Pulls mountedDropdown / requestId / debouncedSearch / dropdownState ownership
out of Tiptap's suggestion.render() closure into one createMentionController().
render() becomes a thin adapter: onStart→open, onUpdate→update, onExit→close.
This is the single-owner structure #628 needs for the AC-6 single-dropdown
invariant — the upcoming pencil re-edit affordance opens via the same
controller.open() instead of racing the suggestion plugin over module state.
open() now also bumps the request token so an open-A→open-B sequence discards
A's in-flight fetch (preserved increment-on-open semantics). No behaviour
change for the fresh-@ path — existing browser suite is the regression guard.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Convert the two bare failure echoes (gateway detection, /actuator status) to
::error:: so Gitea renders them as CI log annotations, consistent with the rest
of the deploy steps. No behaviour change. Raised in review (Leonie).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
obs.env documents POSTGRES_HOST but does not set a value, so obs-secrets.env
does not 'override' it — it is the only source. Reword the carried-over comment
to match reality. Raised in review (Tobias).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The unquoted <<EOF delimiter is load-bearing — under a composite action secrets
come from $VAR (env), not Gitea ${{ secrets }} substitution, so a re-quote to
<<'EOF' would write literal $VAR strings and the five-key non-empty guard would
not catch it. Adds a self-testing grep guard (matching the ci.yml 'Assert no X'
convention) so a future re-quote fails CI instead of shipping broken obs auth.
Raised in review (Felix, Sara, Tobias).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A failed cp/mkdir in the deploy-configs step was previously swallowed (the step
had no set -e), so a broken config copy could still reach the validate step. The
five-key guard catches empty secrets but not a failed copy. -u also catches a
typo'd env var name. Raised in review (Sara, Tobias).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds a Composite actions section covering the checkout-first ordering rule, the
secrets-via-inputs + unquoted-heredoc constraint (with the five-key guard and
shell: bash requirement), and a step-by-step for adding an input. Notes that the
inline Reload Caddy example now lives in the reload-caddy action.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Records the decision to extract the shared obs-deploy/reload-caddy/smoke-test
logic into three composite actions instead of a reusable workflow or shared
shell script. Numbered 029 (028 was taken by the pdf.js wasm ADR on main since
the issue was filed).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The reload-caddy pinned alpine digest moved out of the workflow files into a
composite action. Add .gitea/actions/** to the manual-review digest rule so the
digest stays watched and never silently goes stale (#603).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replaces the four inline obs steps with one uses: ./.gitea/actions/deploy-obs,
and the Caddy reload + smoke test with one uses: each (host
archiv.raddatz.cloud, postgres_host archiv-production-db-1, PROD_* secrets).
Removes all three '# Keep in sync with nightly.yml' comments — the shared
definition now enforces the invariant.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replaces the four inline obs steps with one uses: ./.gitea/actions/deploy-obs,
and the Caddy reload + smoke test with one uses: each (host
staging.raddatz.cloud, postgres_host archiv-staging-db-1, STAGING_* secrets).
checkout@v4 stays the first step; the #526 /import mount guard stays inline.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Five required, no-default inputs (incl. grafana_db_password for the #651
read-only reader role). Four named run: blocks keep the four CI log sections:
deploy configs, validate, start, assert health.
Secrets map to env: and are written via an unquoted <<EOF heredoc ('$VAR'
expands at the shell layer; a quoted delimiter would write the literal var
name and config --quiet would pass anyway). A five-key non-empty guard runs
right after the write, and chmod 600 is the final operation so the file is
never world-readable. ADR-016 absolute paths and the two-file --env-file
ordering are preserved.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parameterises the public-surface smoke test by host (one required input,
mapped via env: HOST). Keeps the three checks verbatim — login reachable,
HSTS value pinned, Permissions-Policy present, /actuator -> 404 — plus the
/proc/net/route gateway-detection and RESOLVE-array rationale.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
First composite action in the repo (establishes the convention). Lifts the
Caddy reload step verbatim from nightly.yml/release.yml — DooD privileged
sibling + nsenter to systemctl reload caddy, pinned alpine digest, reload
not restart. No inputs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Promote the future-CSP constraint from an inline Caddyfile comment to a
durable ADR-028: serve the pdf.js wasm decoders same-origin (never a
CDN), any future CSP must allow 'wasm-unsafe-eval' + worker-src 'self'
blob:, and the build-time guard keeps the wasm shipping. Caddyfile now
points at the ADR.
Addresses re-review: Markus (constraint should be an ADR, not a comment).
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Address the remaining UI/UX polish: add a warning-triangle icon so the
failure is signalled by shape, not colour alone (WCAG 1.4.1); give the
recovery download link a full 44px tap/focus target (inline-flex
min-h-[44px]); and soften the message copy in de/en/es.
Addresses re-review: Leonie (colour-only, undersized link, copy warmth).
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Enable svelte/no-target-blank so reverse-tabnabbing is caught at lint
time instead of relying on review (the very gap that left the viewer
download link exposed). Repo is already clean — all existing
target="_blank" anchors carry rel="noopener noreferrer".
Addresses re-review: Nora (optional detection-for-free).
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The error block was a colour-only, visually-small dead end. Add
role="alert" so screen readers announce the failure, bump the message to
text-base and the recovery download link to text-sm with a py-2 tap
target — the only escape hatch, sized for the archive's older readers.
Addresses re-review: Leonie (a11y of the error state).
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The "render failure" test rejected getDocument().promise — the load
path, not the render path — and only asserted a template constant. Now
the fake loads the document successfully and rejects the page render
(the actual #708 wasm-decode failure class), plus a negative companion
asserting the message is absent on a successful render. Also reset
renderTask to null on the render-error path.
Addresses re-review: Felix, Sara (mislabeled test / asserted a constant).
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The render path was localized but loadDocument still stored the raw
pdf.js message (and an untranslated English fallback), contradicting the
"never leak raw error text" principle. Both load and render failures now
set the localized doc_render_failed message.
Addresses re-review: Felix, Nora (raw error leak on the load path).
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The in-browser pixel-render fixture test was green locally but flaky in
CI: the real pdf.js worker could not fetch /pdfjs-wasm/ in the CI
Chromium container, so the CCITT canvas stayed blank (0 sampled pixels)
and failed the suite — green locally, red in CI, root cause not locally
reproducible. A flaky gate is worse than none.
This bug is a build/serve parity failure, so guard it deterministically
where it actually breaks: a postbuild assertion that jbig2.wasm and
openjpeg.wasm shipped into build/client/pdfjs-wasm/ (non-empty). It runs
after `npm run build` — including the Docker build stage — and fails the
build loudly if a future pdfjs bump makes the static-copy glob match
nothing. Combined with the getDocument(wasmUrl) unit guard and the
negative-path render test, the regression is covered without CI flake.
Addresses re-review: Tobias (no automated parity check), Sara (pixel
test not pinned). Render-decode correctness verified manually via
`node build` serving /pdfjs-wasm/jbig2.wasm as application/wasm.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
If a Content-Security-Policy is ever added, it must permit
'wasm-unsafe-eval' (script-src) and 'self' blob: (worker-src) or the
pdf.js wasm decoders and worker break and scanned PDFs render blank.
Forward-looking note so the future CSP author doesn't silently
reintroduce #708.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Render committed synthetic fixtures through PdfViewer with the REAL
pdf.js loader and assert the canvas is non-blank (sampled dark-pixel
count). The CCITT (G4 fax) fixture exercises the shared jbig2.wasm
decode path — the same module pdf.js uses for JBIG2 — so it transitively
covers the JBIG2 acceptance criterion (the archive sample found zero
true JBIG2 docs and jbig2enc is unavailable to synthesize one). The
JPEG/DCTDecode fixture guards against regressing the natively-decoded
path. Verified the CCITT case goes red when wasmUrl is removed.
Fixtures are hermetic, committed assets (~2-5 KB each), generated with
ImageMagick — never fetched from staging at test time. CI browser mode.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The error-state download link opened with target="_blank" but no rel,
exposing the opener to reverse tabnavbabbing. Add rel="noopener
noreferrer". Same-origin so low severity, but a one-token fix in a file
this issue already touches.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The error state showed a hardcoded German string ("Fehler beim Laden
der PDF" / "Direkt öffnen") to all users regardless of locale. Use the
localized doc_render_failed and doc_download_link messages so the
recovery path (message + working download link) is honest in de/en/es.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
renderCurrentPage swallowed every render rejection with a bare return,
so a decode failure left a blank white viewer with no feedback. Now a
non-cancellation rejection sets a localized doc_render_failed message,
which routes into the existing error UI (message + download link).
Cancellation (page-nav / zoom) still returns silently — no error.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Localized message shown when a PDF page cannot be rendered, so users
never see a blank canvas or a raw English pdf.js string. de/en/es.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
getDocument was called with a bare src string, so pdf.js 5.x had no
`wasmUrl` and could not initialise the JBIG2/CCITTFax wasm decoder —
CCITT (G4 fax) scans painted a blank canvas. Pass
{ url, wasmUrl: '/pdfjs-wasm/' }; the directory URL (trailing slash
required) is the single source of truth next to the worker config.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
pdf.js 5.x moved the JBIG2/CCITTFax/JPEG2000 image decoders into
WebAssembly. The wasm lives in node_modules and was never web-served, so
those decoders failed to initialise and CCITT (G4 fax) scans painted
blank in production while rendering fine in dev.
Add vite-plugin-static-copy (devDependency) to copy
node_modules/pdfjs-dist/wasm/* into build/client/pdfjs-wasm/, so the
assets are emitted into the SvelteKit client build and survive the
production Docker image — not just `npm run dev`. Verified that
`node build` serves /pdfjs-wasm/jbig2.wasm with 200 + application/wasm.
Refs #708
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
With the visible "Originaltext" line gone from every view, the
date_original_label message has no remaining references — remove it from
de/en/es. Also drop the now-inaccurate comments in documentDate.ts that
described the raw cell as "preserved separately as the visible secondary
line"; the raw cell now only feeds the SEASON word and is never shown.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The detail drawer's date cell rendered DocumentDate whenever a date OR a
raw cell was present (`{#if documentDate || metaDateRaw}`). For an
undated, raw-only document that meant the verbatim import text leaked
into the view. Tighten the guard to `{#if documentDate}` so such a
document shows "—". The raw prop is still passed through for the SEASON
word on dated documents. Covered by a new test.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DocumentDate rendered an "Originaltext: <raw>" secondary line for
UNKNOWN/SEASON/APPROX dates, gated by a showRaw prop. Drop the visible
line, the showRaw prop, the showRawLine derived, and the now-unused
date_original_label message import. The raw prop stays — it still feeds
the SEASON word in formatDocumentDate, which only ever maps a fixed
German season token (never emits raw text), so no XSS surface remains.
Update both DocumentRow call sites to drop the now-gone showRaw={false}
and the comment that justified it. Remove the two DocumentDate tests
that asserted on the deleted DOM sink (the UNKNOWN secondary line and
its XSS-escaping); the DAY/MONTH coverage stays.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The "Originaltext:" line in WhoWhenSection rendered the verbatim import
cell (metaDateRaw) as static text plus a hidden input that re-submitted
it on every save. Editors mistook it for an editable field. Remove the
visible line, the hidden round-trip input, and the now-unused rawDate
prop (here and at the DocumentEditLayout call site). The backend's
partial update preserves the stored value, so no data is lost.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The density chart is an interactive filter control; a 5-minute private
browser cache let it show stale month counts after an edit/upload/re-tag.
The in-memory aggregation is sub-200ms p95 over ~5k docs, so there is no
load reason to cache. Removing the explicit header lets Spring Security's
default no-store directive apply, so the response is always fresh.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Addresses Sara's review concerns:
- Add a negative Testcontainers test: saveAndFlush of a RANGE with end < start
throws DataIntegrityViolationException, proving chk_meta_date_end_after_start
actually fires (H2 wouldn't) and exercising the backstop's trigger end-to-end.
Guards against silent app/DB drift if the service guard ever regresses.
- Tighten updateDocument_acceptsRange_whenEndAfterStart to assert the persisted
end value, not just that save was called.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Addresses Tobias's review concern: the generic DataIntegrityViolation
backstop turned every integrity violation into a silent 400 with no
constraint name, no stack, no Sentry — an unanticipated write bug would
fail invisibly in production.
Now extract the constraint NAME from the cause chain (schema metadata, safe
for Loki) and log it parameterized at WARN, so the failure is debuggable.
Still never pass `ex`/`getMessage()` (SQL + values, CWE-209) and still no
Sentry — the response stays generic, so the response logic is not brittle.
New test proves the WARN names the constraint but never carries the SQL.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Testcontainers integration test persisting a RANGE doc with end == start
against real Postgres + Flyway, which (unlike H2) enforces the V69
chk_meta_date_end_after_start CHECK. Pins the app guard's isBefore
semantics to the actual >= constraint, guarding against app/DB drift (AC2).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add an endBeforeStart $derived to WhoWhenSection (lexicographic ISO compare,
no Date object) that renders an inline error on the end-date field —
border-red-400, aria-invalid, aria-describedby, and a #end-date-error <p>
inside the existing aria-live region — with a ⚠ glyph so the cue is not
colour-alone (WCAG 1.4.1). Save is not disabled; the server stays the gate.
Wire ErrorCode INVALID_DATE_RANGE through errors.ts getErrorMessage and add
the single key error_invalid_date_range to de/en/es, so the same translated
string is used inline (client) and via getErrorMessage (server fallback).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add @ExceptionHandler(DataIntegrityViolationException) returning 400
VALIDATION_ERROR with a fixed constant message, so any integrity violation
that slips past the upstream guards (a future constraint, or the import
path) becomes a clean 400 instead of a 500 + Sentry alert (AC9).
Deliberately generic — it does not inspect which constraint failed. Never
echoes ex.getMessage() (constraint name + SQL, CWE-209), logs at WARN
without passing the exception (would re-leak the SQL to Loki), and does not
call Sentry.captureException.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add the second validateDateRange predicate mirroring
chk_meta_date_end_only_for_range, so a direct API client that sets an end
date without RANGE precision gets a clean 400 INVALID_DATE_RANGE instead of
a 500 (AC6). Shares the code with the end-before-start branch.
Also fix updateDocument_preservesStoredPrecision_whenDtoOmitsIt: its stored
fixture (MONTH + end date) is a state the DB CHECK forbids, so the
carried-over-state guard correctly rejects it. Switched to RANGE + end —
the only DB-valid non-null-end combo — preserving the test's intent.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Cover AC2 (end == start), AC3 (open-ended, end null) and AC4 (null start +
end set, which must not reject or NPE), plus end-after-start. Guards the
guard against future over-rejection that would diverge from the DB CHECK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add ErrorCode.INVALID_DATE_RANGE and a validateDateRange guard on
DocumentService.updateDocument, run right after applyDatePrecision so it
fires before any save (updateDocumentTags persists earlier in the method).
Mirrors the V69 chk_meta_date_end_after_start CHECK: end >= start with a
null start allowed, using isBefore so equal dates stay valid. Turns a user
date typo into a clean 400 instead of a 500 + Sentry alert.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Mirror the useTranscriptionBlocks pattern: makeCsrfFetch(options.fetchImpl
?? fetch) wraps both the default and any injected fetch, so CSRF protection
holds regardless of how the hook is constructed — defense-in-depth against a
future caller injecting a bare fetch. Simplifies the CSRF test to assert on
the injected path instead of stubbing global fetch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The OCR trigger POST went through bare `fetch`, so it carried no
X-XSRF-TOKEN header. Spring Security rejected it and the UI showed
"Sitzungsfehler. Bitte laden Sie die Seite neu." (CSRF_TOKEN_MISSING).
Default the job controller's fetchImpl to csrfFetch — matching the
autosave hook — so mutating requests are CSRF-protected while GET
polling passes through unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The desktop AC8 test flaked in CI: it asserted replaceState was never
called after a tap, but the mount-time URL mirror fired late with the
unchanged default view (cx=0&cy=0&z=1), tripping the assertion. Assert on
the rendered viewBox instead — a pure function of the view state — so a
recentre shows as a shifted origin and a desktop tap leaves it identical,
with no dependence on the noisy mirror-effect timing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sara/Elicit noted AC8 was proven only as recentreAbove geometry, never as
wired behaviour. Add route-level tests that mock window.matchMedia: a tap
recentres the canvas (mirror effect re-fires) when the mobile breakpoint
matches, and leaves the view untouched on desktop where the side panel is a
flex sibling that never overlaps the canvas.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The shared parent-pair child loop read group.childIds[i] while iterating
the filtered childCenters, so a child without a position would desync the
id from the centre — and that index now also drives the active-connector
lookup. Ride the id on the mapped {id,x,y} centre so the two never drift;
a positionless child drops out of both together.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sara/Elicit flagged that AC5 was proven only at the isConnectorActive
predicate level. Add render-layer assertions: no connector group carries a
dim opacity when nothing is selected, and selecting Vater dims exactly the
vertical feeding the collateral child Tante. Exercises the shared
parent-pair per-child <g opacity> wiring.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bump DIMMED_OPACITY 0.4 -> 0.45 so dimmed outlines/labels stay legible
against bg-surface in both themes (dark mode dims already-light mint, the
riskier case). Import the constant into StammbaumTree.svelte.test.ts so the
node-opacity assertions track it instead of a hard-coded '0.4'.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Group opacity on the node <g> made the whole node translucent — including
its card fill — so the connector lines drawn beneath a dimmed node showed
through it. Render the card fill at full strength outside the dim group and
move the lineage focus+dim onto an inner content group (outline + labels)
only. The focus ring also leaves the dim group, so a dimmed-but-focused
node keeps a full-strength ring.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
