The preview ladder is a published contract, not a shared constant

Ce contenu n’est pas encore disponible dans votre langue.

0071 — The preview ladder is a published contract, not a shared constant

Context

The preview pipeline renders a ladder of raster variants per asset. The default is col (320² cover), preview (1024 contain), screen (1920 contain), hires (4096 contain) — but the ladder is operator-configurable via sysconfig, and DefaultPreviewConfig() is a default, not a contract.

For most of the project’s life the client did not know this. Cards hardcoded /variants/col as their only image URL, because col was the one rung they could assume existed. That assumption cost real features:

Responsive srcset was deliberately disabled in PostCard — the code and its sizes machinery were left in place with a comment saying they awaited a signal that did not yet exist. Every card served a 320px square regardless of viewport (#502).
Widescreen art displayed as a square centre-crop, visibly disagreeing with its own hover-scrub animation, which used true aspect (#589).
A proposal existed to add a landscape-specific rung to fix this — solving a knowledge problem with more storage.

The naive fix — publish the four default keys as a shared constant — reproduces the bug one layer out. An install that drops hires to save storage, or adds a rung, would have a client confidently requesting URLs that 404.

Decision

The ladder is published data, and neither side hardcodes it.

The server states what exists per asset. ladder_available is true iff every configured rung is stored for that asset AND the caller passes the content plane (ADR 0064 — a restricted asset reports false, never 403, so the flag cannot become an oracle). It is computed against sysconfig’s configured variants threaded into the query as a parameter — never a literal list. One exported SQL fragment serves all call sites so the guard cannot be dropped in one of them.
The server states what the ladder is. GET /previews returns each rung’s key, fit and max_dim. A flag saying “the whole ladder exists” is not actionable without knowing what the ladder contains: srcset needs the keys for URLs and max_dim for width descriptors.
GET /previews is public-mode governed, not unauthenticated. It is registered in auth.PublicSurfaceRoutes: anonymous on a public install, 401 on a private one. Deliberately not excused the way /appearance is — fonts render the login card, so an install that refused them could not draw its own sign-in page, whereas nothing before sign-in needs image rungs.
The client caches the ladder and degrades to col. A 401, an offline failure or a malformed response all mean the same thing: no ladder, use col only. The failure direction costs a feature, never a 404.
fit: cover rungs are excluded from srcset. col is a square crop; offering it as a width candidate for a contain-mode slot would letterbox or distort. The grid’s contact-sheet mode still uses col directly and deliberately.
The ladder’s SOURCE shape is recorded per asset, in asset_field_value under the existing pixel_width / pixel_height definitions, set_by = 'computed' (#757).

The client can size a tile before any byte arrives only if the server states the shape. Points 1–5 tell it which rungs exist and what they are; none of them says what shape the picture is, and a layout that has to wait for naturalWidth reflows every tile on load. Masonry rendered a wall of identical squares for three merged PRs because that fact had a reader, an API field, a bucketer and a CSS rule — and no writer.

The quantity is the shape of the image the contain rungs are built from, not “the source file’s pixels”. Half a catalogue has no source pixels: a 3D model, a font, an audio file and a plain-text document have none, yet each produces exactly one image on its way through the pipeline — a turntable frame, a glyph specimen, a waveform, a rendered plate — and fans it across the ladder. That image is what a card renders, so its shape is what a tile reserves. A 2048×384 waveform is a 5.33:1 tile; nothing in the file it came from says so.

It is not recorded on storage_variants. That table is keyed by object hash and describes stored rungs, and the ladder source is not one of them — col is 320² for everything, so a per-variant row would either need inventing for an object that does not exist or would answer with the crop’s shape rather than the picture’s.

Consequences

Adding or removing a rung is an operator action with no code change. Both sides discover the ladder at runtime.
A hardcoded rung key anywhere is a bug, on either side of the wire. This is the rule most likely to be violated by a future change that “just needs the hires URL”.
Landscape tiles required no new variant and no backfill — #589 collapsed into “request a different rung”, which is the clearest evidence the diagnosis was right.
ladder_available and preview_available are different questions and both are needed: the latter means “a servable col exists” (render a thumbnail at all), the former means “the full ladder exists” (safe to build a srcset).
The flag is nearly co-extensive with preview_available on a healthy install — 1004 of 1007 assets on the reference dataset have both. That is expected and not a reason to alias them: the distinction is what lets the client stop guessing, and it degrades correctly for partially-rendered or failed assets.
A handler that never reaches the ladder never records the shape. Eight handlers short-circuit a non-forced re-queue before decoding anything (the same early exit that leaves a pre-#645 thumbhash unhealed), so the backfill for those is aa rebuild-previews --force. Raster and video are the exceptions — both reach the stamp on an ordinary re-queue, so they backfill without re-encoding a single rung.
The recorded pair is post-EXIF-rotation, because the ladder source is. That is the shape a viewer sees and the shape the rungs were cut from; the EXIF extractor’s own write is the on-disk pair, and where both have run the preview value is the better one.
Related trap, recorded in ADR 0008’s amendment: a derived variant under a stable content hash is exactly what preview regeneration rewrites, so ladder URLs addressed by asset id are not immutable. Cache validators must derive from the stored bytes.

Amendment 2026-08-11 — §3 generalises: “public-mode governed” is the default shape for a new anonymous read (#709)

§3 decided this for GET /previews. #709 needed the same call for GET /browse-views — the operator’s enabled browse layouts — and the reasoning transferred without modification, which makes it a rule rather than a one-off. Recording it here so the next anonymous read does not re-derive it, and because the brief for #709 got it wrong in the direction that looks safer.

The rule. A new read that anonymous callers need takes security: [] plus registration in auth.PublicSurfaceRoutes — anonymous on a public install, 401 on a private one. Being excused from that governance, the way /appearance and /site-text are, requires a specific justification: the endpoint renders the login card. Fonts, branding and UI strings qualify because an install that refused them could not draw its own sign-in page. Nothing else does by default.

Where the #709 brief went wrong. It specified unconditional security: [], citing /site-text as the precedent, on the reasoning that “a logged-out visitor browses too.” The premise is true and the conclusion does not follow: a logged-out visitor browses only on a public install, which is exactly the case public-mode governance answers. On a private install the browse switcher is already behind auth, so there is no anonymous surface for the layout set to serve — and answering anyway would hand an unauthenticated caller the operator’s configuration for nothing. The coding agent caught this and applied §3 instead. Verified on a live stack: public install + anonymous → 200, private install + anonymous → 401, private install + signed in → 200.

The guard that makes this hard to get wrong is a build-time one, and it is worth knowing about. security: [] on an operation that is not named in PublicSurfaceRoutes (or in notGovernedAnonymousOps) fails TestPublicSurfaceCoversAnonymousOperations. So the brief’s version would not have shipped quietly — it would have failed CI in front of the author. That is the property publicmode.go describes as recovering the allowlist guarantee “at BUILD time instead of request time,” and #709 is the first time it fired on a genuinely mistaken instruction rather than an oversight.

⚠️ The corollary for anyone reading security: [] in openapi.yaml: it does not mean “ungated.” It means “this operation’s gate is not a credential” — the gate may still be the public-mode toggle. The /browse-views operation carries a comment saying so at its security block, because the two-word declaration reads exactly like the thing it is not.

Amendment 2026-08-02 — the announcement flags are DB-first BY DESIGN, and reconcile keeps them truthful (#829)

preview_available / ladder_available are answered from storage_variants rows, not from backend stats — that is load-bearing for the zero-console-404 contract (the server must be able to answer cheaply on every list row) and is not changed. What #829 adds is the missing half of the bargain: the render/skip path now heals the rows (ADR 0008’s amendment), so a restored backup or any bytes-without-rows state converges back to truthful announcements on the next requeue instead of deadlocking every card into the placeholder. Serving itself was and remains backend-first (Download → Backend.Get); only the announcements were ever at risk.

Amendment 2026-08-02 — `scrub_available` extends the announcement pattern to the hover sheet (#835, #832)

The hover sprite-scrub was the one preview surface still doing what this ADR exists to stop. Its gate was the file extension (isVideo || is3D), and its geometry was a grid hardcoded per extension (10×10 for video, 6×6 for the 3D turntable). Both were guesses about storage made from a filename, and both were wrong in both directions:

A video whose expensive preview.video job has not drained yet has a col (from the cheap poster job, #818) and no sheet, so the card requested one and 404’d — the exact class ADR 0064 / #471 removed everywhere else.
A format that DOES have a sheet but was not on the list could never scrub. Animated GIFs (#832) are the case that forced this: they now render a sheet and the extension list would have hidden it.
The grid is not always full. writeSpriteSheet floors the cell interval at 0.2s, so a clip shorter than cells × floor cannot fill it and ffmpeg’s tile filter pads the remainder with black. A 5s clip fills 25 of 100 cells; the client cycled all 100 and spent three quarters of the hover on padding it could not distinguish from a dark frame.

Two additions, both instances of rules this ADR already states.

scrub_available — “a servable sprites.vtt exists for this asset AND the caller passes the content plane”. Identical ADR 0064 contract to preview_available / ladder_available, computed from the same ContentReadable decision in the same pass at every site that produces one (browse list, single GET, collection members, post members), so the three can never disagree for a restricted asset. It is keyed on the cue file, not the sheet: both are written together by all three producers, so either would answer “is there a sheet”, but only one answers “can I drive from it”.
The cue file is the geometry, and the client reads it. sprites.vtt already declared one cue per populated cell with an exact #xywh rect. Cycling the cues rather than a grid makes video, 3D and GIF one code path, gets short clips right, and needs no constant on the client at all — the same “ask, don’t assume” move rung keys made in the original decision. spriteCellBox has already moved once (160 → 240, #811) without the client noticing; nothing about the grid may be knowable to it either.

Deliberately not extended to FeaturedItem. The featured rail draws its own tile and has no hover scrub, so the flag would be a field nothing reads. When the rail adopts CardThumb it inherits the requirement, and the required-field contract in cardAsset.ts is what will make that a type error rather than a missing animation.

The client-side compatibility half is the point. The cue-driven cycle fixes every sheet already in storage with no re-render, because the truncation was already encoded in the stored VTTs — the client simply was not reading them. One filter carries that: pre-#835 VTTs end with a zero-length cue (the old writer emitted the cue and then broke on start >= duration), which addresses ffmpeg’s first padding cell, so the parser drops any cue whose window is empty. The writer no longer emits one; the filter stays for the sheets that already have it.

Amendment 2026-08-02 — the scrub cycles what the VTT declares, and GIFs are first-class (#836)

The hover scrub’s geometry contract is now cue-driven: the client fetches sprites.vtt and cycles exactly the cues it declares, using their #xywh rects. The per-kind hardcoded grids (10×10 video, 6×6 turntable) are gone from the client; a sheet’s real frame count is whatever the VTT says, which is what lets a short clip stop at its last real frame and an animated GIF’s sheet (new preview.gif kind, #832) join the same code path with no third special case. The parser drops empty-window cues — pre-#836 writers emitted a trailing zero-length cue on short clips, and dropping it client-side is what fixed existing sheets without any re-render. Scrub availability is signalled (scrub_available) rather than probed, preserving the zero-console-404 contract this ADR established.