# Decisions One line per decision: date · decision · alternative · reason. | Date | Decision | Alternative | Reason | |---|---|---|---| | 2026-10-02 | The prototype dump `docs/reference/titano-prototype.md` was not supplied; algorithms are implemented from the brief's descriptions and the cited literature, with the §20.4 thresholds as acceptance tests. | Wait for the dump | Not available in the repository or uploads; the brief fully specifies behaviour and gates. | | 2026-10-02 | Custom WPF theme dictionaries (`Titano.Ui/Themes`) implementing DESIGN.md tokens instead of `Application.ThemeMode` Fluent. | Fluent theme + overrides | DESIGN.md fixes exact colours/sizes; Fluent is still flagged experimental (WPF0001) and would need restyling of most controls anyway (§2.1 permits in-house dictionaries). | | 2026-10-02 | Added shared WPF library `src/Titano.Ui` (theme, icons, controls, `Loc` markup extension) used by App, Uploader and Setup. | Duplicate resources per app | One visual system for three programs. | | 2026-10-02 | Added `tests/Titano.Core.Windows.Tests` for WIC / Media Foundation / DXGI tests. | Put them in App.Tests | Keeps engine interop tests independent of the WPF app. | | 2026-10-02 | Localization areas extended with `Common`, `Jobs`, `Uploader`, `Setup`. | Fold into `Shell` | Smaller, owned files; same key scheme. | | 2026-10-02 | Resources live in `i18n//[.].resx`, embedded into `Titano.Core` (satellite assemblies), base name `Titano.Core.i18n.`. Keys marked `invariant` need no translation. | Raw XML loader | Uses `ResourceManager` as the brief requires; satellites bundle into single-file publishes. | | 2026-10-02 | Program icons (ICO) are generated at build time into `artifacts/icons` by `Titano.Tools make-icon`, invoked from `Directory.Build.targets`. | Commit generated ICO files | Brief forbids committing generated binaries and raster sources. | | 2026-10-02 | In-house CRC-32 (`Titano.Core.Hashing.Crc32`). | System.IO.Hashing | That assembly is a NuGet package, not part of the shared framework. | | 2026-10-02 | Windows projects target `net10.0-windows` (platform version 7.0) without a Windows SDK projection TFM; Windows 10 APIs are reached by P/Invoke/COM. | `net10.0-windows10.0.19041.0` | Avoids the CsWinRT projection; toasts use the in-window toast host plus tray balloon. | | 2026-10-02 | Test projects use an explicit `TestMain` entry point. | Top-level statements | Avoids the `Program` type clash with the server's top-level program. | | 2026-10-02 | SQLite P/Invoke binding lives in `Titano.Data` with a library resolver (winsqlite3 on Windows, system libsqlite3 elsewhere for tests/dev); the Linux server still refuses SQLite at configuration time. | Binding in Core.Windows only | Lets repository tests run on Linux CI against both dialects. | | 2026-10-02 | Client preferences needed before the database opens (language, theme, window placement) are stored in `settings.json` next to the database. | Settings table only | Language and theme must apply before the first window appears. | | 2026-10-02 | Server Windows-service hosting will use an in-house advapi32 P/Invoke host. | Microsoft.Extensions.Hosting.WindowsServices | That assembly is a NuGet package, not part of the shared framework. | | 2026-10-02 | Admin panel files are embedded resources served by an in-house middleware with ETags and SPA fallback. | ManifestEmbeddedFileProvider | Its manifest generator ships in a NuGet package. | | 2026-10-02 | Container health check runs `Titano.Server --healthcheck` (in-process HTTP probe). | curl/wget in the image | Keeps the runtime image free of extra tools. | | 2026-10-02 | Container drops privileges with `setpriv` to PUID:PGID; `/data` is chowned, `/archive` only checked for write access. | Recursive chown of the archive | The archive may hold terabytes on NAS shares. | | 2026-10-02 | Release workflow publishes to `RELEASES_REPO` (default `/titano-releases`) when the `RELEASES_TOKEN` secret exists, otherwise to this repository. | Fail without the token | First releases work before the public repository is set up. | | 2026-10-02 | Release notes are generated by `Titano.Tools changelog` from `git log` piped on stdin. | Tool spawning git | Keeps the tool free of external process calls. | | 2026-10-02 | CI `dotnet format` gate checks whitespace formatting. | Full style/analyzer verification | Analyzer rules already fail the build (warnings as errors); style preferences stay suggestions. | | 2026-10-02 | Test runner exits 0 for a project with no tests and 2 when a filter matches nothing. | Always 2 | Empty projects during bootstrap are not failures; mistyped filters are. | | 2026-10-02 | COM via `[ComImport]` + explicit `ReleaseComObject`; flat APIs via `LibraryImport`. | `[GeneratedComInterface]` | Deterministic release keeps files unlocked; the apps are not trimmed. | | 2026-10-02 | WIC decodes to 64bppRGBA and linearises with the exact sRGB LUT (128bpp float only for float sources). | Always 128bppPRGBAFloat | WIC's integer→float gamma is not the exact sRGB curve and premultiplication darkens edges. | | 2026-10-02 | Exif orientation is applied in managed code while de-interleaving. | `IWICBitmapFlipRotator` | Microsoft documents quadratic behaviour of the flip-rotator over scanline decoders. | | 2026-10-02 | Volume watcher uses a hidden top-level tool window plus 3 s polling. | Message-only window | Message-only windows do not receive broadcast WM_DEVICECHANGE. | | 2026-10-02 | Camera RAW is "native" via WIC only when the Raw Image Extension package is installed. | Decoder CLSID alone | The in-box RAW codec shares the CLSID but only reads embedded previews. | | 2026-10-02 | HEVC/AV1 decoder presence is detected with `MFTEnumEx`. | Store package names | Also covers OEM and hardware decoders. | | 2026-10-02 | Math code lives in `src/Titano.Core/Math` with namespace `Titano.Core.Mathematics`. | `Titano.Core.Math` | That namespace would shadow `System.Math` in every `Titano.Core.*` file. | | 2026-10-02 | CCT = McCamy refined to the nearest Planckian point (Krystek, CIE 1960 uv), valid for \|Duv\| ≤ 0.02. | Plain McCamy | McCamy alone misses the 100 K gate above ~9000 K. | | 2026-10-02 | Deflicker keeps (1 − strength)·ρ·S of a frame's deviation (ρ own Tukey weight, S window surviving fraction). | Blend by S only | S alone barely isolates outliers. | | 2026-10-02 | LOWESS uses the k nearest valid samples with bandwidth = farthest distance + one spacing. | Fixed half-width | Constant sample count at sequence edges. | | 2026-10-02 | Clock jump = backward step > 3× median interval or a ±1 h offset from cadence (DST); counts as one interval. | Any whole-hour step | Scheduled multi-hour pauses look the same. | | 2026-10-02 | Time zone for solar labels: recorded offset → camera−GPS clock (15 min) → longitude/15 with a "time zone estimated" finding. | No day phases | Usable labels, honestly flagged. | | 2026-10-02 | Holy-grail step is "received" when the detrended median jump ≥ 50 % of the metadata step (same sign); metadata amplitude is redistributed with smoothstep ramps lengthened until ≤ 0.3 EV/frame. | Measured amplitude | Flicker makes the measured jump noisy. | | 2026-10-02 | Flicker fundamental = first autocorrelation peak within 80 % of the strongest; window ∈ [15, 151], odd. | Strongest peak | Avoids harmonics at fractional periods. | | 2026-10-02 | Optical flow: dense pyramidal Lucas–Kanade with box filters, neighbour-flow correction, σ = 1 px pre-blur; grid step 4/2/1 px for Draft/Balanced/Best. | Farnebäck | Separable, SIMD-friendly, stable and as accurate on smooth motion. | | 2026-10-02 | Phase correlation: Gaussian-shaped peak + log-parabola sub-pixel fit, both patches in one complex FFT. | Foroosh sinc fit | Exact for Gaussian peaks, half the FFT work. | | 2026-10-02 | Stabiliser: noise-corrected texture gate, exhaustive pairwise similarity hypotheses (≤ 36) + weighted LS, path smoothed by Gaussian-weighted local linear regression (σ = 12). | RANSAC / plain Gaussian | Deterministic with nine patches; keeps intentional pans. | | 2026-10-02 | Interpolation chooses velocities by forward–backward consistency with per-sample flow-agreement weights (τ = 3 px). | Super SloMo combination | Better at motion boundaries (36.8 vs 33 dB). | | 2026-10-02 | Virtual-camera zoom is interpolated geometrically; trail length L ↦ decay 0.01^(1/L). | Linear zoom / half-life | Even perceived speed; the trail is gone at its stated length. | | 2026-10-02 | Motion estimation, sharpness and noise use sRGB-encoded luminance. | Linear light | Shadows keep texture weight; less exposure dependence. | | 2026-10-02 | Archive presets use 0.30 bpp (≈60 Mb/s at 4K24 HEVC 10-bit); share presets use the brief's per-codec defaults (0.10/0.07/0.05). | 0.07 bpp everywhere | An archive must survive later re-encoding; star fields are the hardest content. | | 2026-10-02 | All metadata fallback chains (time, ISO, serial, lens, rating, size) live in `MetadataBuilder.Build`; parsers only collect candidates. | Fallbacks inside each parser | One place for priority rules; parsers are order-independent. | | 2026-10-02 | `MetadataReader.Read(path)` throws when the file cannot be opened; malformed content becomes warnings (with a parser-fault safety net the fuzz tests require to stay unused). | Swallow everything | An unopenable file is not malformed input. | | 2026-10-02 | HEIF `irot`/`imir` override Exif orientation (imir axis 0 = top↔bottom). | Exif first | HEIF decoders apply the transforms themselves; both would rotate twice. | | 2026-10-02 | Main image size comes from the container's own size field (SOF, IHDR, ispe, VP8X, RAF crop record, RW2 borders, CR3 CMP1) before raster IFDs. | Always the largest raster IFD | That is what each container defines and what decoders output. | | 2026-10-02 | Only MPF large-thumbnail entries count as previews; XMP date-only values are never used as capture times. | Every secondary image / midnight | Gain maps are not previews; midnight would be invented. | | 2026-10-02 | Media Foundation interop uses `[GeneratedComInterface]` with explicit release; threads entering MF call `Win32Com.EnsureInitialized()` (MTA unless already STA). | `[ComImport]` like WIC | Hardware encoders stall when sample references wait for GC; source-generated COM does not initialise apartments, so it is done explicitly. | | 2026-10-02 | MP4: 16-byte `free`+`mdat` placeholder rewritten to a 64-bit header past 4 GiB; negative composition offsets shifted to ctts v0 plus an edit list; parameter sets moved into avcC/hvcC; fast start on by default. | Always 64-bit mdat / ctts v1 | Small files stay maximally compatible; wider player support. | | 2026-10-02 | GIF: Wu-criterion median cut, frame differencing against the source last drawn, identical frames merged by holding one back. | Plain median cut / diff against dithered output | Max palette error 37 → small; 945 KB → 52 KB at equal PSNR; exact delays. | | 2026-10-02 | AVI always carries OpenDML indexes plus idx1 (1 GiB RIFF segments). | idx1 only | Files beyond 2 GB stay playable. | | 2026-10-02 | MF encoder capabilities: software H.264 ≤ 4096×2304, hardware assumed 4096² (H.264) / 8192² (HEVC/AV1); an unsupported size fails at media-type negotiation and the next encoder is tried. | Querying limits | MF has no reliable capability query. | | 2026-10-02 | Full-size JPEG IDCT is a port of libjpeg's accurate integer IDCT (bit-identical to djpeg); reduced-size (1/2–1/8) decoding uses float IDCTs with triangle chroma upsampling. | SIMD float IDCT everywhere | Float produced rare ±3 differences after YCbCr; reduced sizes gain 6.7 dB with triangle upsampling. | | 2026-10-02 | Codecs throw only `InvalidDataException` for bad structure or unsupported features; damaged scan data yields a partial image plus warnings. | NotSupportedException for features | One exception type for callers and fuzzing. | | 2026-10-02 | Linear DNG export: raw image in IFD0, lossless-JPEG tiles, ColorMatrix1 + ForwardMatrix1 for Rec.709 D65; source camera Make/Model kept when Exif is supplied. | Preview IFD + SubIFD raw | Simplest valid DNG; LibRaw renders test patches back exactly. | | 2026-10-02 | Resampler keeps negative/overshoot values in linear light. | Clamp to 0–1 | Ringing stays within each kernel's exact bound; clamping happens at encode. | | 2026-10-02 | Codec fixtures (each < 12 KB) are generated by `tests/fixtures/codecs/make-fixtures.sh` with libjpeg-turbo/libtiff dev tools and committed. | Fetch at test time | Reference outputs must be stable; the tools are not build or runtime dependencies. | | 2026-10-02 | Installer payload "TPK1": deduplicated, deflate-ratio-bounded entries after the setup stub; footer aligned to 8 bytes and found before an Authenticode table so signing works before and after packing. | Spec fields only | Every offset is bounds-checked; signtool can append without breaking the payload. | | 2026-10-02 | Setup runs `asInvoker` and relaunches itself elevated (`runas`) only for all-users installs; the unelevated parent starts Titano afterwards. | requireAdministrator | Titano never runs elevated. | | 2026-10-02 | Uninstaller self-removal via a %TEMP% copy started with `/CLEANUP`; running programs are only asked to close (never killed; silent mode exits 5). | MoveFileEx alone / kill | Per-user installs have no admin rights; user data safety. | | 2026-10-02 | Setup has its own `Program.Main` (silent and cleanup modes run without WPF) and is published with single-file compression (70 MB stub). | App.xaml-generated Main | Reliable exit codes; smaller installers. | | 2026-10-02 | SQLite: one in-process async write lock with `BEGIN IMMEDIATE`, shared-cache in-memory mode for tests, fixed-width ISO-8601 UTC text with microseconds and lower-case UUID text. | busy_timeout alone / epochs | Writers queue instead of failing; text order equals time order on both dialects. | | 2026-10-02 | PostgreSQL: parameters typed by OID, results in text decoded by OID, named prepared statements (128 LRU), session settings pinned in the startup packet, cancel = CancelRequest then socket abort after 5 s, transient retry once for idempotent reads only. | Untyped params / tracking ParameterStatus | Predictable decoding, no "could not determine data type", a dead server never hangs a caller, writes never repeat. | | 2026-10-02 | Migrations: all pending ones in one transaction under `pg_advisory_xact_lock`. | One transaction per migration | Atomic and safe when several servers start together. | | 2026-10-02 | Case-insensitive search/sort via lower-cased `search_text`/`title_key` columns (and `COLLATE "C"` on PG). | ILIKE / NOCASE | Identical results on both dialects. | | 2026-10-02 | Job claim: `UPDATE … WHERE id = (SELECT … FOR UPDATE SKIP LOCKED) RETURNING` on PG; immediate transaction on SQLite. | Advisory locks | Exclusive under concurrency, as specified. | | 2026-10-02 | One `JobEngine` in Titano.Data runs every long operation for both the client and the server: handlers per `JobKind`, progress persisted at most every 5 s, a failed persist (job no longer running) stops the handler, a shutdown leaves jobs running so the next start requeues them from their checkpoint. | Per-feature background loops | One queue, one cancel/pause/resume path and one crash-recovery rule everywhere. | | 2026-10-02 | Job throughput is a moving average over windows of at least 2 s. | Average since start | ETA follows speed changes (cache warm-up, network) without jitter. | | 2026-10-02 | Visual controls (card grid, filmstrip, charts, preview, curve editor) keep their layout and math in WPF-free classes; colours are theme brush keys re-resolved on theme switch; charts, filmstrip and preview never mirror in RTL. | Math inside controls / brush properties | Testable on Linux; view models hold no brushes; time axes and images must not flip. | | 2026-10-02 | `VirtualizingWrapPanel` does its own keyboard navigation and recycling; the focused container is kept alive. | ListBox navigation | Direct targets, wrapping across rows; focus survives scrolling. | | 2026-10-02 | Library card slot 236 × 208 with a 16:9 thumbnail. | Mockup's 150 px thumbnail | DESIGN §4 specifies 16:9. | | 2026-10-02 | Uploader sessions: key = 8 hex chars of the source folder's SHA-256 + first-frame time; leading test shots (first group < 8 frames) join the next session; watch-folder files always use automatic sessions. | Group by time only | Simultaneous cameras stay apart; unattended files never land in a sequence chosen for a manual drop. | | 2026-10-02 | Uploader: a watch file counts as processed only after the server confirms it; offline, rejected-token and changed-certificate failures do not use up the three retries; a changed certificate signs the user out. | Mark at enqueue / count every failure | Clearing the queue loses nothing; an outage does not exhaust a large batch; never upload to an unverified server. | | 2026-10-02 | Delete-after-upload also requires size and write time unchanged since hashing. | Server confirmation alone | BRIEF rule 8. | | 2026-10-02 | Uploader settings are an in-window panel; single instance via a per-session mutex and a current-user named pipe; native tray menu (`TrackPopupMenuEx`). | Dialog / WM_COPYDATA / WPF ContextMenu | §12.1 allows modals only for login and destructive confirmations; works while hidden in the tray; dismisses reliably. | | 2026-10-02 | Network watch folders are polled (default 5 min); local folders every 60 s plus `FileSystemWatcher` hints. | Watcher everywhere | SMB change notifications are unreliable. | | 2026-10-02 | The Dockerfile uses BuildKit's built-in frontend (no `# syntax=docker/dockerfile:1`). | Pinned external frontend | Every feature used is built in; the build no longer depends on Docker Hub (a 502 there failed CI). Base images come from mcr.microsoft.com. | | 2026-10-02 | SQLite sessions run every blocking call on the thread pool (`ThreadPoolHop`) and cancel through a per-connection flag polled by `sqlite3_progress_handler` (every 1000 VM instructions) in addition to `sqlite3_interrupt`. | Callers offload themselves / interrupt only | A repository awaited on the UI thread never freezes it; an interrupt that lands before `sqlite3_step` starts was silently lost and the query ran to the end. | | 2026-10-02 | Deep analysis persists per-frame shifts from the nine-patch `Stabilizer` and smooths the path itself; per-frame data is stored as columns (~80–110 bytes/frame) with a JSON-lines checkpoint that drops a torn last line. | `Stabilizer.Solve` / row objects | `Solve` cannot resume from a checkpoint; the report stays small for 100k-frame sequences. | | 2026-10-02 | Sky/land mask from a pre-pass over 15 sampled frames before the main pass. | Segment after the pass | Cloud coverage and per-region luminance need the mask during the pass. | | 2026-10-02 | Scene distance: chi-square on histograms smoothed by ⅛ EV and levelled to the same mean; flow statistics use brightness-levelled images and forward–backward-consistent vectors only; noise is median-based (Immerkær). | Raw chi-square / raw flow / mean noise | Flicker and exposure steps no longer read as scene changes or motion; stars no longer inflate noise. | | 2026-10-02 | Shake and motion thresholds are judged in output pixels (stabilise above 0.3 px RMS). | Source pixels | That is what the viewer sees. | | 2026-10-02 | Automatic output stops at 4K UHD; no automatic stacking advice. | 8K / temporal-median denoise | 8K stays a manual choice; a median erases moving stars and stillness cannot be verified reliably. | | 2026-10-02 | Apply records a missing-data reason only where no advice exists; Optimise returns modules to Auto and only adds to the user's hero, exclusions, offset or output cap; a preset's own choices are stored as manual values, content-derived values stay Auto. | Overwrite | Manual values stay in force; re-analysis does not undo a chosen preset. | | 2026-10-02 | Advisor namespace is `Titano.Core.Advice` (class `Advisor`). | `Titano.Core.Advisor` | A class named like its namespace needed an alias in every `Titano.Core.*` caller. | | 2026-10-02 | CI also force-pushes each same-repository PR run's screenshots as one commit to the `ci-screenshots` branch (separate job, only it has `contents: write`). | Artifact only | Artifacts need an authenticated download; a branch can be fetched with plain git for visual review. The branch never merges and holds a single commit. | | 2026-10-02 | Export fps is the encoded grid; `Timing.FrameRate` is the playback rate (source frames per video second, default = export fps). Durations are quantised to output periods on rounded 90 kHz boundaries; whole frames are sampled at the period centre. | One fps / free durations | Presets keep their meaning, Interpolated can synthesise slow motion, AVI and image sequences stay exact. | | 2026-10-02 | Adaptive timing is proportional up to 3× the median interval; longer gaps show for 3 + log2(r/3) frames (max 5); pauses ≥ 20× median and ≥ 120 s become a 0.2 s cut. Interpolated fills gaps of up to 3 missing/excluded frames. | Fixed clamps | Bursts stay faithful and long gaps cannot stall the video. | | 2026-10-02 | A non-monotonic time-remap curve is repaired (sorted, clamped, made non-decreasing) with a warning; virtual-camera keyframes are keyed to source frame positions. | Reject the curve / output frame index | The remap is always monotone; keyframes stay on the content when timing or fps change. | | 2026-10-02 | Per frame: exposure → one warp (stabilisation⁻¹ ∘ camera ∘ crop) to output size → flow, synthesis and blur in output space → stacking → knee. Stabilisation zooms by 1/max(crop scale, 0.75), warning below 0.75. Modules run only when Auto resolves true. | Flow in source space / no zoom | A single resample; flow sees stabilised content; borders stay hidden. | | 2026-10-02 | Blur fraction = BlurFraction(target × speed, real) / steps, speed clamped to 4, no blur across cuts. | Fixed fraction | The look survives speed changes. | | 2026-10-02 | Encoding overlaps processing by one frame; image sequences encode up to one frame per core into a hidden `.tmp` folder renamed at the end, named `/_000001.ext`. | Strictly sequential / `frame_%05d` | 1080p PNG 1.0 → 2.5 fps; crash-safe and editor-friendly. | | 2026-10-02 | JPEG/MJPEG use 4:4:4 at quality ≥ 90, else 4:2:0; HEVC/AV1 → H.264 fallback rescales the bit rate by bits-per-pixel; size capped to the highest level, then the encoder maximum; managed bit depth moves to the shallowest supported depth ≥ the request; linear DNG tiles are sized to the frame (≤ 256 px). | — | Predictable quality and sizes; small DNGs are ~3× smaller. | | 2026-10-02 | `Auto` carries `HasAdvice`; `Resolve` falls back only when there is no advice. | Treat `default` as "no advice" | A value type's advised 0/false was indistinguishable from "no advice". | | 2026-10-02 | `i18n check` also treats any C# string literal shaped `Area.Group.Name` with an existing area as a used key. | Marker functions such as `LocKey(…)` | Reason and warning keys travel as data; 53 render/export keys had no resources and the check did not notice. | | 2026-10-02 | Sidecars store sequence and frame ids, original name, UTC offset, time source, `verifiedAt` and session keys; compact JSON with numeric flags and named enums that read as Unknown; a sidecar with a newer `schemaVersion` is read but never rewritten. | Ids only in the DB / string enums / rewrite | A rebuild returns the same ids; unknown bits survive; data safety. | | 2026-10-02 | Frames stay in capture order: ingest inserts in place and renumbers later frames; thumbnails carry a source fingerprint and proxies an Exif capture time so stale ones rebuild themselves. | Append in arrival order / mapping file | Parallel uploads arrive out of order. | | 2026-10-02 | Ingest batches one sidecar write per sequence for concurrent commits, protected by a write-ahead journal; "new sequence with title" uploads map to a session key of user + title. | One write per file / new sequence per call | Large sidecars are not rewritten per frame; a crash between move and sidecar write recovers; parallel uploads with one title share a sequence. | | 2026-10-02 | Import sessions split per camera body and file format; the session key hashes body, format and first capture time; duplicates offer Skip or Import again, never overwrite. | Body only / random key / overwrite | RAW+JPEG would double every frame; an interrupted import continues its sequence; originals are never replaced. | | 2026-10-02 | Card erase re-hashes both the card file and the archive copy and refuses files touched since the dry run; adoption reads light metadata only and Verify fills the hashes. | Trust import-time hashes / hash during adoption | User data rule 8; adoption stays fast. | | 2026-10-02 | Index rows without a sidecar are kept and reported as orphans (only Rebuild or an explicit fix removes them); scan stamps are 100 ns ticks; a copied sequence folder gets new ids; index edits win over the sidecar when the index's settings version is ahead. | Auto-delete / timestamps / shared ids / sidecar always wins | A temporarily missing folder loses nothing; microsecond DB precision would make every scan see changes; duplicate ids corrupt the index. | | 2026-10-02 | Writers coordinate through lease files (2 min, renewed every 40 s, stale leases renamed aside); thumbnail passes take their own `thumbs-` lease. | OS file locks / the sequence lease | Locks are unreliable over SMB and across processes; long thumbnail passes must not block ingest. | | 2026-10-02 | One `thumbs.pack` format: the archive's `ThumbPack` (frame index + source hash, append-only entries, index + footer, walk-back recovery). The decode workstream's second pack format and proxy builder were removed; its LRU `ThumbnailCache` stays for the UI. | Keep both | Both wrote `TTPK` files with incompatible layouts to the same path. | | 2026-10-02 | Native RAW: the sensor saturation point only marks clipping (never rescales); white balance before demosaic, colour matrix after; highlight recovery lifts clipped channels to the unclipped mean then fades to grey; camera matrices match the exact model name; CR2 black level from masked columns; Sony as-shot WB falls back to daylight. | Per-frame white adjustment / prefix match / SR2 decryption | Per-frame rescaling would flicker a timelapse; D750 ≠ D7500. | | 2026-10-02 | Frame-window spill stores 32-bit floats; DNG DefaultScale stretches the axis with the larger factor; ForcePreview uses an embedded JPEG preview only when it is at least the requested size. | Half floats / shrink / always preview | Renders stay bit-identical under memory pressure; full resolution; no blurry thumbnails. |