The container dropped privileges to the non-root snapotter user via gosu
(external) and s6-setuidgid (embedded), both of which preserve the
environment without setting HOME. The app therefore kept root's HOME=/root,
which is not writable by snapotter, and PaddleOCR died with
PermissionError: '/root/.paddlex/temp' -- breaking the ocr tool at default
quality in every non-root deployment. Prior GPU QA ran the app as root, which
masked it.
Fix: export HOME=/data/.home (persistent, writable, hidden) at every
privilege-drop point:
- entrypoint.sh external gosu path and non-root tini path (the latter uses
$DD/.home so a DATA_DIR override stays consistent).
- the s6 snapotter/run service (scoped there, not globally before /init, so
postgres/redis do not inherit a snapotter-owned HOME).
The root preflight creates /data/.home and the existing chown sweep owns it as
the PUID/PGID-remapped snapotter; the dir is added to both ensure_writable
probes so an unwritable HOME fails fast with the storage-permission guidance
instead of crashing late. The Dockerfile passwd home moves from /app
(read-only) to /data/.home as the getpwuid fallback when HOME is unset.
Because bridge.ts forwards HOME to the Python sidecar, this also repairs the
expanduser("~") caches in inpaint/outpaint/restore/noise_removal/remove_bg,
not just PaddleOCR.
Also fixes a test-harness inconsistency: tool-default-settings passport-photo
countryCode "us" -> "US" (the route exact-matches uppercase PASSPORT_SPECS
codes; the UI already sends "US", so users were never affected).
Claude-Session: https://claude.ai/code/session_01XGB4pGvTvb7sUX4JN745U7
Closes the "e2e never runs in CI" hole. Adds per-PR e2e smoke gate,
nightly full-suite workflows, parallel vitest forks (per-fork DBs),
Playwright parallel/serial/visual projects against production builds,
metadata-generated test suites (drift guards, hostile inputs, format
matrix, pairwise settings, property-based fuzz), Stryker mutation
testing, Schemathesis API fuzz, coverage ratchet, and fixes for three
session-poisoning bugs that caused 200+ serial-bucket failures.
Bug fix included: favicon/split/bulk-rename could hang clients forever
when ZIP streaming failed after reply.hijack().
