at the top. Reopening develop is the one thing that requires an exact match, because a canvas on a path that no longer resolves is a filename over an empty frame. The record lives in `dr-types` beside `Settings` and the store lives here beside `SettingsStore`, for the reason `dr-types`' manifest gives: a JSON serialiser in `core/` would be paid for by every crate there. Two files and two lifetimes, though -- resetting preferences must not forget where you were. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 20:39:50 +02:00
dtourolleandClaude Opus 5	8bf5e13faf	Centre the photo roll on the frame it opens with ... The roll brought the open photograph into view by the shortest move, which is right for stepping along it and wrong for the first look: a frame near either end of the loaded window arrived hard against an edge, with nothing on that side to give it any context. It now centres on the first settle of a develop session and steps minimally after that. A one-shot request that the strip itself clears -- the only thing that knows the request has been honoured is the code honouring it -- rather than something recomputed on creation, because the strip is created far more often than a session begins: leaving develop for Settings and coming back rebuilds it, and re-centring then would undo a roll the user had scrolled by hand. Raised on the two ways into develop from the grid, and not on a pick along the roll, which is a step within a session rather than the start of one. Centring is clamped to the ends: the third photograph of a window cannot be centred without scrolling empty space in beside it, and a strip that begins with a gap reads as broken rather than as centred. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 20:35:55 +02:00
dtourolleandClaude Opus 5	0ff1e01ec3	Come back from develop on the photograph you were editing ... Leaving develop returned to where the *grid* was, which after a walk along the photo roll can be a thousand rows from the frame you had just finished. So the one photograph you were certainly interested in was the one the grid came back without. Two positions, and the rule is not to pick one of them. The grid seeks to the remembered position, then reveals the keyboard cursor -- which is now put on the open photograph, and which moves the viewport as little as will bring its row into view. A frame inside the remembered screenful moves nothing at all; one outside it scrolls exactly far enough. One rule, both behaviours. The cursor rather than the selection, deliberately: `place_cursor` also rewrites the selection, and a set of forty photographs assembled in the grid must survive having one of them opened. `reveal()` now also runs on the grid's `init`, since `cursor-row` is initialised rather than changed when the subtree is rebuilt and no handler would otherwise fire. Both it and the roll's centring defer while the element has no height yet -- `init` runs before layout, where a height of zero makes every row look off screen -- and a latch brings the first real height back to the cursor without letting every later resize haul the viewport around. The capture-time marker follows the same move, for the same reason: `load_window` rebuilds the axis only when the scope, the filter or the total has changed, and none of them has. It is the same library seen from a different row. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 20:35:23 +02:00
dtourolleandClaude Opus 5	53dc2d171e	Say where the grid is on the capture-time axis from the first frame ... The sidebar's marker rested greyed at mid-track until the first scroll or scrub. The reasoning was that anchoring it would imply a choice the user had not made -- but that reads the marker as reporting an intention, and it does not. The sidebar's whole claim is to say *when* you are, and that is known from the first frame: the grid is at the top of the library, or wherever it was last left. So a launch opened with the marker halfway down an axis whose visible photographs were all from the wrong end of it. Dimmed rather than absent, which made it look like a reading rather than the absence of one. Seeded in `refresh_timeline` -- the one place that decides what the marker says, and the one that runs on every route which builds the axis -- and only when nothing has claimed it, so a scroll or a scrub still speaks for itself. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 20:34:11 +02:00
dtourolleandClaude Opus 5	eed27eb36d	Keep the grid's place when another screen covers it ... Opening Settings, Import or People and coming back landed at the top of the library however deep in it you had been. The grid is gated on an `if` in the markup, so every route away from it destroys the subtree and rebuilds it. A Flickable being destroyed passes its viewport through zero on the way out, and that reaches `on_library_scrolled` looking exactly like the user having flung the grid to the top. The handler already guarded against it -- but on `show-library`, which means "the library rather than develop" and stays true while any of those four screens replaces the window. So the guard covered the develop route and none of the other three: `resume_at` was overwritten with 0 on the way out, and the position was gone before anything could restore it. The condition the `if` is actually spelled with is now computed once, in `app.slint`, and Rust reads that. The two cannot drift apart again because there is only one of them. That fixes the overwrite. The second half is that nothing replayed the position on the way back in: `on_back_to_library` does it by hand, and Settings, Import, People and the launch screen do not go through it. Rather than teaching three more modules to call it, `scroll-to` is now kept current on every scroll. It is read by `seek()`, which runs on a token change and on `init`, so writing it without bumping the token cannot move the grid on screen -- and is exactly what the next grid reads when it is built. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 20:33:34 +02:00
dtourolleandClaude Opus 5	4dc954f01a	Take the photo roll's grab band off the buttons that end a mode ... "Done Cropping", "Done Repairing", "Done Masking" and "Fit" float over the foot of the canvas. So does the photo roll's swipe handler, and a gesture handler is not a layout box — it is an input surface. A press inside one is delayed, then offered to that handler's own children and to nothing else: `input_event_filter_before_children` returns `DelayForwarding`, which aborts the hit-test traversal outright, and the replay afterwards visits only the handler's subtree. Everything behind it is never asked, hover included. The band was `strip-height + reach` — 136px along the bottom — whether the roll was out or away. So the button that ends a mode was drawn, was lit, and did nothing for as long as a library was open, which is the whole time anybody is developing from one. The tool rail kept working because it is a sibling of the canvas rather than behind the roll, which is exactly why this looked like two dead buttons rather than a dead region. The band now goes where the roll goes. The handler carries the strip instead of standing still while the strip animates inside it: closed, only `reach` is on screen and the rest hangs below the window where nothing can press it; open, it still covers the thumbnails, which is what lets a swipe down anywhere across them put the roll away. The 180ms travel moved from the strip onto the handler, so the drawn positions in both states are what they were. The controls are then positioned against that band rather than against the bottom of the canvas, and ride up with the strip when it comes out. Reordering them in front of the roll would have been the other fix, and it is the wrong one — the band would become the thing that cannot be reached, and a gesture nobody can start is worse than a button with a second way out. `roll-strip` and `roll-reach` are tokens now, because two files have to agree on where that band is for either of them to keep out of it. The bottom of the photograph comes back with it: the crop's lower handles and a repair placed near the bottom edge were inside the same 136px and had the same fault. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:42:17 +02:00
dtourolleandClaude Opus 5	3d248cfb79	Export the photograph, not the canvas ... Zooming the develop view changed the exported file. `Framing::view` is kept out of the sidecar, out of `is_active` and out of `output_size` precisely so that it cannot — but those exclusions keep it out of the *edit*, and an export is a *render*. `visible_rect` deliberately folds the view into the single rect the fused shader's prologue samples, so `render_for_export` inherited it: at 4:1 it wrote the middle of the frame, magnified to fill the file at the full output size, with the detail kernels scaled four times over because `render_scale` folds the view in as well. `render_thumbnail` did the same to the grid. `render_uncropped` already suspends the view for this exact reason, so the fix is its pattern: one `render_the_file` that both file-producing paths go through, composing inside the suspension since the view reaches the shader as a uniform baked at composition time. Restored whatever happens — leaving the graph un-zoomed after a failed export would throw away where the photographer was looking. Nothing caught it because the guard checked the wrong things. `zooming_does_not_change_the_exported_image` asserted the output size and the crop; both held perfectly throughout. Renamed to `zooming_does_not_change_the_size_or_the_crop`, which is what it tests, and the pixels are now guarded where pixels exist. The new test uses a ramp rather than quadrants deliberately: a four-quadrant frame is self-similar under a centred zoom, and the first version of this test passed against the bug because of it. Traces FR-EXP-9, which asks for the full-quality pipeline "regardless of what the display was showing". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:30 +02:00
dtourolleandClaude Opus 5	a1e361e35a	Measure the native path against the one it replaces ... Everything argued for this change so far was read out of a catalog after the fact: crop_px across 18,671 faces the old code had already stored. That is evidence about what the previous implementation did. It is not evidence that the new one does better, and the difference matters because a landmark left in the detector's coordinates, or a box filter with an off-by-one in its source span, would both produce faces that look entirely plausible until somebody counted the pixels behind them. So examples/face_native.rs renders one file and indexes it twice, native and from a 1024 proxy, changing nothing else. Fourteen originals from the reference library, 5472x3648 CR2 and DNG: native 9 faces, mean crop 287px 1024 proxy 5 faces, mean crop 75px Crops 3.8x larger, and across the line that decides whether the crop is photographed or interpolated: 75px is below ALIGNED_EDGE, so the proxy path was upsampling into the embedder on average where this one downsamples into it. Fourteen images and nine faces is enough to show a direction and to catch a wrong scaling; §7b says so rather than quoting the ratio as a library-wide figure. It also corrects something §7b asserted two commits ago. I wrote that detector input resolution cannot affect recall, because §4.1 letterboxes everything to 640. Native found nine faces to the proxy's five, including four on files where the proxy found none, so it plainly can. The two paths differ in their resampling as well as their size, and this experiment does not separate those, so §7b now records the result as evidence for the double-resampling hypothesis rather than as its proof. M4 still owns settling it. The audit-summary test went stale when the ready/to-fetch split was collapsed and is updated to assert the single number, including that the old wording is gone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:30 +02:00
dtourolleandClaude Opus 5	4af3b93dfa	Index faces from the native render, not from a preview of it ... Implements the FR-CULL-8 written two commits ago. The sweep fetched the JPEG preview embedded in each RAW and used that one buffer for both detection and the crop; it now fetches the original, renders it through the same path export uses, reduces that for the detector, and warps the crop back out of the native frame. Three pieces, and each exists for a reason worth stating. dr_face::Pixels lets the warp sample 8-bit RGBA directly. A 24 MP native frame is 96 MB as RGBA and 288 MB converted to the f32 RGB align.rs was written against, and the warp reads about forty thousand pixels out of it. Converting the whole frame to sample 0.2% of it is NFR-RES-2's budget spent on a copy, per image, for a whole library. The variant costs one branch per sample and a test asserts both layouts produce identical crops. The detector gets a box-filtered reduction to 1600px, not the native frame and not a point-sampled one. Averaging rather than sampling because the detector's job is finding small faces and decimation is precisely the operation that removes them: at 4x, fifteen of every sixteen pixels are discarded and a 40px face survives or not depending on where it falls relative to the sample grid. 1600 rather than 640 leaves the letterbox a mild 2.5x rather than a 9x, and bounds the f32 buffer at 20 MB. Landmarks come back in the reduction's coordinates and are scaled to native in one place before any crop pixel is read. This is the failure mode that would not announce itself -- unscaled landmarks put every crop near the top-left corner, which yields faces of something else, cleanly embedded and confidently clustered. The sweep fetches SWEEP_LANES-wide and renders sequentially. Not a placeholder for a parallel version: there is one GPU, so concurrent renders queue on it regardless, and each materialises a native frame. Overlapping them would multiply the one allocation that threatens the memory budget while buying parallelism that does not exist. The chunk drops from 96 to 6 for the same reason -- 96 held 8 MB previews, this holds whole RAWs. The stored edit is deliberately not applied, which is where this departs from export::render_from_library. Face geometry is normalised to the frame, so indexing a cropped render would record boxes against a frame that changes whenever the user changes their mind, and every stored box would quietly become wrong. Orientation is applied: that is a fact about the file rather than an edit. examples/face_native.rs renders one file and indexes it both ways, so the claim behind all of this can be checked against photographs rather than re-read out of the catalog it came from. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:30 +02:00
dtourolleandClaude Opus 5	9ddc1273c0	Make a sweep that fails everything say so ... A run over 169 images failed all 169, in fourteen seconds, and reported "0 face(s) in 0 image(s)" -- the same sentence a run that indexed nothing because there was nothing to index produces. Three separate places dropped the information on its way to the screen. The progress count only moved on success. FaceSweepMessage had no failure variant at all, so a pass where every image failed sat at 0/169 from the first tick to the last: the receiver was told the total, told nothing, and told the pass had ended. That is indistinguishable from a hung job, and it is what it was taken for. Finished already carried a failed count and identity_ui matched it with `Finished { .. }`, throwing the number away and printing the tidy success line regardless. And the reason each image failed was logged at debug, which is off, so 169 consecutive failures left no trace of why anywhere. Failed { images } now carries the count back per lane batch, the progress counter advances on it, and both the running status line and the finishing activity row say how many could not be read. A batch rather than one message per image because failures come back lane-sized and the useful number is how many. Also renames the store sweep's guard to MIN_CROP_EDGE with the rest of that constant's move, since the two touch the same lines. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:20 +02:00
dtourolleandClaude Opus 5	e7b526c550	Specify face indexing at native resolution, and say what the proxy cost ... FR-CULL-8 said detection runs against the thumbnail or proxy tier and never a full decode, and faces.md §5 said the aligned crop is sampled from that same proxy. Both are wrong in the same place: they treat detection and cropping as one resolution problem when they are two, with opposite answers. Detection does not care. §4.1 fixes the graph's input at 640x640 and letterboxes whatever arrives, so a face filling 2% of the frame reaches the model at 12px whether the buffer handed over is 1024px or 6000px. Every pixel above the detector's own input is discarded before inference. The crop cares about nothing else. §5's warp produces the fixed 112x112 ArcFace sees, so source resolution converts directly into whether those 112 pixels were photographed or interpolated. Reading crop_px across the 18,671 faces the proxy-tier implementation stored: 47.3% were upsampled to reach the embedder, 314 of them by more than 2x, the smallest from 34 source pixels. An upsampled crop does not fail loudly -- it yields a confident embedding of detail that was never there, and the damage appears three stages later as clusters that will not separate. So FR-CULL-8 now specifies four stages with the resolutions named separately: render native through FR-EXP-9's pipeline, downscale for the detector, map boxes and landmarks back to native, crop and align from the native render. The affordability the old rule bought is met instead by when the pass runs -- background, preempted, resumable -- and the requirement says plainly what it now costs on a remote library: the original rather than FR-NC-3's byte range, 412 GB across the reference library's 19,107 images, so a whole-library pass is a transfer under FR-NC-6 rather than something that may start on its own. MIN_CROP_EDGE replaces the MIN_DETECT_EDGE this branch briefly had. Same number, guarding the quantity that turned out to matter. faces.md §7b records both measurements, and marks the second as unexplained rather than dressing it as a finding. Grouped by the buffer detection ran against, faces per image was 0.078 at 1024 or below and 1.82 at 2048 or better, controlled for file type and size. That gap is real and reproducible and I cannot account for it, because the letterbox above says detector input should not matter. M4 is where it gets settled. The crop measurement does not depend on it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:20 +02:00
dtourolleandClaude Opus 5	144d2e4e84	Revert the detection floor: it guards the wrong resolution ... Reverts 53f7cdf and e92d22d. The floor those added sat on Detector::detect, refusing any buffer under 1025px on the reasoning that a small buffer finds no faces. That reasoning does not survive §4.1: the detector letterboxes every input to 640x640, so a face occupying 2% of the frame presents at 12px to the model whether it is handed a 1024px buffer or a 6000px one. Detector input is precisely the quantity that does not matter. Worse than merely useless, it blocks the design FR-CULL-8 now specifies, where the detector is deliberately fed a downscale and the crop is taken from the native render. A guard on detect() rejects exactly that call. What the measurement actually supports is a floor on the *crop* source, which is where resolution converts into embedding quality, and which faces.crop_px already records: 47% of the reference library's faces were upsampled to reach 112x112. That floor is a separate change against the native-resolution path and does not belong on the detector. The 23x faces-per-image gap by source_edge that motivated the original commit is kept in faces.md §7b, restated as the unexplained observation it is rather than the causal claim it was written as. V12 stands: those runs cropped at 1024 whatever detection did, and that is reason enough to look at them again. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:12 +02:00
dtourolleandClaude Opus 5	41c655c176	Stop the local-store pass pretending it can still index ... spawn_store_face_sweep detects on what the thumbnail store holds, and the store's largest tier is FACE_TIER -- 1024, which the floor now refuses. Left alone it would list every outstanding image, decode every proxy it had, and report every single one as failed: twenty thousand refusals all saying the same thing, with the real explanation buried at debug level. There is no repair available inside this pass. It has no larger tier to read; the pixels the detector needs have to come off the server, which is spawn_face_sweep's job and always was. So the honest behaviour is to check the tier against the floor once, say plainly which pass to use instead, and stop. It is only reached from examples/face_index.rs, so this costs the tool its --run mode and no shipped behaviour. FACE_TIER keeps its value and loses its meaning. It is now the tier a stored crop is *cut from*, which 1024 is entirely adequate for -- the face has already been located and the crop only has to be looked at -- and no longer the tier faces are *found* on, which is the thing that was returning 0.078 faces per image. IndexAudit's ready/awaiting_proxy split goes the same way. It existed because one of the two passes could only do images that already had a proxy; now that detection refuses that proxy's size, both halves cost the same fetch, and a status line reading "169 ready to index" implies a distinction that no longer decides anything. One number. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:04 +02:00
dtourolleandClaude Opus 5	d0ebc9f571	Count the same images in the progress figure that the sweeps index ... The Identity screen said 4,593 images were left to index and stayed there for hours across repeated runs, which is what a stuck job looks like. It was not stuck. 4,424 of those 4,593 are shadowed -- the JPEG half of a RAW+JPEG pair -- and no sweep will ever index one, because every work list is built on VISIBLE, which excludes them. They are not separate photographs and the grid does not show them either. But faces::coverage counted them: its denominator was "images WHERE trashed_at IS NULL", with no shadowed_by clause. So the outstanding figure had a floor of 4,424 that no amount of work could bring down, and Coverage::is_complete could never once return true no matter how completely the library had been indexed. A progress number that cannot reach its own target is worse than no progress number. The fix is to count the population the sweeps actually draw from, in all three places that were describing it differently: coverage's denominator and its indexed join, and audit's split of the outstanding set, which had the same gap and fed the same status line. On the reference library the denominator goes from 23,531 to 19,107 and outstanding from 4,593 to 169 -- the second of which is a number the user can watch go down, and which turns out to be a real and separate fetch failure worth chasing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:04 +02:00
dtourolleandClaude Opus 5	6783d0c723	Settle the images whose best preview is under the floor ... The floor introduces a state the sweep had no arm for. An image whose largest embedded preview is genuinely smaller than 1025px now comes back from index_preview as an error, lands in the generic Err arm, and is counted as failed -- which means no face_index row, which means it is still outstanding, which means the next sweep fetches exactly the same bytes and refuses them again. For ever, on every run, at one range request each. The previous behaviour was wrong but at least terminated; this would not. So ProxyTooSmall gets its own arm, and it records a marker at the true edge rather than nothing. That is the difference between "we looked and found nothing" -- which would be a lie, since nothing was looked at -- and "this was examined at 900px, which is the best this file has". The first is unrecoverable; the second is a fact source_edge was added to carry, and a later floor or a bigger proxy can select on it deliberately the way V12 just did. Counted separately from failures all the way up, because they are not failures and reading them as such would misdescribe a library of small scans as a broken network. The summary line says how many and why. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 19:41:04 +02:00
dtourolleandClaude Opus 5	7981718d83	Time the phases of a launch, because the tablet has no profiler ... Two things are now off the launch path and the rest of it is unmeasured. There is no way to attach a profiler to an Android launch, and the window that matters — from `android_main` to the first `poll_events` — is over before anything on the device can be asked a question about it. So a line in the log file is the only measurement anybody gets. Three of them: the GPU open, the window build, and the total to the event loop. The last is the one that matters, because it is the figure the input dispatcher is counting against — anything approaching five seconds there is the next ANR whatever the phases above it say. The GPU open is timed rather than moved. It is a Vulkan instance, an adapter enumeration and a device request, and on the desktop it cannot be deferred at all: it selects the Slint backend, and creating a window selects one for us. On Android it could be, because nothing shares that device with the compositor (TD-1) — but "could be deferred" is not "costs enough to be worth deferring", and there is no number yet that says which. This is the line that will produce one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:41:34 +02:00
dtourolleandClaude Opus 5	d48e9f6033	Open the catalog on a worker, so a launch is not a page-by-page read ... The second thing standing between `android_main` and the first `poll_events`, and the one that grows with the library rather than with the APK. `library_ui::open` called `show_catalog_now`, which called `Catalog::open_verified`. That runs `PRAGMA quick_check`, which reads every page of the database, and then `Catalog::open`, which takes a full SQLite backup of the file before a migration and rewrites its structure afterwards. On a 50,000-image library that is tens of megabytes of I/O on a tablet's flash, and it happened before the window had painted anything — so on Android it was counted against the five seconds the input dispatcher allows, and on the desktop it was a launch that sat on a blank window. `dr_catalog::recovery`'s own module documentation says the check is affordable "at startup, where a failure has a user in front of it who can answer a question". That was the intent and it was not true: there was no interface yet in which to ask. Now there is, because the open happens on a worker and the answer arrives on a channel drained by a timer — the same shape the scan, the thumbnails and the login already use. The gate the synchronous call provided is kept, and is the reason the scan moved with it. `Catalog::open` succeeds on a damaged file whose header survived, so a scan running beside an unanswered recovery question writes ETags and image rows into damaged pages and turns a catalog that had a backup into one where the backup is the only copy left. So the scan now starts from the drain, on the two answers that permit it, and not at all on `Corrupt`. `library-scanning` stays true throughout, which hides the Rescan button and stops the gate being merely advisory. What the user sees while it runs is a third empty state. The grid already refused to conflate "still scanning" with "scanned, found nothing"; "opening the library" is a third answer and it gets its own sentence, because a grid saying "Scanning…" while nothing is on the network is the same kind of lie the other two were separated to avoid. `show_catalog_now` stays, unchanged and blocking, for `recovery_ui`. That call site has the event loop running, has just replaced the file under a `forget_catalog`, and has `recovery-busy` on screen — the same reasoning `recovery_ui::answer` already gives for doing its file copy in place. The part both paths share is now `adopt_catalog`. One consequence worth naming: the cache-usage figure on the settings page was read at startup from a catalog that is no longer open by then. It moves to the page's `on_open` closure, beside the face coverage, which is read there for exactly the same reason — it is only ever looked at while that page is on screen. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:41:34 +02:00
dtourolleandClaude Opus 5	1c5849ebe8	Unpack the bundled models on a worker, not on the way to the first frame ... Launching v0.10.0 on the tablet produces an ANR: "Waited 5000ms for MotionEvent", 5,827 ms on one input sequence, over a window that has never painted. The app recovers and sits at 0% afterwards, so it is a startup cost rather than a hang. The structural fact behind it is that `android_main` runs with the activity's input channel unserviced. Nothing drains it until Slint reaches `poll_events`, and Slint does not reach `poll_events` until `dr_ui::run` calls `window.run()` on its last line. Every millisecond before that is a millisecond the input dispatcher waits on, so five thousand of them is an ANR whatever the work happens to be. The largest single piece of that work was here. `install_bundled_models` copies 41 MB on the first launch after an install — 24.9 MB of scene model, 13.6 MB of embedder, 2.5 MB of detector — each read whole out of the APK into a `Vec` and written to `/data`, in a loop, on that thread. v0.10.0 is the release that added the scene model, which is 60% of that total, and it is the release the ANR appeared in. The 8,010 minor faults in the report are about what 41 MB of freshly touched pages costs. So it moves to a detached thread and the function returns as soon as the thread is running. Nothing on the launch path wanted the result: the only two things that read these files are the People screen and the scene tab, both of which are reached by hand, minutes later, from workers of their own. What that costs is a window in which a model looks absent. `library::face_models` and `library::scene_model` decide availability on `is_file()`, so during the copy both report their feature unavailable — which is the same answer they give a build shipping no weights at all, the ordinary case both were written around. Briefly pessimistic rather than wrong, and the temporary-name-then-rename that was already there is what keeps it from being worse than that: a lookup never sees a half-written file, only an absent one. Both call sites now say so. A completion line reports the bytes copied and the milliseconds taken, including when it is zero, so the second launch after an install can be told from the first in a log rather than by inference. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:41:34 +02:00
dtourolleandClaude Opus 5	4efab496c2	Put the refinement on a slider, per layer ... The evidence was being gathered and spent immediately at one strictness nobody could see or change. This makes it a control. `MaskLayer::refine` is shaping, like the feather, and it lives on the layer for the layer's reason: two layers may sit on the same category and want different amounts of it, and the model ran once for both. ## What the segmentation now stores `CategorySummary` keeps the **coarse** mask and the `Refinement` beside it, rather than a refined mask. That is what gives the control an off position that is bit-for-bit the model's own weighting, and what stops a strictness change from needing the model again. `category_mask_at` borrows at zero and wherever no refinement could be fitted, so a layer nobody has touched costs nothing over the old path. The fit moved to the far side of the orientation permutation. The verdict is a per-pixel field over the same grid as the mask it gates, so fitting it upright would mean permuting a proxy-sized buffer afterwards to match — a second rotation, and a second chance to get one wrong. One logit cell is the same number of pixels either way: the letterbox scales by the longer edge and a permutation does not change which edge that is. The two frames became a `Frames` struct rather than six parameters. This function reads the picture twice for opposite purposes — the model needs it upright or it recognises far less, the refinement needs the sensor's grid — and a transposed pair produces a plausible mask over slightly the wrong pixels, which is the failure this module is most prone to. ## It rebuilds the field, and the signature says so `refine` is mixed into `subject_signature`. Unlike a feather, which is read off a field that is already correct, this changes which pixels are in the mask at all — so it changes the coverage the field is measured from. Omitting it is the bug where the slider moves and nothing happens until some unrelated control invalidates the cache. That puts it in the same cost class as a close or an open, which is why the row takes `SliderRow::changed` — already once-per-gesture, since that row takes `SliderTrack`'s `committed` internally — rather than a live stream. The slider is offered only where there is something to move: a category source, *and* a refinement the frame actually gave enough to fit. A control that moves and does nothing is worse than an absent one. ## Two defaults that are deliberately different A layer added from the panel starts at 4.0, because a category's edges are twenty proxy pixels wide before anything is done to them and a photographer adding a sky mask wants the sky rather than the sky plus every chimney in it. A layer read from a sidecar with no `refine` key starts at **zero**. A file written before this control existed has to render as it did then, and a default of 4 on absence would quietly re-grade every stored category mask in the catalogue. `a_categorys_refine_strictness_survives_and_defaults_off` holds both halves, and `an_out_of_range_refine_is_clamped` holds the file to the scale — past the top of it every colour fails and the mask deletes itself, which reads as lost work rather than as a bad file. `MAX_REFINE` is dr-pipeline's own constant mirroring `dr_segment::STRICTNESS_MAX`, following `Falloff` and `Morphology`: this crate holds the description of an edit and must not depend on the crate that runs a model. dr-ui is where the two meet, and the only place that converts. Verified: fmt clean, clippy --workspace -D warnings clean, 488 dr-pipeline and 60 dr-segment tests. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:30:40 +02:00
dtourolleandClaude Opus 5	4f4abd335f	Cut a scene category back to the pixels that agree with it ... A flag in the sky came out weighted as sky, and no feather setting fixed it. The scene model's logits are `[1, 150, 80, 80]`, so one cell is eight input pixels; at the 1600px proxy the letterbox scale is 0.4 and **one cell is 20 proxy pixels**, which `rasterise`'s bilinear then spreads across one more either side. A flag is a handful of cells whose softmax is dominated by the sky around it. The information was never in the grid, so nothing downstream of the grid can recover it. Tiling is the answer for an instance and is not available here: a category has no bounding box to tile over — sky is wherever the sky is. But the photograph is at full proxy resolution even though the weights are not, and it knows exactly where the flag is. So the model says *what*, and the pixels say *which of them*, which is the division of labour arm C already draws between the instance model and the watershed. ## Seeds, and why the erosion radius is not a guess Threshold the weights high, take `signed_distance`, and keep what is more than 1.5 cells inside. One cell *is* the model's resolution and the bilinear spreads it across one more, so the band either side of the boundary is smear rather than evidence. Deriving the radius from `Scene::cell_pixels` rather than picking a pixel count means it stays right if the proxy edge or the export changes. The mirror of that set is a confident *exterior*, free from the same field. ## Dropping small modes is the step that makes it work Four k-means modes per side, not one Gaussian: sky is blue at the zenith, white where the cloud is and pale at the horizon, and one blob over all three rejects two of them. Then modes holding under 3% of a side are discarded, and without that step the whole thing fails on the case it was built for. A small flag deep in the sky has both a high weight and a large distance from the boundary, so it lands in the interior sample and teaches the model its own colour. It cannot be excluded geometrically. It can be excluded by share. Luminance is weighted at a quarter against chrominance for the same reason the watershed's gradient is. Sky's variance is dominated by luminance, so at equal weight the distribution is a long bright streak that a mid-grey flag sits comfortably inside. A flag is separated by chrominance; a cloud is separated by luminance alone. Not zero, or a dark bird against a bright sky survives. ## Two tests, because either alone is wrong Absolute — is this colour plausible under the category, as a chi-square on the Mahalanobis distance. Comparative — is it likelier inside than outside. A pixel must pass both. The absolute test is what catches the flag, whose colour is far from *both* sides and which the comparative test alone would leave at even odds. The comparative test is what stops the absolute one needing a constant tuned per category. ## What this cannot do, written down rather than left to be discovered An intruder large enough to hold its own mode is kept. By share, a flag over a fifth of the sky and a cloud bank over a fifth of the sky are the same object, and colour does not separate them either — a white cloud is as far from blue sky in chrominance as many intruders are. So `min_cluster` is not a threshold with a correct value waiting to be found; it is the trade-off itself, set where a photographic intruder falls. Both ends are pinned by tests — `a_flag_in_the_sky_is_removed` and `an_intruder_larger_than_min_cluster_survives` — so that moving the number reads as moving the trade-off rather than as fixing a bug. The case left open is a large unrecognised object in a clean category, which wants the boundary snapped to watershed basins and is a different mechanism. ## Safe to apply without a control It is subtractive: the output is the input times a factor in `0..=1`. The worst failure available to it is losing part of a real sky, never gaining a region, so a blue car below the horizon that was never in the mask cannot be pulled into it. And a factor in `0..=1` cannot raise a sum, so `scene.rs`'s partition still holds when every category is refined independently — the weight taken off the flag lands in the unlisted remainder, which is where a flag belongs, ADE20K having no class for one. Every path without the evidence to judge returns the weights untouched and says which path it took. A refinement that silently did nothing is indistinguishable from the feature being off, and an empty seed set fitted to a distribution would reject every pixel. The signature is deliberately unchanged: categories are addressed by name, not by index, so a sharper mask cannot create the stale-index hazard the signature exists to guard against. The example writes `<prefix>-<category>-refined.ppm` beside the coarse one, never instead of it — whether this is an improvement is a comparative judgement and one image cannot answer it. Verified: fmt clean, clippy -D warnings clean, 57 dr-segment tests. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:30:16 +02:00
dtourolleandClaude Opus 5	bfdb6d5e4b	Cut the People rail's portraits off the blocking path ... Opening Identity cut a portrait for every person before the screen was allowed to appear. Measured against the reference library: 3.6 seconds, of which 3.2 is 931 full 1024px proxy decodes — the fallback for faces indexed before crops were stored beside them. The rest of the wait was the rail building 14,268 rows, fixed in the commit before this. `refresh` now draws with the portraits already cached and hands the rest to `fill_covers`, which cuts them 8ms at a time — under half a frame — behind a screen that is already up. Blocking work on that path goes from 4,965ms to 16ms on this library. A timer rather than a thread. The work is a catalog query and a decode against a `ThumbStore`, and both handles live on the UI thread; a worker would need its own connection to the same file, which is what the sweep and the regrouping pass do because they run for minutes and would otherwise be unbounded. This is seconds of small, independent pieces, so slicing answers the same question more cheaply. `pending` is in rail order, and the rail is sorted by confirmed faces, so the portraits the user is looking at are cut first. It patches single rows rather than reloading: a reload would rebuild the model on every tick, and a model replaced underneath the `ListView` is what the slicing exists to avoid. Each patch checks the row still holds the person it was started for — a stale index would draw a face beside somebody else's name — and stops the fill when it does not. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:30:16 +02:00
dtourolleandClaude Opus 5	d7b851f622	Build the People rail rows you can see, not the ones you cannot ... `Flickable { VerticalLayout { for ... } }` instantiates every row. On the reference library that is 14,268 row subtrees, which measures at 4.2 seconds before a pixel is drawn — and it was paid again on every action, because every action reloads the model. Roughly half the wait between clicking "Identity" and the screen appearing was this. Slint's compiler has a virtualising path for a `for`; it is what makes `std-widgets`' `ListView` cheap. It keys on the parent element's base being *named* `ListView` and exposing the five lengths its layouting code writes back, and a custom base is explicitly allowed. So `widgets.slint` grows one, and `std-widgets` stays out of the file that establishes our style. Measured on the real screen: 200,000 rail rows now render in 62ms. Three things had to be true together, and each was silent on its own. **The row height must be constant.** The rail hid a set-aside person with `height: cond ? 52px : 0px`, a height that reads the model — so the layout cannot place row N without building rows 0..N, and Slint builds them all. The filtering moves to Rust, where the toggle was already reloading anyway. **The list must not be wrapped.** It carries its own stretch and preferred size, having no natural height to offer; a Rectangle in between hands the layout that Rectangle's constraints, which are taken from the list and are therefore nothing. **The panel must let it fill.** `Panel` lays its children out with `alignment: start`, which gives each its preferred height — right for a column of sliders, wrong for anything that scrolls. Hence `Panel.fill`. Get any of them wrong and the rail renders empty, with no error and a model full of people. All three were, in turn, before a headless render of the real screen showed a blank rail. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:30:16 +02:00
dtourolleandClaude Opus 5	f26f1ab694	Withhold the empty groups that fill the People rail ... `load_people` returned every live person, and on the reference library that was 14,268 rows of which 11,739 held no faces at all — 85 per cent of the rail naming nobody and able to do nothing. They are not a mystery. A regrouping pass creates a person per cluster; the next pass moves those faces elsewhere and leaves the person it emptied behind. `faces::prune_empty_unnamed` exists for exactly this and runs only at the end of a pass, so nothing clears what accumulates between them, and the Identity screen never prunes at all. Each one cost a `for_person` query and a built row on every reload. This withholds precisely the set the prune already treats as disposable — empty, unnamed, not set aside — and no more. Filtered rather than deleted: a screen is being drawn, not a catalog repaired. Nothing is lost, a sync cannot resurrect what was never removed, and the prune stays the one place that decides these can go. An empty group with a *name* still shows. That one is not debris but the symptom of a real failure — a named person whose faces were regrouped out from under them — and hiding it would take away the only way to merge them back. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 18:30:16 +02:00
dtourolleandClaude Opus 5	94b1b9e0cc	Reconcile what four branches each built separately ... Three seams, found by the first compile after the merge. Two branches implemented 'can this scope be reordered' independently: one from the collections model, one from the catalog through orders_manually, on every re-read and excluding the trash. The second is the better answer and is what survives; it only needed to set the property app.slint declares. Two lints from scene-mask-ui, which was merged mid-flight and had never been through -D warnings: an is_none check spelled out where clippy wants ?, and a return in a cfg block's tail. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 16:51:41 +02:00
dtourolle	19b56ad85c	Merge: collection ordering, and a range that says where it ends ... Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> # Conflicts: # docs/traceability.md # ui/dr-ui/src/collections_ui.rs # ui/dr-ui/src/library.rs # ui/dr-ui/ui/app.slint # ui/dr-ui/ui/library.slint # ui/dr-ui/ui/widgets.slint	2026-08-30 16:38:08 +02:00
dtourolle	328fda6f7c	Merge: mask a whole category, not just one instance ... Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> # Conflicts: # apps/darkroom-desktop/Cargo.toml # docs/traceability.md	2026-08-30 16:37:31 +02:00
dtourolleandClaude Opus 5	86260b5028	Bind a category layer to its own distance field ... A category mask showed nothing and its adjustment covered the whole photograph. Both from one line: the loop in `MaskPass::rasterise` picks a distance field by matching `layer.source`, that match named only `Subject`, and a `Category` layer fell through to `_ => (&self.empty_subject, 0)` — a 1x1 placeholder. No field, so nothing to draw and nothing to confine the adjustment. The comment three lines above the arm I missed describes the failure I then shipped: an absent mask that defaults to "everything" would apply the adjustment to the whole photograph There are *two* matches on `layer.source` in that loop — one choosing the field, one building the params. Adding the category to the second and not the first compiles, runs, and is wrong in exactly the way the first one warns about. ## Also: a missing mask must still be the right size Both model-backed arms of `ensure_subject_fields` used `unwrap_or_default`, which yields an empty `Vec` when the coverage is gone. `SubjectMasks::upload` rejects a wrong-sized field and fails the whole batch, so `self.subjects` becomes `None` and *every* layer in the stack loses its mask — one stale reference silently unmasking the others. Pre-existing, and it mattered less when the only model-backed source was a subject: an instance index goes missing rarely. A category name goes missing whenever the descriptor is edited, which is a thing the descriptor exists to allow. A full-size empty field costs one layer instead of all of them. Neither of these is reachable from a test on this machine — both live past a GPU adapter and a real segmentation — so they surfaced the only way they could, by someone opening the app and looking. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 14:28:27 +02:00
dtourolle	80f0a0eeac	Merge branch 'master' into feat/library-toolbar ... # Conflicts: # docs/gestures.md # docs/traceability.md	2026-08-30 14:18:21 +02:00
dtourolleandClaude Opus 5	677146883f	Decide what the header gives up when it cannot hold everything ... Three faults, all invisible in the source and all found by looking at the window. `alignment: start` on the selection bar. Slint only hands space to `horizontal-stretch` children under the default `stretch` alignment; under `start` every child takes its preferred width and the stretch is ignored without complaint. That was harmless while the bar held four controls. Now that it holds every selection verb it is the difference between a count that gives way and a row that can only scroll, so the alignment goes and the stretch does its job. The title collapsed to "…". An eliding Text has a minimum of nothing, so once the buttons had taken their own minimums the title was the cheapest thing in the row to give away — the window stopped saying which library was open while the byte counts beside it stayed. It gets a 90px floor. And the status line does not get one. After the sidebar takes its 232, a 1100pt window leaves this header about 868, and the buttons want most of that before a character is drawn — so something must degrade, and the order is the whole question. A floor here bought a readable status by pushing Settings off the right edge, reachable only by knowing to flick-scroll, which is precisely the fault the row's own Flickable comment warns about. A control you cannot see is worse than a sentence you cannot finish. The status line is also the most redundant thing in the header — the sidebar states the library's count and the filter chips state it again — so it is what gives, and it grows back the moment there is room. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 14:05:24 +02:00
dtourolleandClaude Opus 5	9994bb4ce7	Regenerate the matrix over the third wave ... Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 14:02:40 +02:00
dtourolleandClaude Opus 5	7f350e8c4c	Offer the categories in the masks panel ... Everything before this was reachable only from an example that writes PPMs. This is the part a photographer can touch: a list under the subjects, click one, get a mask layer for every pixel of that category. ## Under the subjects, and the order is the argument Clicking the photograph is how a local adjustment usually starts, so the things the model *found* come first. A category is the move you reach for deliberately — grade the sky, not this one bird — and putting it second says so without a word of explanation. Coverage is shown for the same reason a subject's score is: it tells the photographer whether a category is worth a click before they spend one finding out. ## Not a new tab, which is what was asked for A develop tab is derived from `Attribute`, not declared — the tabs exist because operations claim an attribute, and no amount of Slint adds one. A seventh attribute would have meant duplicating every adjustment once per category, and the combinatorics get silly by the third. Reached through masks instead, a category composes with every adjustment that already exists, and inherits feather and falloff rather than needing its own. What was described as "per-category sliders with feather and decay" is exactly what this is; only the door is different. ## Verified as far as it can be here Compiles, populates, round-trips, 560 dr-ui tests green. **Not clicked** — synthetic input is blocked on this setup, so how it looks and feels is unverified and wants a human at the window. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 14:02:34 +02:00
dtourolleandClaude Opus 5	763dfd353a	Weigh the categories in the same precompute, and mask with them ... The scene model shipped with a decoder and no caller. This runs it. ## Beside the instance pass, not instead of it `compute` now does both on the same upright frame and lays both back down the same way, so instance masks and category masks index into one grid — the sensor's. A failure in the scene half is logged and dropped rather than propagated: no scene model is an ordinary state, and a photograph that can still be masked by subject should not become unopenable because the categories are missing. Categories under half a percent of the frame never reach the cache. A control that does nothing when moved is worse than an absent one, and each one it skips is a proxy-sized buffer not allocated. ## The shader needed nothing A category reaches `dr-gpu` as a soft coverage buffer at proxy resolution, turned into a distance field — which is exactly what a subject is. So they share `MODE_SUBJECT`. That is not a shortcut taken for speed: the shader has no way to tell them apart and no reason to want one. What differs is only which model produced the coverage, and that has already happened by then. Feather, falloff, dilation and erosion therefore work on a category on the day it arrives, because they were never subject-specific. ## Where the weights come from `scene-model` compiles the graph in and the desktop app takes it; Android leaves it off and reads the copy `install_bundled_models` unpacks, because 24 MB of constant is worth avoiding in a mobile install and not worth the plumbing to avoid on a desktop one. Embedded is tried first — a build that has the weights compiled in should not be silently overridden by a stale file in a data directory. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 14:02:22 +02:00
dtourolleandClaude Opus 5	d70dcf78d1	Merge: recover a damaged catalog, and capture a crash locally ... Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-30 13:45:18 +02:00
dtourolle	c48896bd95	Merge: touch selection and drag, from the gallery-selection branch ... Verified before merge: fmt clean, clippy -D warnings clean, 563 dr-ui tests. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> # Conflicts: # docs/traceability.md	2026-08-30 13:45:18 +02:00