Owner review 1, phase 2 (docs/requests/2026-10-15-owner-review-1.md, R2).
Jobs now run one at a time on the desktop and on the server. Parallelism
(Settings > Rendering, TITANO_PARALLELISM) is how many photos the running
job processes at once: thumbnail and proxy passes draw that many photos
together and still write the pack and checkpoints in frame order, decode
tiers are probed that many at a time, the deep analysis measures that many
frames and a render decodes that many ahead. Wherever originals are
developed whole (TIFF, PNG, RAW without a usable preview), fewer run at
once within the memory budget, which the running job now has whole; the
server's frame window is sized to the host instead of 1 GiB.
A new job waits its turn and pausing the running job starts the next one.
An optional priority queue gives each job kind a level (renders, imports
and downloads high, verifies and proxies low by default): a queued job of
a higher level pauses the running one without asking, and that job goes
back to the queue with its checkpoint and resumes by itself afterwards.
Jobs of the same level, and a job's own follow-ups, never interrupt it.
The Jobs screen and the admin panel show the parallelism and the queue;
Settings > Rendering and the admin panel's Settings set them, at once.
TITANO_WORKERS and the Workers key of server.json still set the
parallelism (1 to 64, with a warning at start for the variable).
Titano.Tools gains i18n prune, which removed the renamed keys from the 36
machine-translated languages; the new strings are translated in all 38.
Found while testing: uploads and imports failed when a pass read the
sidecar while they wrote it (atomic writes now retry the rename while a
reader holds the file, and never for a read-only one), and taking a lease
just released while another process read it failed on Windows (the create
now waits for the reader). Five adversarial review rounds checked the
phase before this commit.
Also updates BRIEF (12.8, 14.6, 14.8, 14.9, 24), DECISIONS, CHANGELOG,
ARCHITECTURE, STATUS, the request file and the compose/Unraid templates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner review 1, phase 1 (docs/requests/2026-10-15-owner-review-1.md).
Thumbnail and proxy jobs over several sequences counted sequences while
they checkpointed inside one: the rate decayed until the time left
overflowed TimeSpan, the job failed while publishing progress and again
while saving it, and stayed Running. The throughput is now seeded by the
first window with progress, the time left is bounded (JobEta), saving
progress never throws, image passes count frames, and job speeds use the
right unit (bytes for verify and import).
The debugger stops in WicDecoderProvider.cs when presets changed were
cancelled preview decodes leaving synchronous thread-pool delegates; the
editor's sources, LatestRunner and the render node's proxy task now keep
them in their own tasks, and a cancellation is never an unreadable frame.
GoPro DNGs converted by Adobe software were classified GPR and read only
through their 8-bit preview, and Windows' preview-only DNG decoder was
reported as a native decode. A DNG is GPR only when its raw image is
VC-5, WIC decodes RAW only through the Raw Image Extension and never DNG,
analyses and previews take an adequate embedded preview by choice while
renders develop natively, and frames already indexed as GPR are
reclassified by the next light analysis.
Also records the owner's request (docs/requests), the BRIEF amendments
(§8.1, §8.2, new §24), DECISIONS, CHANGELOG, STATUS and ARCHITECTURE.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each tag upsert locks its row until commit, and tags were written in the
order the caller listed them. A PATCH listing them in another order than
an upload, a scan or another edit of a sequence sharing two of them could
lock the rows in opposite orders, and PostgreSQL aborted one writer with a
deadlock that nothing retries. Tags are now written in the order of their
keys, and a set the sequence already carries is not written again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
ArchiveIndexer.SyncAsync and the ingest's in-step append (and the first
frame of an API sequence) read the sequence row and upsert the whole row
from that read in one transaction. On PostgreSQL (READ COMMITTED) a
PATCH accepted between the read and the upsert was written back over:
its settings were lost and settings_version went back, so the PC that
pushed them later pulled the old settings over its edit.
The row is now read with SequenceRepository.GetForUpdateAsync (FOR
UPDATE, as EditAsync takes it). A PATCH committed first is then moved by
the insertion like any position; one arriving later sees the new frames
version and gets 409 frames_conflict.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
- Preview sizes come from probes and unscaled decodes: a size-limited decode
picks the smallest preview that reaches its edge (CR3 PRVW 1620x1080), so
renders and deep passes record its tier without a size and the output cap
keeps the largest preview's.
- A read error is not an undecodable file: ImageDecoder.Decode throws the
IOException when a provider could not read the file and none produced
pixels. Deep passes never flag frames they could not read or that
vanished, and frames only an interrupted earlier run found undecodable
are decoded again before they are flagged.
- Renders resolve stand-ins only for frames the pipeline asks for; frames
the read-ahead meets on the way (skipped by speed-ups) are left alone.
- A GoPro GPR finds its JPEG in the index when an import or upload kept it
in a sequence of its own (same original stem and capture time), for
probes, deep analysis, renders, thumbnails, proxies and the editor;
GPRs whose JPEG is indexed later are probed again.
- Thumbnails judge an embedded preview by its own JPEG header.
- PostgreSQL writes tier fields 1000 frames per statement.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
BRIEF §8.2 tiers 2 and 3 had no writer: PreviewOnly and Undecodable were never set,
so the advisor could not cap the output at the preview size and one unreadable frame
failed a whole render.
- ImageDecoder.Probe names a file's tier without decoding it (JPEG, PNG and TIFF
without opening the file); the light analysis probes frames whose tier is unknown,
and the deep analysis, renders from originals and thumbnail/proxy passes record
what decoding actually found (FrameTiers). Sidecar frames gain "tier", "previewW"
and "previewH"; the index gains frames.preview_width/height and
sequences.preview_only_count (migration 0005, SQLite and PostgreSQL).
- SequenceContext takes the preview size from preview-only frames, so the advisor's
output cap applies; the light analysis reports format tiers (preview-only frames as
information, undecodable ones as a warning under the deep analysis' key, listed
once); a sequence that gains undecodable frames warns at once.
- Only a file in a format the host decodes natively is flagged undecodable; files
beyond the host's decoders keep their tier without the flag, so render nodes and
proxies can still supply them.
- A frame that stops decoding during a render is replaced by its nearest planned
neighbour and logged instead of failing the job; the render fails only when no
planned frame decodes.
- The Library card shows a preview-only chip with the count (local and server
sequences: SequenceDto.PreviewOnlyCount, FrameDto preview size).
- GoPro GPR decodes through the JPEG the camera writes beside it when that is larger
than the embedded preview (BRIEF §8.1).
- Thumbnails use a stored embedded preview only when it reaches 320 px; a smaller one
sends the frame through the host's decoders instead (BRIEF §8.5).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
SequenceQuery.Rendered and SyncStates back the Library's Rendered, Not
rendered and Synced status choices. ISyncRemote.FirstFrameAsync reads one row
of the frames page (default: the whole list), so a remote card without a hero
frame no longer downloads the server's manifest for its thumbnail.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Server sequences that are not in the local index open through a RemoteSequenceSource:
proxies from the cache, project from the server's settings with their version (a 409
merges again), analysis report from the server, analyses and renders posted as server
jobs (one per preset) and followed through the server's events. Set hero now reaches the
index (SequenceEdit.HeroFrameId) so the Library card shows it. The Export tab gains the
Sync option (remembered in the project) and restores "Save both" copies through undo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Union of the Tooltips resources; sample data keeps the sync server and the
simulated cards; the import test names App.Infrastructure.Formatting fully now
that a test namespace of the same name exists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
- SQLite restore holds every pool slot instead of the write lock first,
removing a lock-order deadlock with sessions that begin to write.
- PostgreSQL sessions always return their connection on dispose, a
rollback while disposing never masks the original exception, and a
busy session keeps its transaction until it is discarded.
- The pool disposes a connection whose health check the caller cancels.
- Listener commands keep their lock until the server answers and never
tear a write with the caller's token; closing and pending no longer race.
- Tag replacement locks the sequence and the tag row; unused-tag cleanup
locks candidates and re-checks them with a fresh snapshot; render
counts and stats refreshes lock the sequence row on PostgreSQL.
- Sequence upserts keep rendered_count, which renders maintain.
- AuthenticationOk after an unfinished SCRAM exchange is rejected.
- Oversized read buffers are released; failed backups leave no .tmp;
a zero idle lifetime is rejected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Shared abstract suites run against in-memory/temp-file SQLite and, with
TITANO_TEST_PG, a private PostgreSQL schema per test: migration
idempotency and version refusal, keyset paging with ties for every sort,
statistics, exclusive job claiming under 8 concurrent claimers, and the
100k-frame bulk insert under 15 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC