DarkRoom

Author	SHA1	Message	Date
dtourolle	4a04496c78	Merge: focus peaking, so a frame can be judged without zooming to 100% FR-CULL-3's peaking half. The raw histogram and raw clipping indicators remain unbuilt -- what exists is a display histogram tagged FR-DSP-7, counting AdjustPass's 8-bit output, which reports a highlight as gone precisely where FR-CULL-3 needs it to report the highlight recoverable. Verified before merge: fmt clean, clippy --workspace --all-targets -D warnings green, 11 focus GPU tests, 79 baseline dr-gpu tests, 511 dr-ui tests. The cfg(target_os = "android") arm is unverified -- the host-target clippy never compiled it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> # Conflicts: # ui/dr-ui/src/lib.rs # ui/dr-ui/ui/app.slint	2026-08-29 21:43:35 +02:00
dtourolleandClaude Opus 5	2168cdd1c4	Mark what is in focus, so a frame can be judged without zooming to 100% FR-CULL-3's focus peaking. One compute dispatch measures local contrast in WGSL and writes an overlay texture; on desktop it reaches Slint through the same zero-copy wgpu import the canvas uses, so nothing per-pixel touches the CPU on the frame path. With peaking off the cost is zero and structurally so: focus_overlay opens with `let settings = self.peaking?;` before the frame is touched, and clearing drops both overlay textures, so no VRAM is held either. NFR-P14 is met by construction rather than by measurement -- one dispatch, no second render, no pipeline compile after session open, and a test asserting allocations stay at 2 over eight frames. The budget test asserts 50ms at 4K rather than a tight bound, deliberately: a tight bound fails on a loaded machine and gets deleted, which is worse than a loose one that still catches the regression that matters. TD-1 is amended rather than joined by a TD-6: on Android the overlay rides the readback that already exists there, roughly doubling that transfer while peaking is on, and TD-1's own "Done when" removes both because both are the same missing capability. Verified: cargo fmt clean; clippy --workspace --all-targets -D warnings green, which also compiles peaking.slint through dr-ui's build.rs; 11 focus GPU tests and 79 baseline dr-gpu tests pass; 511 dr-ui tests pass. Not verified: the cfg(target_os = "android") arm, which the host-target clippy never compiled. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 21:43:08 +02:00
dtourolleandClaude Opus 5	7ad75dd905	Let a manual collection be put in the order the photographer wants `collections::set_order` and `Sort::CollectionPosition` have been in the catalog since collections were, and nothing above dr-catalog has ever called either. A manual order existed, could not be seen, and could not be set. This is the half that was missing. Three pieces, because it needed all three to be visible at all. The catalog gains `orders_manually` and `members_in_order`. The first is the rule about when a manual order means anything, kept in one place with one name: a collection must be manual, and it must have no children. A set shows its descendants' images, and positions are only ever assigned within one collection — so two children's positions are unrelated integers, and ordering by them would sort the grid by a coincidence. The existing comment on `read_cells_scoped` already argued this; now something enforces it. `members_in_order` returns the whole membership rather than the filtered view, because `set_order` renumbers exactly what it is handed. Reordering a filtered list would renumber those and leave every hidden image on a stale position — two images sharing one, and a grid that rearranges itself the moment the filter comes off. The grid reads position where the scope qualifies and capture time everywhere else. Manual order joins the member row rather than testing membership with `IN`, which is safe from fanning out rows because that branch is a single collection. The gesture is a DropArea over the viewport, drawn only where a reorder means something, with a caret in the gap the photographs would go into — a line between two images rather than a highlight on one, because lighting up a cell would say the drop replaces it. The trap worth naming: `DropEvent.position` is in window coordinates. Slint maps it through `map_to_window` when the drag begins and hands every target the same event untranslated, so a target inside a Flickable has to subtract its own `absolute-position`. Getting that wrong is invisible until the grid is scrolled, because at the top the two frames coincide. `reordered` is pure and names its destination by the image it goes before rather than by an index, because the grid can only name a gap in what it is showing and the ids are what survive a window swap. A drop that changes nothing returns the order untouched: that counter is what a cross-device merge resolves by, and spending a revision on a no-op makes this device win an argument it did not have. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-29 20:49:00 +02:00
dtourolleandClaude Opus 5	bb35665bd2	Let a paste carry some kinds of edit and not others FR-DEV-6 asks for presets "covering a subset of the edit graph". What landed with the named presets covered two subsets: everything, and everything but the crop. "Match the colour but not the sharpening" had no way to be said. `Scope` is now a set of `Attribute` — the same six kinds every operation already declares and the develop panel already builds its tabs from. The photographer ticking "tone and colour" is naming the groups they navigate by, and neither this module nor the interface has to name an operation to do it (FR-DEV-3c). The pleasing part is what left. Framing used to be excluded by an explicit test against one operation's id; it is now excluded because Geometry is not in the default set. The special case dissolved into the general rule, and the argument for it — a crop is a decision about this photograph, and carrying it across forty destroys forty compositions — is now a statement about a kind of edit rather than about a node. All thirty-three existing preset tests pass unchanged, which is the evidence that the generalisation kept its promises. One decision that is a field rather than a rule, because the two cases genuinely differ. An operation this build cannot classify — from a newer version, arriving over sync — travels under "everything" and "everything but the crop", because those are claims about the whole edit and an unrecognised operation is part of it (FR-NC-8). It does not travel under a hand-picked set, because that is a claim about kinds, and an unknown kind is not one of the kinds that were ticked. The settings page's "Copy crop and rotation" checkbox is gone, replaced by the same chips the preset sheet draws. It asked the right first question — geometry is the kind whose accidental travel destroys work — but it was the only question a boolean could ask. The field stays in `Settings`, read exactly once to seed the new set, so anyone who had ticked it keeps their behaviour. The chips are deliberately not in the develop column. Six of them there would set the width of the whole sidebar, which is the bug `ChipGrid`'s comment records at length. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-29 20:38:19 +02:00
dtourolleandClaude Opus 5	d3dadbd725	Group the frames of one moment, by when they were taken and what they look like A burst is the commonest thing in a cull and the least interesting: twelve frames of the same gull at 10 fps occupy twelve cells, are scrolled past twelve times, and end with the photographer keeping one. FR-CULL-5 asks for them to collapse to one representative and be judged as a unit. Two signals, because neither alone survives a real library. Time alone groups a whole wedding ceremony -- a photographer working steadily never leaves the gap that would end the run. Similarity alone groups a studio setup shot across two days, which is a project rather than a moment. Together they are specific: adjacent in time and looks like the frame before it. Two seconds is the time bound, and the reason is worth recording because the figure looks absurd next to a 10 fps camera. `images.captured_at` is whole seconds -- EXIF's DateTimeOriginal has no sub-second field and SubSecTimeOriginal is optional and widely omitted -- so a burst arrives in the catalog as ten frames sharing one timestamp. Any threshold finer than a second is a threshold on information that is not there. Where the pace really is faster, the similarity bound is what separates the frames. Similarity is a 64-bit difference hash over a 9x8 box-averaged reduction, compared between adjacent frames only. Chained rather than anchored on the first frame, because by frame twenty a camera following a bird has nothing in common with frame one while no two neighbours differ by much; the time bound is what stops the chain running away. There is no all-pairs step and there must never be one -- that is what turns a grouping pass into something nobody can afford to run over 50k images. Nothing here ranks a frame. FR-CULL-5 names the failure it is avoiding, which is rejecting the only frame of an important moment because somebody blinked, so there is no sharpness score and no best-of-burst. The representative is the earliest frame -- a fact about the clock, not a judgement about the photograph -- and the user's own choice lives in its own table so that rebuilding the grouping cannot erase it. Same argument `people.ignored` makes one subsystem over: nothing short of remembering a decision survives re-clustering. A newly found burst is recorded open. Collapsing on discovery would be tidier, and would also mean a background pass taking photographs off the screen part way through a cull. The pass marks; the user folds. It is a pass rather than a job kind for the reason catalog.md 10.2 gives for face clustering: a burst is a property of a run of frames and has no natural subject_id, so a per-image job would rebuild the world once per photograph. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 20:37:54 +02:00
dtourolleandClaude Opus 5	ab0ef6a26d	Say which requirements the code was already satisfying Thirteen requirements were surveyed as built but untagged. Eight of them were: R3, R6, FR-DEV-1, FR-UI-6, FR-NC-6d, NFR-OPS-3, NFR-PORT-2 and NFR-SEC-3. Each was read against its full text in requirements.md and against the code before the tag was added, because a tag that is wrong is worse than an absent one — it turns a visible gap into an invisible one. The five that were refused, and why, because the reasoning is the part worth keeping: R2 carries "(figure TBD)" in its own acceptance criterion and asks for a stated prefetch margin and cache-hit rate; neither figure exists anywhere in the tree and neither quantity is measured, while TD-2 and TD-3 both describe the thumbnail path falling short of it. R5 asks for three things and the code does one. The display pipeline does run at viewport resolution, but "only visible tiles are computed" and "panning recomputes only newly exposed tiles" need a tile scheduler that does not exist — and frame_budget.rs currently argues for striking tiled computation from the interactive path rather than building it. FR-RAW-2 asks for a trait taking a SourceRef, so that a second decoder can be added without changing callers. What exists is free functions over &[u8]. That meets the requirement's stated purpose — the same decoder serves a local file, a SAF document and a byte range, which is exactly why it takes bytes — but there is no trait and no second implementation seam, so the requirement should probably be amended rather than tagged. NFR-ARCH-1 asks for named executors with stated thread counts. architecture.md §7.1 states the table; nothing implements it. Workers are twenty-odd ad-hoc std::thread::spawn sites, each building its own one-worker tokio runtime, with no decode pool, no GPU-submit executor and no I/O pool. The requirement's own text says R4 and NFR-P9 "assert an outcome with no stated means", and that is still true. NFR-SEC-4 is satisfied by absence — there is no telemetry — and absence has no module to tag. A tag would point at nothing. NFR-OPS-3 was the closest call of the eight taken. The store is single, separate from the catalog, survives a catalog rebuild and does not sync between devices; it has no version field, deliberately, and settings.rs argues why and names the condition that would need one. The substance is met and the reasoning is recorded where it belongs. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 20:23:16 +02:00
dtourolleandClaude Opus 5	5a8327824f	Keep an edit under a name, not just on the clipboard FR-DEV-6 asks for three things — named presets, copy/paste between images, and batch-apply to a selection. The last two have been here for a while; this is the first. The format is the sidecar's, deliberately. A preset is the non-default half of a version, so the lines are the same lines keyed the same way, which makes the two files diffable against each other and lets someone debugging an edit paste a block from one into the other. One file rather than one per preset: a preset per file makes the name a path, and every name then has to survive a filesystem — a `/` becomes a directory, a name differing only in case collides on one platform and not another, and renaming becomes two operations that can half-fail. As a key in a document it is none of those. Unknown parameters needed no machinery. `Preset` already holds whatever keys it is given and resolves them against the descriptors only at apply time, so one written by a newer build survives by being stored. Only lines that are not `op.param = float` at all are preserved verbatim, which is the sidecar's version-skew promise made here too. Applying is the paste path with a different source, so a preset reaches a selection through the sidecar read-modify-write that was already there: no graph, no decode, no GPU, forty files or one. Two smaller decisions worth the record. A library that fails to parse is held empty in memory and not written back over — settings regenerate themselves and this is work, so a parse failure must not be the moment it is destroyed. And every save persists immediately and rolls the in-memory copy back if the write fails, so the sheet never lists a preset the file does not have. The grid's "Presets" button is gated on the selection alone, unlike the "Paste to 40" beside it. That button needs a clipboard armed this session; the preset list is whatever was saved last month, and hiding it behind an unrelated action is what makes a feature only its author knows about. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-29 20:07:14 +02:00
dtourolleandClaude Opus 5	574107bc39	Let a manual collection be put in the order it is meant to be seen in `collection_members.position` and `Sort::CollectionPosition` have been in the catalog since collections were, and nothing above dr-catalog has ever written or read either: `collections::set_order` had no callers, and the grid ordered everything by capture time whatever it was scoped to — dr-ui does not construct a `Query` at all, it has its own `GRID_ORDER` constant. So a manual collection was a set with an order nobody could see or change. Three pieces, because it could not be fewer: `grid_order_for` decides the ordering from the scope, and both readers take it from there. That is the load-bearing part. An ordinal only names a photograph relative to an ordering, so the window read and the span read have to agree — a shift-click resolved through a different ORDER BY than the cells were drawn with selects a different run than the one on screen, and the user finds out when the export runs. `read_ids_span` already stated that invariant about `GRID_ORDER`; this widens it to an ordering that depends on the scope. Only a single manual collection has one. A set draws its descendants' images too, and two children's positions are unrelated integers that interleave arbitrarily; a smart collection has no member rows to carry a position at all. Both fall back to capture time and refuse the drop rather than pretending. The drop is on the cell, on whichever half of it the finger landed — the trailing edge is the only way to name the last place in a collection, since there is no cell beyond the last one to drop in front of. `reordered` is pure and the membership is rewritten whole. `set_order` sets the positions it is given and leaves the rest, so a partial write would interleave the moved run with rows nobody touched; and it is read unfiltered, so what the filter is hiding keeps its place relative to what the user can see. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 20:04:17 +02:00
dtourolleandClaude Opus 5	5133e53bc8	Give back what a straighten took, when the angle comes back Build and test / Desktop (Linux) (push) Successful in 2h16m6s Details Build and test / Layer separation (push) Successful in 52s Details 🐳 Android image / Build and push (push) Successful in 4s Details Build and test / android-image (push) Successful in 4s Details Traceability / Requirement traces (push) Successful in 51s Details Build and test / Android (aarch64) (push) Successful in 47m10s Details The auto-crop only ever shrank. Straighten to 20 degrees and the corners are cropped away correctly; come back to 3, or all the way to zero, and the crop stays at the size 20 degrees demanded. Nothing on screen explains why the photograph is still small, and the only way back was undo. The cause was that each correction was computed from the previous correction's output, so it accumulated: every angle the slider rested at took its cut and none was ever returned. The fix is to stop accumulating and recompute. The applied crop is now always the user's own rectangle fitted into the current angle's safe area, so as the angle falls and that area opens up the crop grows back — and stops, exactly, at the rectangle they chose. At zero the safe area is the whole frame and the fit is the identity, which is what carries it the last of the way home. There is deliberately no early exit for the upright case now: that exit is precisely what would strand the crop small. The intent is remembered as a pair, so it repairs itself. The session keeps `(applied, intended)` — what the correction wrote, and what it was derived from — and trusts the remembered intent only while the graph still holds `applied`. Every other route to the crop leaves something else there: a handle dragged, a ratio chosen, a sidecar loaded, a paste, an undo. That mismatch is the signal the memory is stale, and the current rectangle becomes the new intent. The alternative was a write into this field from each of those paths, which is the kind of bookkeeping that is correct until someone adds a seventh path. Dragging a handle therefore is the user choosing, including at a non-zero angle: the correction will not later grow the crop past what they dragged it to. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 19:16:43 +02:00
dtourolleandClaude Opus 5	3bb68cff69	Export at an exact resolution, and offer the panels worth naming Export sizing could bound an image but not fix it. Long edge, short edge and percentage all preserve the aspect ratio by letting one dimension fall where it may, which is right for most work and useless against a display that accepts one resolution and rejects everything else — a television's art mode, a digital frame, a wallpaper slot. FR-EXP-3 has always listed both halves of the answer, and this adds them. Fit box scales to fit inside a width and height, so nothing is thrown away and the result is smaller than the box on one axis unless the crop already matches it. Fill box scales to cover the box and cuts the overhang off the middle, so the file is exactly the pixels asked for. Fill is the only mode in the file that discards image data, so two things about it are worth stating. The overhang comes off symmetrically: the crop tool is where a photographer decides which part of a frame survives, and this stage having an opinion of its own would fight it. And locking the crop to the same ratio leaves nothing here to cut, which is the workflow the two features are meant to be used in. With upscaling off and a source too small to cover, a fill box keeps its shape rather than falling back to the source's: exporting a 3:2 file where 16:9 was asked for is silently wrong in exactly the way the mode exists to prevent, so the box shrinks instead. The existing rule — clamp, never fail — is otherwise unchanged. Four panel sizes are offered as buttons beside the fields. Getting 3840 x 2160 by typing four digits twice is a step at which the mistake is discovered after the upload rather than before it. They fill in the numbers and nothing else, in particular not the fit/fill choice: both are legitimate against a screen, and guessing would discard the edges of a photograph for a user who wanted them. The list is panels rather than platforms, because a screen has one exact pixel count for ever where "what a photo site wants" would rot in the file. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 13:49:20 +02:00
dtourolleandClaude Opus 5	d9eb8faffd	Crop away the corners a straighten exposed, once the slider is let go Turning a rectangle inside its own bounds exposes its corners: there is no source pixel out there, and the shader renders it black. Nothing in the render prevents that, deliberately — a free angle does not change the output size, which is what leaves the frame where the user put it while the slider moves. Correct during the drag; four black wedges on the finished photograph. Letting go of the slider now pulls the crop inside the area the angle leaves defined. `Framing::max_inscribed_crop` already computed that bound and had no caller; this is the caller its doc comment described. Once, at the end of the gesture. Applied per frame it would shrink the crop on every step of the slider and never grow it back, so a user who overshot to 20° and came back to 3° would be left with a crop ratcheted down by the excursion rather than by the angle they settled on. Per gesture it is bounded by the angles actually rested at, and undo steps back through them. The crop is fitted into the bound, not replaced by it. A crop placed deliberately off-centre is a decision, and an automatic correction that recentred it would undo the user's work to fix a problem they did not have. `CropRect::fitted_into` scales only as far as the bound demands and then slides the rect the shortest distance needed to be inside — so a ratio locked in the crop panel survives the straighten too, since the shape is never touched. It returns the rect unchanged, bit for bit, when nothing needed to move. That matters more than it looks: this runs on every release of the slider, including releases at zero, and a rect that drifted by a rounding error each time would be an edit recorded for no reason. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 13:49:19 +02:00
dtourolleandClaude Opus 5	e6ad906bc1	Let the crop be held to a ratio while it is dragged A photographer cropping for a print, a phone wallpaper or a 16:9 frame is not choosing four edges — they are choosing one edge and a known shape. Free-dragging every corner made them do that arithmetic by eye on every drag, and get it slightly wrong. The panel now offers Free, Original, 1:1, 3:2, 4:3 and 16:9, with a Portrait switch for the ones that have two orientations. Original follows the frame rather than naming a number, so it stays right on the next photograph from another body and after a quarter turn. The ratio is of output pixels, and the rect is not. `CropRect` is stored in fractions of a frame that is not itself square, so holding a shape needs the frame's size — `ratio * height / width` of the frame. Skipping that gives a "1:1" crop that is square only on a square photograph, which is the one case nobody would test on, so the conversion lives in `CropRect::with_aspect` where it is explained and pinned by a test that asserts the fractions are not equal. Two decisions worth recording: The reshaped rect grows onto the ratio rather than shrinking onto it, then scales down only as far as the frame's edge demands. Fitting inside instead makes a one-axis drag do nothing at all — the other axis clamps the first straight back, and the handle simply refuses to move. The overlay now reports which corner the drag is holding, because reshaping onto a ratio has to know which corner is nailed down and only the handle that took the press knows that. A move reports no corner and keeps its shape: reshaping about a centre would pull an over-moved rect smaller instead of sliding it along the edge. The lock lives with the window rather than the session. A `DevelopSession` is per image, and cropping a set of frames to one shape is exactly when the lock earns its place. It is not an edit and reaches no sidecar — what is saved is the rectangle it produced. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 13:48:51 +02:00
dtourolle	30162e80df	Merge branch 'master' into clarity-reduced-base # Conflicts: # docs/traceability.md	2026-08-29 13:27:04 +02:00
dtourolleandClaude Opus 5	bff95e25ad	Let clarity's base be computed where it is still fully determined Clarity's Gaussian sigma is 1.2% of the frame's shorter edge, so its radius is a property of the viewport: 52 render pixels at 4K, two separable passes of 105 taps each over 8.3 M pixels. That measured 33.9 ms — seven times the entire fused point chain, for one slider — and is docs/technical-debt.md TD-4. A detail pass may now declare `output_scale`, and clarity's base is computed on a grid a quarter the size on each axis. The pass that combines needs the blur and the full-resolution colour, and a colour that has been through a quarter-scale target is no longer full resolution. So a scaled pass cannot simply join the ping-pong: there are two chains now. The full-resolution one carries the colour and no scaled pass touches it; the reduced one carries the base and reaches the combining pass through a second binding as `reduced_at()`. The reduce is a dispatch of its own rather than something the first blur half does on the way past, and that is the whole difference between this and the strided kernel the module documentation rules out. A stride samples an image that is not band-limited and aliases high-frequency content down into the base, which is then subtracted, and arrives in the output as mottling across smooth gradients. This band-limits first and samples after. What is discarded is content the base could not represent at any resolution, because a Gaussian at sigma = 26 px holds nothing above one cycle per 26 px and the quarter-scale grid carries one per 8 — so the reduced base is not an approximation of the full-resolution one, it is the same function sampled where it is still determined. Which is also why the scale belongs to the band rather than to the stage. Texture's sigma is a decade finer, so the reduce pass's own box would be wider than the Gaussian it was prefiltering; texture never reduces. And clarity steps 4 -> 2 -> 1 as sigma falls, because a quarter of a small sigma is not a Gaussian either — the case that gives up is the one that was already cheap. `radius` stays in each pass's own pixels and `ComposedDetail::radius` multiplies it back up, so 13 reduced pixels at scale 4 still report the 52 render pixels a tile would have to be grown by. The halo a scheduler sees does not move. The halo tests pass unchanged, which was TD-4's stated bar; they render at 1024 px and so exercise the reduced path rather than stepping around it. Added `crossing_the_reduction_threshold_does_not_change_the_picture`, because nothing yet compared the reduced form against a less reduced one — every other test measures one form against itself. It renders the same edit either side of the 4 -> 2 step-down and holds the peak excursion to 0.03 stops and the reach to 2% of the frame. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 13:12:29 +02:00
dtourolle	3b5d564495	Try every GPU, not only the fastest one Build and test / Desktop (Linux) (push) Successful in 2h7m41s Details Build and test / Layer separation (push) Successful in 46s Details 🐳 Android image / Build and push (push) Successful in 3s Details Build and test / android-image (push) Successful in 3s Details Traceability / Requirement traces (push) Successful in 41s Details Build and test / Android (aarch64) (push) Successful in 21m1s Details `request_adapter` with `HighPerformance` returns one adapter and no second chance. That is right on a healthy machine and wrong on one with a sick GPU, which is not rare: observed 2026-08-29 on a laptop whose discrete card had hit an NVRM assertion failure and a fullchip reset. The driver still advertised it, wgpu dutifully picked it as the highest performing, and the process died on it — while a working integrated GPU and a working external card sat unused in the same enumeration. A photo editor that will not start because the fastest GPU is broken, on a machine holding two that are not, is worse than a slow one. So: enumerate, order by preference, take the first that yields a device. The ordering reproduces what `HighPerformance` meant, so a healthy machine picks what it always picked and pays one enumeration for it. A CPU adapter sorts last rather than being excluded — software rendering is a poor experience and a working one. Which GPU to prefer is now a policy rather than an assumption, because the fastest is not obviously the right one. A 24 MP frame is ~96 MB of RGBA and every upload and export readback crosses PCIe on a discrete card, where an integrated GPU shares memory and crosses nothing — and does not empty a battery. Measured before choosing a default, on this machine's Iris Xe against its RX 5700 XT. The fused colour pass is within 1.5x, which is the shape shared memory suits. The neighbourhood stage is 5-8x slower, and that decides it: clarity at 1920x1200 costs 20 ms on the iGPU, over the budget on its own at the smallest size tested. So `Performance` stays the default and `Efficiency` is offered rather than chosen (`DARKROOM_GPU=integrated`). docs/frame-budget.md carries the table, and says what it does not show: the harness renders from a resident texture and never uploads or reads back, so the transfer cost an iGPU avoids appears in none of it. Import, export and the thumbnail sweeps may well go the other way. What this cannot fix: a GPU sick enough to accept `request_device` and segfault afterwards, which arrives as a driver crash rather than an error. It moves the boundary from "the preferred adapter is unusable" to "unusable and dishonest about it".	2026-08-29 12:43:10 +02:00
dtourolleandClaude Opus 5	0407fb8d2d	Format the two new examples They were written after the last fmt run and CI gates on --check. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 12:33:23 +02:00
dtourolleandClaude Opus 5	39c34d4e44	Say what actually counts as a rival, now that a name anchors too assign's denominator is the identities the user has ruled on, and it reads Cluster::person to find them. Master's "Let a name hold a group together" widened what sets that field: a confirmation, a name, or an ignore, where before it was a confirmation alone. The behaviour is right either way — a named person is exactly the identity a suggestion should be discounted against — but the module note and faces.md §9.1 both said "a confirmation", which is now too narrow. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 12:32:31 +02:00
dtourolleandClaude Opus 5	da20d42d33	Merge master: pluggable storage, and a name that anchors Conflicts were docs/traceability.md alone, and it is generated — so it was regenerated rather than hand-merged. dr-face was untouched on the other side; ui/dr-ui/src/faces.rs and identity_ui.rs auto-merged, the first around recluster's anchoring and the second around load_faces. Worth recording because the two branches met on the same problem from different ends. Master's "Let a name hold a group together" is the fix for the sixteen Catherines — fourteen of them empty — that this branch found while measuring the library and reported without fixing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 12:30:01 +02:00
dtourolleandClaude Opus 5	b2250cc460	Measure a regroup on the tablet, not just on the desktop The GPU question needed a number nobody had: how a regroup divides on the hardware whose CPU is weakest. dr-face carries no weights and touches no display, and dr-catalog's example needs only a catalog file, so both run under adb shell against a copy of a real library. On the same 18,143 faces — desktop against the tablet — scan 0.96s / 2.61s, agglomerate 1.69s / 2.16s, score 0.26s / 0.40s. The scan is half the pass on the tablet and under a third on the desktop, because twenty cores of AVX2 pull ahead of NEON much further than the merge engine's single-threaded hashing does. So a GPU GEMM is worth roughly 2× a regroup on the tablet and 1.5× here, and it is the tablet that should decide whether it is built. The two architectures agree exactly: the same 1,531,969 evidence pairs, the same 2,518 groups holding the same 16,246 faces, the same reliability table. That is a better check on the NEON kernel than the unit test can be. Two instruments, both read-only: the example now prints its phases, and dr-face gains scan_bench, which needs no library at all and so can answer "how fast is this machine" on a device with nothing on it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 12:16:48 +02:00
dtourolleandClaude Opus 5	a67402961d	Let the merge engine's dot product use the machine's kernel too Engine::cross is the one place a dot product is computed during agglomeration — when two groups become adjacent through a third and their sub-threshold pairs, never summed because they were never interesting, have to be accounted for. It was calling the portable loop while the scan beside it had AVX2 or NEON, which on the reference library was 1,753,514 dot products taking 0.54s. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 11:58:07 +02:00
dtourolleandClaude Opus 5	b4d39ba33a	File each pair under its component once, not once per component Seeding the merge heaps was 3.06s of a 5.93s regroup on the reference 18,143-face library — more than half the pass, spent before a single merge was considered. Every component scanned the whole pair list looking for the pairs that were its own: 475 components against 804,499 pairs, 382 million set lookups to place 804,499 of them. A pair can only ever join two faces of one component, since that is what a component is, so the union-find that finds the components can file the pairs at the same time and hand each agglomeration the list it needs. The membership set inside agglomerate goes with it — it existed only to run that filter — and the heap can be sized up front now that the pair count is known. Ordering is preserved deliberately: pairs are filed in the order they arrive, which is the global (i, j) order, so the seeded heap breaks its ties exactly as before and the merge order is unchanged. Same 2,518 groups holding the same 16,246 faces on the reference library, at 3.5s rather than 5.9s. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 11:57:29 +02:00
dtourolleandClaude Opus 5	e596eb0657	Give the similarity scan the machine's SIMD, and its cache The scan is O(n²) dot products and nothing else, so its speed is the face subsystem's speed — and it was running at 0.7 flops per cycle. Two separate faults, both measured over the reference 18,143-face library on twenty cores. It walked the whole embedding array once per row, ~336 GB of traffic, where a column tile that fits in L2 is read once per tile of rows: 4.64s → 2.81s. And the workspace builds for baseline x86-64 — SSE2, no FMA — into which the portable loop was not being vectorised at all: 2.81s → 0.86s, 195 GFLOP/s. So the dot product is now chosen per machine. AVX2 + FMA where is_x86_feature_detected! finds it; NEON unconditionally on aarch64, since Advanced SIMD is in that baseline and every Android device the app builds for has it — with the explicit vfmaq, because LLVM will not fuse a multiply and an add without being told to. The portable loop stays as the definition the others are tested against, and the_fastest_kernel_agrees_with_the_portable_one is the only check the NEON path gets on a machine that is not aarch64. Faces::embeddings is one flat buffer rather than a Vec per face: the pointer chase defeated both the prefetcher and the tiling, and it is also the layout a GPU pass would want. Behaviour is unchanged and that is checked rather than asserted — the same 1,531,969 pairs from all three kernels, and on the real library the same 2,518 groups holding the same 16,246 faces with the same confidence distribution. A full regroup there goes from 10.0s to 5.9s; the rest is the agglomeration, which is a sequential heap walk and is where the next look should go. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 11:33:15 +02:00
dtourolleandClaude Opus 5	ebb7d3cf5c	Score a suggestion against the people the user has named The number beside a suggestion was the mean calibrated probability between the face and the rest of its group, which measures the wrong thing twice. It punishes coverage: a person with two hundred faces over fifteen years is meant to have members a given photograph is orthogonal to, so a correct suggestion onto a well-photographed person scored low for being well photographed. And it never asked who else the face might be — a face matching Anna at 0.95 and nobody else, and one matching Anna at 0.95 and her sister at 0.93, came out identical, when the second is the only one worth the user's attention. dr_face::assign answers both, and multiplies them: the mean of the best ten calibrated matches into the identity (the old mean, capped, which is what stops coverage counting against it), times that identity's share of the evidence against every named rival. Only named people compete, and per person rather than per group. Both halves of that had to be measured on a real 18,000-face library rather than reasoned about. Normalising across every group made the number useless — median suggestion 21%, four in five under half — because clustering leaves one person spread over many groups, so a face competed against itself; and keying rivals by group left Catherine competing with Catherine, median 39%. Per named person: median 99.5%. Rivals are gathered below the merge threshold, down to even odds: a named person matching at 0.6 will never be merged into but is exactly the competition to discount for. That would be a second similarity scan, the expensive half of regrouping a library, so cluster_scored scans once at the looser floor and hands the merge engine the subset at or above the threshold — pair for pair what it would have scanned for itself, held to that by a test. Leave-one-out over that library's 2,702 confirmations across 54 named people: 99.33% of faces placed on the right person against the old mean's 99.15%, and the number shown for the right person moves from a median of 90.4% to 99.3%. It errs low — 100% correct wherever it states 80% or more — which is the safe direction, and docs/faces.md §9.1 says plainly that the low bands are not calibrated. The example that measures it comes too: this is a claim about a library's numbers, and nobody should have to take it on faith. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 10:44:01 +02:00
dtourolle	4168d67cfa	Do not push a catalog over one we could not read `sync_catalog` is a read-modify-write over a file another device also writes: take theirs, merge, push the union. It was shaped if let Ok(bytes) = backend.get(&RemoteId::Path(target), None).await { which folds every failure into "there is no remote catalog" and carries straight on to the upload. On a placeholder library the snapshot in `.darkroom-derived/` is dehydrated like anything else, so the read failed every time and each sync pushed our catalog over theirs unmerged — taking the other device's collections and their members with it. The same shape as the sidecar bug, and the same fix: a read that fails for anything other than `NotFound` stops the upload and says why. An unreadable or unopenable snapshot stops it too — "will not parse" is not "is not there". This is what `NotFound` and `NotMaterialised` being separate errors is for: one means ours is the whole truth, the other means do not dare. Shard downloads go through the same fetch-on-demand read. They logged and skipped before, which on a library the client keeps dehydrated is every shard, every pass, and a peer's thumbnails and faces silently never arriving. And `put` over a placeholder now replaces it rather than refusing. Refusing was over-cautious of me: derived state lives inside the library folder, so a folder the client had dehydrated could never be written to again. An unconditional write replaces the whole file, so there is nothing in the stub to keep — content first, then the placeholder, since in a synced tree an absence is a deletion that propagates. `IfMatch` still refuses, because a stub's validator describes the stub; `IfAbsent` fails, because the file is there and only its content is not.	2026-08-29 10:36:26 +02:00
dtourolle	702d83c218	Measure the borrow, rather than asserting it bounds the disk The claim that hydration-as-a-borrow makes peak disk the working set rather than the library was so far an argument. `--example vfs_cycle` runs it: 100 photographs of 25 MB, 90 dehydrated, a pass over all of them through the real engine. Peak 275 MB — the resting set plus one photograph — against 2,500 MB had the pass simply fetched everything. Back to 250 MB afterwards, and all ten files the user already kept still there, which is the half of the contract that matters more. Recorded in docs/storage.md §6.3 and ARCH §9.0a, because a bound argued from a number nobody measured is one that gets quietly lost.	2026-08-29 09:57:53 +02:00
dtourolle	5768100816	Borrow the library to index it, and give it back The passes that need every photograph's bytes — thumbnails, face indexing — now borrow each one and release it at the end. On a placeholder library that is the difference between peak disk being the working set and being the whole library. Including on cancellation, which was nearly missed: the face sweep returns mid-loop when the user presses Stop, and without releasing there the disk is spent and nothing is delivered for it. `materialise` now answers whether it fetched the content. The pool used to work that out by listing a file's parent directory — one listing per file across a library — when the backend already had to `stat` it to decide whether to ask. One syscall instead of a directory walk, and it removes the bug class the tests found earlier: a file at the library root has no `parent()`, so every one of them read as already-downloaded. Pinning is the retention control, and it drives the model the catalog already had rather than a second one. `tier_desired` is what the user asked to keep hydrated, `pending_pins` is the resumable work list, and a pinned collection is never dehydrated for the same reason it was never evicted. It was in fact broken here before: `get` on a stub failed, and the pin worker logged "one unreadable file must not abandon the whole pin" and silently did nothing. Pinned originals on such a library are recorded with `path = NULL` (`Cache::record_in_place`) rather than copied under `originals/`. Two reasons, and the second is the important one. A copy would hold every pinned photograph twice, with the budget able to evict the half that was not costing the disk. And `release` deletes the file a row names — so a row that names none cannot delete anything, which puts the one catastrophic operation out of reach by construction rather than by remembering not to call it. Deleting a materialised file inside a synced tree removes the photograph from the server and every other device. Handing disk back is `spawn_dehydrate`, which asks the client. Two gaps written down rather than papered over (docs/storage.md §7): a hydrating pass cannot yet quote its cost, because a stub reports no size; and the two sweeps hold separate pools, so a library indexed for both fetches twice.	2026-08-29 09:57:53 +02:00
dtourolle	c102ba9df2	Treat a placeholder as the photograph, not as a one-byte file The folder connector was pointed at a Nextcloud VFS tree and got three things wrong, the first of which loses work. A dehydrated sidecar read as absent. `a.drsc` does not exist when the client has dehydrated it — only `a.drsc.nextcloud` does — so `get` missed, `.ok()` swallowed the `NotFound`, and the sidecar writer took that for "there is no sidecar yet" and wrote a fresh document over the existing one. Every edit another device had put there went with it. That function's own doc comment calls this the exact loss the format's unknown-key preservation exists to prevent. A stub was catalogued as a 1-byte image, and ARCH §9.0 measured this machine at 121,785 placeholders against 10,267 real files — so a folder library on a synced tree was ~92% broken rows. Identity changed on hydration, so downloading a photograph looked like a delete and an add, orphaning its thumbnail and its face rows. Entries now carry the photograph's own name and a `materialised` flag; `get` on a stub returns the new `RemoteError::NotMaterialised`, which is distinct from `NotFound` precisely because the sidecar writer must treat them differently — it fetches the sidecar and merges, or leaves the entry queued. Hydration is a borrow. `BorrowPool` records what was on disk before it asked, so `release_all` dehydrates only what a pass brought and leaves what the user already had. Reference counted: the thumbnail pass and the face pass meet on the same RAW, and without counting the first to finish dehydrates the file the second is reading. A borrow against a plain folder or a server does nothing, so a pass written for VFS runs everywhere. Releasing means asking the client to dehydrate and never deleting: a deletion inside a synced tree propagates to the server and removes the photograph from every device. Not a second backend — the capability is per connection, not per type, since the same folder hydrates only while the client runs. The convention arrives through a detector the registry supplies, so `dr-sync-folder` still knows nothing about any client's protocol. ARCH §9.0a records this as an amendment: finding 3 rejected hydration because it costs 100× a range read, and that comparison assumed a connector was available. A folder library has none.	2026-08-29 09:57:52 +02:00
dtourolle	cbe5c4fcde	Measure the folder walk against a real tree, not a claim The folder connector declares `LocalEtags`, which means the engine walks the whole library on every scan with no pruning. That is the honest capability, and the argument for it being affordable was so far an assertion about `stat` versus `PROPFIND`. `--example scan` runs the real path — `dr_sync::scan` over the connector, then a ranged read of the kind the thumbnail worker makes. Read-only; it never writes into the folder it is pointed at. 2,299 images across 233 directories in 137 ms, and 380 across 13 in 29 ms. Against 34.1 s for 17,185 RAWs over WebDAV with pruning available. Recorded in docs/storage.md §5.2 and ARCH §8.4a, because a capability trade-off argued from a number nobody measured is the kind that gets quietly reversed later.	2026-08-29 09:57:52 +02:00
dtourolle	f12aece07e	Make storage pluggable, and prove it with a folder backend `RemoteBackend` existed from the first release and bought nothing it was designed for. Seven files in `dr-ui` constructed a `NextcloudBackend` directly, an account was a server URL beside a DAV user id, the local cache directory was named after a hostname, and the launch screen knew that signing in meant a browser handshake. The trait was real; the seam was documentation. A trait over operations is only a quarter of it. Pluggable storage needs four things, and this adds the other three: - Capabilities — already there, and the reason the engine can drive two backends at the speed each actually runs at. - Configuration — `dr_sync::Account`: where a library lives, in whatever form its connector addresses, with no server in it. Loads every existing config unchanged (`backend` defaults to `nextcloud`, `endpoint` is stored under its historical `server` key), and `Account::namespace()` reproduces the old catalog directory byte for byte, because changing it would abandon a catalog, its thumbnail shards, and the sidecars holding unsynced offline work. - Registration — `BackendProvider` and `BackendRegistry`. `ui/dr-ui/src/remote.rs` is now the only file above `dr-sync` that names a connector. `Connection` (an account plus an optional `Secret`) replaces the credentials-and-user-id pair that was threaded through fifteen signatures in an order that could be swapped. `Secret`'s inner string is reachable only through `expose()` and its `Debug` prints `Secret(***)`, so the indirect leak — a `{:?}` on anything holding one — no longer compiles into a leak. Nextcloud is unchanged and keeps every peculiarity: propagating ETags, chunked upload v2, `oc:fileid`, the `oc:permissions` probe on a refused PUT, the 423 retry classification, Login Flow v2. Those are what the capability model exists to serve, not something to hide. `dr-sync-folder` is the second connector: a local disk, a network mount, an external drive, or a folder a Nextcloud client already syncs. No account, no credential — the route that works where no secrets daemon does. It declares `LocalEtags` rather than claiming propagation a POSIX directory cannot provide, which costs nothing because 50k `stat` calls are not 50k PROPFINDs. Identity is a path hash, not an inode: an inode survives a rename but differs between devices and is reused after a delete, so two machines would disagree about which photograph a thumbnail belonged to. Re-deriving a thumbnail is a cost; showing the wrong one is a bug. docs/storage.md is the contract — the traits, the four steps to add a backend, and what each connector declares. ARCH §8.0 and §8.4a, and FR-NC-13, say why.	2026-08-29 09:57:52 +02:00
dtourolleandClaude Opus 5	c8e05831f4	Show the confidence, and say which curve it came from FR-CULL-9 was read as "no fit, no number", so every library without 200 confirmed positive pairs showed "Confidence unavailable" on every suggestion — which is every library, until enough confirmations exist to fit one. The confirmations are made on this screen, ranked by the number it was withholding, so the degraded state was also the permanent one. There has always been a curve: Calibration::default is the reference implementation's fitted MBF sigmoid, which is what clustering already operates at. It is a published operating point, not an invention, and what the requirement forbids is presenting it as though it were measured on this library. So the percentage is shown, and the screen says once, above the grid, where the curve came from. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-29 09:57:38 +02:00
dtourolleandClaude Opus 5	f41b3f6e8e	Answer "what would the other device end up with" without the other device Build and test / Desktop (Linux) (push) Successful in 2h7m52s Details Build and test / Layer separation (push) Successful in 59s Details 🐳 Android image / Build and push (push) Successful in 3s Details Build and test / android-image (push) Successful in 4s Details Traceability / Requirement traces (push) Successful in 37s Details Build and test / Android (aarch64) (push) Successful in 22m22s Details A tablet showed 200 of a person's 611 faces after syncing, and the obvious suspects — that suggestions deliberately do not travel, that the cross-device face match was too strict — were both wrong. Finding that out meant reading a catalog on a release-signed Android build, which cannot be done. So this stands the second device up locally: an empty catalog, given the images a scan would have found, the shards adopted into it exactly as a sync does, and the real catalog merged in as the remote. Then it counts, per person, against what the source holds. person source here Catherine 611 611 Me 242 242 Ian 219 219 Which settled it: the merge carries everything, and the shortfall was transfer — shards that never finished arriving. Worth keeping, because "did the sync lose this or has it not got here yet" is a question that will come up again, and guessing at it cost most of an evening. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 23:09:28 +02:00
dtourolleandClaude Opus 5	d2af6a3981	Time out on a stalled transfer, not on a slow one The HTTP client had a 60-second total request timeout. That is not a hang detector; it is a floor on link speed. A face shard runs to 25 MB, so it demanded a sustained 425 KB/s or the transfer failed — and having failed it was retried on the next pass and failed again, for ever. A tablet on ordinary wifi could therefore never finish taking in a library's faces, and nothing said why: each attempt looked like a network blip rather than an arithmetic impossibility. The catalog snapshot is 36 MB and has the same problem. `read_timeout` fires when no bytes arrive for the period, which is the condition actually worth failing on. A slow transfer that is still moving now finishes, however long it takes; a connection that has genuinely died is still caught in a minute. The connect timeout is unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 23:08:16 +02:00
dtourolleandClaude Opus 5	2c84aa1224	Write face shards the way the rest of the catalog writes The face store was the one part of the catalog still on SQLite's default rollback journal at `synchronous = FULL`. The catalog itself runs WAL at `NORMAL` (`schema::configure`) and so does the thumbnail store; nothing decided this one should differ, it was simply never set. Measured on this project's own filesystem, that is 21.3 ms per commit against 0.05 ms — four hundred times. And an export commits four times per photograph: the shard's transaction, then three separate autocommitting writes to the index. Ten thousand images is on the order of fourteen minutes spent doing nothing but waiting for fsync, before a byte goes to the server. That is the "checking faces…" that appeared to hang. So: WAL and `synchronous = NORMAL`, matching the rest, and the three index writes fold into one transaction. `NORMAL` is the same trade the catalog makes — a shard is derived data, and losing the last commit to a power cut costs one image re-exported. WAL brings an obligation with it, because a shard is uploaded by reading its file: the newest commits live in a `-wal` sidecar that no upload sends, so without a checkpoint the server would receive a database missing exactly the faces just written, and a peer would adopt it and see nothing wrong. `checkpoint` folds the logs back in, with `TRUNCATE` rather than the default passive mode, which gives up when a reader holds the log and would leave the same gap while reporting success. Two tests: that the store is in WAL like everything else, and — the one that matters — that a checkpointed shard copied without its `-wal` still holds every face. That second one fails without the checkpoint, which is how it was confirmed to be testing something. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 23:08:10 +02:00
dtourolleandClaude Opus 5	a1e8f494b9	Say what the face sync is doing while it does it The face pass set the status to "checking faces…" once and then said nothing until it was finished. On a library whose first export after a re-index is 9,849 photographs and 85 MB of shards, that is eight minutes of a progress bar sitting still — which is indistinguishable from a hang, and was reported as one twice. Nothing was wrong with the sync. The only fault was that it was silent. Three places now report, which are the three that take real time: - Preparing, per image with a count, since this is the long one and the only one whose length the user cannot guess from anything on screen. - Sending, per shard with its size, because a face shard carries crops and runs to tens of megabytes — one of them is a visible wait on any connection. Announced before the upload rather than after, since the wait is the upload. - Taking in a peer's shard, which is a download and then a row-by-row merge. The export reports every 25 images rather than every one, so the channel behind it stays lost in the write it accompanies. `export_to_shards` keeps its old signature and delegates, so the callers that do not want progress do not grow a parameter for it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 21:56:16 +02:00
dtourolleandClaude Opus 5	0729dfa359	Stop the face sync opening a database per photograph "Checking faces…" never finished. `export_to_shards` asks, for every indexed image in the library, whether the shard store already holds that image at that index time — and `indexed_at` answered by opening the shard database, running its six-statement schema batch and two `pragma_table_info` queries, then querying. Once per image. 9,849 times for this library, on every sync pass, before a single face had been written. The index time now lives in the store's `index.sqlite` alongside the shard number, so the question is one indexed lookup on a connection that is already open. It stays in the shard as well — that copy is the one that travels — but nothing reads it from there on the hot path. The write side had the same shape: `put_image` opened the shard afresh for each image, which mattered little when exports were a handful of new photographs and matters a great deal now that a re-index sends thousands. The handle is kept and reused, invalidated by shard id so sealing a full one and moving to the next drops it without anything having to remember to. `INDEX_SCHEMA` is `CREATE ... IF NOT EXISTS` like the shard schema, so the new column is added on open for an index already on disk — the same trap, caught the same way. Two tests: that the index time survives reopening the store, and that an index written before the column can still be opened and written to. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 21:19:56 +02:00
dtourolleandClaude Opus 5	76eece8500	Add the columns an existing shard never got `SHARD_SCHEMA` is entirely `CREATE ... IF NOT EXISTS`, which does exactly nothing to a table that already exists. So `crop` and `indexed_at`, both added to that batch, never appeared in any shard that had been written before — and the `INSERT` naming them failed with "no such column". Which took face export down completely, on every library that had ever synced a face. Silently: `export_to_shards` returns the error, `sync_face_shards` logs it at warn, and the sync goes on looking successful while the catalog fills with faces no other device will ever see. This library's shard sat frozen at 1,807 faces with 15,194 in the catalog, and the reason was this rather than anything in the export logic. Shards are upgraded on open now: both columns are additive and nullable, so catching up is one `ALTER` each. There is deliberately no version counter — "does this column exist" is the question actually being asked, and asking it directly cannot fall out of step the way a counter can. A peer's shard is opened read-only and cannot be repaired, so one written before crops is read as it stands, with a `NULL` standing in for the column. An adopted face simply has no crop, which is the truth about it. Four tests, built against the pre-crop schema written out in full rather than derived from the current one — the point being that it is not the current schema and must not track it. Verified against the real 1,807-face shard on this machine: the ALTERs apply, writes succeed, and nothing already in it is lost. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 09:50:55 +02:00
dtourolleandClaude Opus 5	4ed10f7b23	Let a person cross from one device to another The face shards carry boxes, landmarks and embeddings. What they deliberately do not carry is who anybody is — the person rows, their names, and the assignments joining the two. Those travel in the catalog snapshot, which is a whole-file copy and does contain them. But the snapshot is merged, not adopted, and this merge only ever looked at collections and keywords. `face_shard`'s own module note says people travel in the snapshot; nothing implemented it. So a second device received every face and no people at all, and drew an empty People screen over a full catalog. Exactly what a tablet showed after syncing thousands of faces from a laptop. What travels is what the user decided, following the rule the rest of this module already follows — judgements travel, inference is rebuilt: - People, by uuid on `revision`, exactly as a collection is: the name, and whether the group was set aside. - Confirmations, and rejections — "this is not her" is a fact too, and is why re-clustering does not put it back. - The suggestions inside an ignored group, which are otherwise ordinary inference but are what anchors the ignore. Without them a group set aside on one device reappears on the other, the same fault that made "Not interested" not stick locally. Ordinary suggestions are not carried. Both devices hold the same embeddings and clustering is deterministic, so each recomputes them and arrives at the same answer; shipping them would double the merge for no new information. A face has no cross-device identity, and unlike a collection there is no uuid to give it one. Both devices do agree on `oc:fileid` and roughly on the box, so a remote face is matched to the local face on the same photograph whose box overlaps it most, above 0.5 IoU. That is not a new rule — it is the one `record_detections` already uses to carry a confirmation across a re-index, and it is loose on purpose: the question is "the same face in the frame", not "the same rectangle". A local confirmation is never overwritten. Two devices confirming one face as different people is a real disagreement and an assignment carries no revision to settle it with; taking the remote's answer would let a sync undo what the user just did on the device in their hands. The remote's schema is probed rather than assumed: `remote_is_mergeable` admits any catalog at or below this version, so one written before faces existed, or before V10 added `ignored`, is ordinary. An absent table skips this half instead of aborting a merge that would otherwise have succeeded. Nine tests, including that the name lands on the overlapping face and not its neighbour in the same frame, that a set-aside group stays set aside, that an ordinary suggestion does not travel, and that merging twice changes nothing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 09:27:17 +02:00
dtourolleandClaude Opus 5	cf614efa61	Send a re-indexed image's faces to the other devices `export_to_shards` asked `store.contains(file_id)` and skipped anything the shard store had already heard of. So an image was exported exactly once, and re-indexing it updated the catalog and nothing else — every other device kept the first answer for ever. That is not hypothetical. This library was re-indexed after the detection floors changed and crops were added, going from 1,807 faces to 15,194; the shard store still held the original 1,807, written before any of it. Nothing the re-index produced could reach another device. The shard's `indexed` table now carries the catalog's own `indexed_at`, and the export compares against it. A re-indexed image goes again; an unchanged one still costs nothing. Copied from the catalog rather than stamped when the shard is written, because a shard-local write time advances even when nothing changed and could not answer the question. The column is nullable so a shard written before it still reads: absent means "cannot vouch for it", which forces one re-export and then settles. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 09:27:16 +02:00
dtourolleandClaude Opus 5	1d4c3348af	Let a 32-pixel face count, and move the blur floor with it 64 source pixels was too strict: it threw away 70% of everything the detector finds, and plenty of what it took were faces a person could name. Lowering it is not a one-line change, because the two floors are coupled. A face under 112 pixels is upsampled to reach the embedder and upsampling invents no edges, so a small face scores low on sharpness however crisp the original was. Re-measured over the reference library with `face_index --quality`: min crop min sharp size cut blur cut kept 32 0.000 44% 0% 56% 32 0.002 44% 3% 53% 32 0.005 44% 8% 48% 32 0.010 44% 16% 40% 32 0.020 44% 27% 30% 64 0.020 70% 7% 23% Holding the blur floor at 0.020 while dropping the size floor to 32 would have rejected a further 27% — for being small rather than for being blurred — and kept only 30%, barely more than the 23% the strict pair kept. Most of the point of lowering the size floor would have gone straight back out through the other gate. 0.005 removes 8% of what the size floor leaves, which is the same job 0.020 was doing at 64 (7%): the large-but-soft face this gate exists for. Together they now keep 48% of what the detector finds, against 23% before. The box pre-filter follows down to 24, staying below what the real floor accepts so it cannot reject a face that would have cleared 32. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-28 08:30:09 +02:00
dtourolleandClaude Opus 5	a1790c3e67	Stop indexing faces too small or too blurred to be anyone The library was storing faces at 52 source pixels and embedding whatever came back. There was a size floor, but it was 40 pixels on the bounding box, and there was no blur gate at all — so a subject walking through a half-second exposure detected confidently, aligned cleanly, and produced a perfectly ordinary-looking 512-vector. Nothing downstream can tell that apart from a real face, and because blurs resemble each other more than they resemble the people they were, they cluster together and weld unrelated identities into one group. Two floors, both measured rather than guessed. `face_index --quality` runs the detector over real proxies with both gates disabled and prints the distribution; over 1,503 faces in 600 images of the reference library: percentile crop px sharpness 1% 16 0.0006 25% 23 0.0025 50% 38 0.0071 75% 76 0.0284 99% 352 0.4282 The median face in a personal library is 38 pixels. Most of what the detector finds is background: people across a square, a face on a poster, a stranger at the next table. They are real detections and useless identifications. Size, on the crop rather than the box. "At least 64x64" has to mean the pixels the embedder sees, and the box is not that — the ArcFace template reaches past it for forehead and chin, so the aligned crop spans roughly 1.3x the box's shorter edge. The floor is therefore `min_source_px` on the aligned crop, applied after the warp fixes the scale, and `min_face_px` drops to 48 as what it always really was: a cheap pre-filter set low enough that it cannot reject a face the real floor would have kept. Sharpness. Variance of the Laplacian divided by the variance of the luma it was taken over. The division is the part that matters: raw Laplacian variance scales with contrast, so a threshold on it would quietly discard every backlit portrait in the library. The ratio asks how much of the crop's variation is edges rather than broad gradients, and is invariant to exposure. What each pair removes, cumulatively, of everything the detector finds: min crop min sharp size cut blur cut kept 64 0.000 70% 0% 30% 64 0.010 70% 3% 27% 64 0.020 70% 7% 23% 80 0.010 76% 2% 21% 64 and 0.020. The size floor does most of the work, and the blur floor removing only 7% on top of it is the point rather than a disappointment: at 64 pixels most faces are already sharp, and what it takes out is the large-but-soft one — precisely the face that would otherwise contribute a confident, wrong embedding. The two gates are not independent and the doc comments say so: a face under 112 pixels was upsampled to reach the embedder, and upsampling invents no edges, so small faces score low on sharpness even when the original was crisp. That is why `--quality` prints them together. This will re-index. Around 70% of what the current settings store falls below the new floors — faces between 20 and 40 pixels that nobody could identify. The People screen gets shorter and every group in it gets better. 66 dr-face tests pass, including that a blurred crop scores below a sharp one, that halving the contrast does not move the score, and that an upsampled face scores below the same face at full size. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-27 22:17:38 +02:00
dtourolleandClaude Opus 5	79c0520506	Keep the face, not just a way to find it again A face was drawn by decoding the 1024px proxy it was found on and cutting the box out again, every time the People screen opened. That made the screen a derivative of the thumbnail cache: evict a proxy — which the cache may do at any moment — and the cell goes blank, with no way back short of re-fetching the original over the network and re-detecting it. It also cost a full JPEG decode per image, per visit, to show a 96px cell. So the crop is cut once, when the pixels are already in hand at detection time, and kept. A 160px JPEG is a few KB against the ~250 KB proxy it replaces reading. Where it lives is the interesting part. The catalog snapshot is uploaded whole on every sync and downloaded by every device, so a crop column there would put tens of MB on every round trip — the exact cost `face_shard`'s 25 MB cap exists to bound, and the reason bulk per-face data lives in shards already. Crops therefore travel in the face shards, beside the embeddings, and `snapshot_for_upload` strips them from the copy it writes. Nothing reads a crop out of a merged remote catalog — the merge touches collections and keywords only — so a receiving device loses nothing. A shard carrying crops holds around 3,500 faces rather than 22,000, which is the price of a second device showing People immediately instead of re-fetching every proxy. The column is nullable and the reader falls back to the proxy, so a face indexed before this still works and the next indexing pass fills it in. V10 also adds `people.ignored`, for a person the user has looked at and does not want to identify. Most clusters in a real library are strangers — passers-by, other people's guests, a face on a poster — and there is no way to tell "not yet looked at" from "looked at, don't care" without recording the second. It is a column rather than a deletion because a deleted cluster comes straight back on the next Regroup: the faces are still there and still similar, and nothing short of remembering the judgement survives re-clustering. Same argument `face_person_rejected` makes one level down. And `prune_empty_unnamed`, for what clustering leaves behind. Regroup creates a person per unanchored group and never removed the previous run's now-empty ones, so pressing it twice added a rail entry per group it no longer believed in. Named people are never touched however empty — a name is user data — nor is a merge tombstone, which must outlive its faces to keep redirecting. 298 tests pass, including that the snapshot carries no crops while the live catalog keeps them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-27 21:17:44 +02:00
dtourolleandClaude Opus 5	7275c020d7	Group people at the threshold the library actually supports 0.90 left a third of the reference library ungrouped: 1,213 of 1,813 faces in a group, and the rest sitting alone in a screen that had nothing to offer for them. "Is 0.90 too tight" is not answerable from the number. It is a probability, and which cosine it lands on depends on the calibration — so the first half of this is a way to ask the question properly. `face_index --tune` runs the real clusterer over the real embeddings at ten thresholds and prints what each one produces. It writes nothing; comparing thresholds by applying them would have each one pollute the next. On the reference library: P cosine groups grouped largest 0.95 0.449 311 62% 51 0.90 0.403 316 67% 51 0.85 0.374 318 70% 57 0.80 0.353 328 74% 69 0.75 0.335 327 77% 69 0.70 0.319 326 79% 81 0.50 0.267 303 85% 90 The count of groups is the signal, not the count of grouped faces. Loosening from 0.95 makes it climb: real people are being assembled out of fragments. It peaks at 0.80 and then falls — and a falling group count while the grouped faces keep rising is the shape of over-merging, separate identities being welded together. That is the FR-CULL-10 failure, and the one the user cannot undo by hand. So 0.80: the loosest setting still building people rather than melting them together. A third more of the library gets grouped than at 0.90, and the largest group grows by eighteen faces rather than by forty. The table is one library, and the doc comment says so — `--tune` reruns it on any other. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-27 21:17:16 +02:00
dtourolleandClaude Opus 5	c10dca984f	Regroup the library without stopping the window Pressing Regroup on a real library did not come back. Clustering 1,813 faces is the textbook agglomeration — compute every pairwise cosine, then repeatedly scan all live group pairs, score each with average link, and merge the best — and the scan is inside the loop. Each merge rescans every surviving pair, and each score is recomputed from scratch over every cross pair. Some 1.6 million pair scores per merge, some 700 merges to do. Three changes, none of which alter the answer. Only above-threshold pairs can ever matter. An average that reaches the threshold must have at least one term at or above it, so two groups with no qualifying pair between them can never merge — not now, and not after any sequence of merges, since merging only adds terms. The new `neighbours` module produces exactly that sparse list: 7,875 pairs rather than 1.6 million on the reference library. It also means the n^2 matrix is never materialised, so memory goes from O(n^2) to O(edges) — 2.5 GB to a few hundred KB at 25,000 faces. Merges cannot cross components, so the connected components of that graph are independent problems: four hundred small agglomerations instead of one large one. Average link is additive — sum(A u B, C) = sum(A, C) + sum(B, C) — so a merged group's scores follow by addition. Kept as running (sum, count) per adjacent pair, a score costs one division instead of a nested loop, and a heap with lazy invalidation replaces the rescan. Measured on the reference library: 0.28s, release, for all 1,813 faces. An exact ANN index was tried and removed, and neighbours.rs records why so it is not rediscovered as a good idea. IVF with a triangle-inequality bound is exact and prunes beautifully on synthetic clusters; on real embeddings it prunes nothing — 946 of 946 cell pairs survive. Median pair angle is 88.5 degrees and the merge threshold is 66.2, so the bound needs cells of radius under ~10 degrees, but two photographs of the same person sit 36-60 degrees apart. No ball-based partition of a 512-d near-orthogonal space can be tight enough. So the scan stayed exhaustive and got an unrolled dot product and its blocks spread across cores instead. Correctness is held by keeping the old implementation as an oracle: three tests run both engines over the same population — plain, under co-occurrence and anchor constraints, and with a size-weighted calibration — and assert the clusters are identical. Determinism is asserted at a size where the threaded path is in play. 62 tests pass. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-27 21:16:55 +02:00
dtourolleandClaude Opus 5	2c56729354	Read a descriptor's variants without taking them Build and test / Desktop (Linux) (push) Successful in 21m4s Details Build and test / Layer separation (push) Successful in 27s Details 🐳 Android image / Build and push (push) Successful in 1s Details Build and test / android-image (push) Successful in 2s Details Traceability / Requirement traces (push) Successful in 24s Details Build and test / Android (aarch64) (push) Failing after 33m42s Details Two agents worked in parallel and neither could see this. The frame-budget instrument matches `ParamKind::Enum { variants }` by value, which was free when a descriptor was `&'static` and everything in it was borrowed for the life of the program. Descriptors are owned now — a declaration parsed at run time cannot hand out a `&'static` — so `variants` is a `Vec` and the arm was moving out of a shared reference. Bound by reference instead. The arm only ever reads the length. The kind of conflict that survives a clean textual merge: git had nothing to report, and the two changes are only incompatible once they are in the same tree. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-27 19:25:15 +02:00
dtourolle	fb1b44ce47	Merge branch 'worktree-agent-a1e5c8cb565255f5b' into master # Conflicts: # docs/traceability.md	2026-08-27 19:21:10 +02:00
dtourolleandClaude Opus 5	8202c05d9d	Format the zoom test the way the gate asks for it Build and test / Desktop (Linux) (push) Successful in 1h22m24s Details Build and test / Layer separation (push) Successful in 2m57s Details Traceability / Requirement traces (push) Successful in 1m3s Details 🐳 Android image / Build and push (push) Successful in 2s Details Build and test / android-image (push) Successful in 2s Details Build and test / Android (aarch64) (push) Failing after 33m41s Details Whitespace only. `cargo fmt --check` is a required step and the FR-DSP-5 test arrived disagreeing with it — kept as its own commit so it can be skipped wholesale rather than read for a change that matters. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-27 19:20:23 +02:00
dtourolleandClaude Opus 5	0cd3ef1b3f	Run a node declaration without compiling it `ops/.yaml` plus `build.rs` has been the class-1 plugin format since the declarative nodes landed — it was simply resolved at build time. Nothing about a declaration requires the compiler: everything it produces is data plus a WGSL string, and the composer already assembles WGSL at run time from whatever operations are active. So this is not a new mechanism. It is the existing one, loaded later (FR-PLG-2). `DeclaredOp` implements `Operation` from an owned `Declaration` — one interpreter over many declarations, where `build.rs` emits generated code per node. The generated path stays, as FR-PLG-2 says it should: a generated `match` is faster than an interpreted one, the built-ins' declared `tests:` have to run under `cargo test`, and generated source is inspectable in a way an interpreter's state is not. The reader is now one file, read by both.* `src/declared/decl.rs` and `src/declared/expr.rs` are `#[path]`-included by `build.rs` as well as being modules of the crate, and they produce a neutral `Declaration` that names no Rust type. The build script's job is reduced to rendering that declaration as Rust; `DeclaredOp` converts the same declaration into descriptors and `Expr::eval` walks the same tree the renderer writes out. There is one grammar, one set of validations and one set of error messages, so "a plugin is the same kind of thing as a built-in" is structural rather than aspirational. What remains genuinely written twice is the pair of backends — an arithmetic node rendered as Rust here and evaluated there — and that is what the parity test stands between. `tests/declared_parity.rs` parses every built-in declaration at run time and asserts the composed WGSL is byte-for-byte what the generated implementation produces, with the uniform block bit-for-bit identical, at both ends of every parameter's range and at four interior points; then again over the whole develop chain with the declared nodes swapped in, which is what covers uniform slot ordering and helper de-duplication between operations. A third test asserts the declared and hand-written nodes partition `ops/` between them, so coverage cannot shrink silently. Bit-for-bit rather than within a tolerance, because a tolerance is where a real divergence hides. The one thing that had to be got right for that to hold is number literals: `expr::as_f32` rounds a decimal exactly once, through the same shortest-round-trip text the compiler is handed, rather than rounding an `f64` a second time. Not in scope, and deliberately untagged: load-time WGSL validation (FR-PLG-11), id namespacing, a plugin directory read at startup, and pass nodes (FR-PLG-2a). Those are separate work, and tagging them from here would be the overstatement the spec's own §7 warns about. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-27 19:08:34 +02:00
dtourolle	3b5bba62f1	Merge branch 'worktree-agent-aa9f4356c13893373' into master # Conflicts: # docs/traceability.md	2026-08-27 19:08:32 +02:00
dtourolleandClaude Opus 5	772a69711d	Prove that zooming to 1:1 reads the source, and only then tag FR-DSP-5 FR-DSP-5 has been satisfied for some time and untagged. `Framing::view` shrinks the sampled region while the render target keeps its size, so a zoom raises the resolution the pipeline works at rather than magnifying pixels already drawn — there is no second full-resolution path because the zoom is that path. Tagging it on that basis alone is what §7 of the display spec warns against: traceability counts a requirement as covered when a comment names it, and checks nothing about the code under the tag. So the tag goes on tests instead, and the tests are built so that removing the behaviour breaks them. Both failure modes were checked by hand: deleting the view from `visible_rect` leaves the 1:1 render flat, and dropping only its offset leaves the render exactly inverted. The assertion message names both, since those are the two ways this can go wrong and the numbers alone do not say which. The fixture is one-pixel black-and-white stripes — the highest frequency an image can hold, and precisely what a proxy discards. A 1024 px source in a 128 px viewport reads source column `8x + 4` for every output column `x`, all the same parity, so the fit render comes out uniform; that is asserted first, because a 1:1 render showing detail proves nothing unless the proxy is known to carry none. What remains is an equality against the source bytes rather than a claim that something looks sharper. The third test takes the arbitrary zoom the requirement also names, and pins `RenderScale` beside the pixels: a zoom that moved the pixels but not the scale would sharpen at the wrong radius, which stays invisible until somebody compares a preview against an export. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-27 19:05:05 +02:00
dtourolleandClaude Opus 5	0c3b8cb1c4	Hand out descriptors a declaration could produce `Operation::descriptor()` returned `&'static OpDescriptor`, and that lifetime is the whole reason a build-time node is free and a run-time node is impossible: only a compile-time literal can satisfy it, so no amount of reading `ops/.yaml` at startup could ever produce a descriptor the rest of the application would accept. FR-PLG-2 says a bundled operation and a third-party plugin are the same kind of thing, differing only in where the file was found — and a lifetime outsiders cannot meet is exactly the second, weaker format that requirement forbids. So a descriptor is now owned and handed out as `Arc<OpDescriptor>`, with `Vec` where it held `&'static` slices. `Arc` rather than a `&self`-borrowed reference because the callers want to keep* it: the develop panel collects descriptors and then mutates the graph, and a borrow would tie the descriptor's lifetime to a borrow of the operation it came from, which is the one thing `&'static` was doing right. The identifier newtypes deliberately did not follow. `ParamId` is `Copy`, is compared in `match` arms against generated constants, is a map key in the sidecar and history, and reaches Slint model rows; an `Arc<str>` there would cost a refcount on every one of those and would take `match id { EXPOSURE => .. }` away from the generated code. They gain an interner instead, which is honest about its lifetime rather than pretending to one — the set of ids is bounded by deduplication and is process-lifetime by construction, because the sidecar on disk names its parameters and an id has to stay resolvable for as long as any edit naming it can be opened. No behaviour changes. Every descriptor that was a `static` is a `LazyLock` initialiser now, `Operation::helpers` borrows from `self` instead of being `'static` so a future run-time node can own its list, and `Warp` and `Framing` follow `Operation` so there is one shape rather than two. The one place a descriptor is read per frame is `compose_full`, which takes `descriptor().id` to prefix each active operation's uniforms, and `dr-ui` composes on every frame it draws. That is a dozen atomic increments beside a composition that is already building several kilobytes of WGSL on the same call; it is noted at the trait method rather than left for a profiler to find. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-27 19:02:25 +02:00

... 2 3 4 5 6 ...