fix: harden install queue/dispatcher lifecycle and repair review-sweep regressions (#395)

Fixes 15 defects found by a max-effort multi-agent review of the last 6
merged PRs (#388, #390, #391, #392, #393, #394), all adversarially
verified before fixing.

Install queue + dispatcher (the serious cluster):

- features.ts: finalize the installer child exactly once. A failed spawn
  fires both "error" and "close", and the second event released the file
  lock and active slot that pump() had just handed to the next queued
  bundle, letting two pip processes write the same venv concurrently.
  Outcome recording now happens before pump() so the next bundle's first
  progress frame cannot race the previous install's bookkeeping.
- feature-status.ts: keep failed-install errors in a per-bundle map
  instead of the single progress slot. With the queue auto-starting the
  next install, the slot was overwritten within seconds and a failed
  install vanished without ever surfacing to GET /features.
- bridge.ts: scope child lifecycle per process (stopped-children set +
  request generation tags) instead of an instance-wide shuttingDown flag
  that the next spawn reset. A stale SIGTERMed child's late close event
  could record a phantom crash (5 of which permanently disable the
  dispatcher), null out the freshly spawned child, and reject the new
  child's pending requests. The request-timeout kill path still counts
  as a real crash.
- install_feature.py: the pre-write disk re-check measured ai_dir's
  filesystem even when budgeting the cross-filesystem copy that lands on
  the venv's disk; now each budget is checked against the filesystem the
  bytes actually land on, so ENOSPC cannot strike mid-write and leave
  site-packages half overwritten.

Behavior regressions:

- embed-subtitles: preserve pre-existing subtitle tracks (0:s?) and MKV
  attachments (0:t?) that the -map 0:v:0/0:a? rewrite silently dropped;
  data streams stay unmapped on purpose (the actual MPEG remux fix). The
  new subtitle maps first so the language tag hits the right stream.
- usage-survey-overlay: fail closed when the settings fetch fails; the
  fail-open path rendered the blocking survey against an unhealthy API
  and soft-locked admins, the lock-out class #392 fixed.
- features-store: queued bundles poll instead of each holding an SSE
  connection (Install All could pin 7 EventSources and exhaust the
  browser's 6-per-origin HTTP/1.1 limit, hanging the whole app);
  listenToProgress closes any prior stream and stops any poll before
  subscribing; installAll skips bundles already installing or queued.

Contracts, tests, i18n:

- openapi.yaml: add "queued" to the features status enum and document
  downloadBytes/installedBytes (Schemathesis conformance).
- feature-lifecycle e2e: queue transcription (~0.5 GB) instead of ocr
  (~6 GB) and give the test a budget that covers both install drains
  (the stacked waits exceeded the old 900s timeout).
- docker-compose.qa.yml: parameterize the host port (QA_APP_PORT) so
  QA_PROJECT_NAME concurrent stacks can actually bind.
- compare + watermark-image: restore per-input error attribution
  ("Invalid first/second image", "Invalid watermark image") lost in the
  shared-handler migration.
- ai-features-section: the "{size} on disk" suffix now goes through
  i18n; key added to all 21 locales.
- watermark-image + content-aware-resize: migrate to the shared
  inputHandlerFor("image") chain like compare/vectorize/compose, fixing
  drift in the inline copies (no SVG sanitize, no RAW extension hint,
  no AVIF probe).

Verified: typecheck across 9 workspaces, Biome clean on all changed
files, 584 targeted unit tests and 249 integration tests green
(including real-ffmpeg embed-subtitles runs). One unit test updated to
the new poll-while-queued contract with a single-EventSource assertion.

Claude-Session: https://claude.ai/code/session_017mR1HiHaf3a1BmUtrHX4j3

packages/ai/python/install_feature.py

@@ -51,7 +51,7 @@ def detect_arch() -> str:
     Only two archive variants are currently published to the bundle repo:
     'amd64-gpu' and 'arm64-cpu' (see deepsafe/feature-bundles). There is no
     CPU-only amd64 variant yet, so amd64 hosts always resolve to 'amd64-gpu'
-    even when no GPU is present -- this downloads working CUDA-capable
+    even when no GPU is present: this downloads working CUDA-capable
     packages, just larger than a CPU-only host strictly needs. Do not change
     this to branch on GPU presence without first publishing an 'amd64-cpu'
     archive for every bundle; requesting a key that doesn't exist in the
@@ -97,7 +97,7 @@ def estimate_extracted(compressed: int, extracted: int) -> int:
     extractedSize (0), the budget would otherwise collapse to just the
     compressed size and under-reserve for the extracted payload; fall back to a
     conservative 3x of compressed (measured extracted/compressed ratios reach
-    ~3x). This is only the early sanity bail -- the accurate guard is the
+    ~3x). This is only the early sanity bail; the accurate guard is the
     real-on-disk re-check just before the destructive venv write."""
     return extracted if extracted > 0 else compressed * 3
 
@@ -499,22 +499,23 @@ def main() -> None:
     # -- Disk re-check before the first destructive venv write --
     # The upfront check ran before the download and used an estimate; now the
     # payload is really on disk, so measure it and verify there's room to place
-    # it before we start writing into the venv. On the same filesystem the move
-    # is a rename (no extra space needed, just a safety floor); across
-    # filesystems it is a copy that transiently needs the payload's size again.
-    # Running here (after the local/remote branches merge) also covers the
-    # offline-import path, which skipped the upfront check entirely.
-    staging_real = dir_size(staging_dir)
-    if same_filesystem(staging_dir, venv_path):
-        recheck_needed = 1024 ** 3  # 1 GB floor for fixups / installed.json / slack
-    else:
-        recheck_needed = staging_real + 1024 ** 3
-    check_disk_space(ai_dir, recheck_needed)
+    # it before we start writing into the venv. Running here (after the
+    # local/remote branches merge) also covers the offline-import path, which
+    # skipped the upfront check entirely. Each budget is checked against the
+    # filesystem the bytes actually land on: when the venv lives on a
+    # different filesystem than staging, the site-packages payload is COPIED
+    # onto the venv's disk, so that disk (not ai_dir's) must hold it. Models
+    # stay under ai_dir either way, moving by rename.
+    disk_floor = 1024 ** 3  # 1 GB for fixups / installed.json / slack
+    staging_sp = os.path.join(staging_dir, "site-packages")
+    if not same_filesystem(staging_dir, venv_path):
+        sp_bytes = dir_size(staging_sp) if os.path.isdir(staging_sp) else 0
+        check_disk_space(venv_path, sp_bytes + disk_floor)
+    check_disk_space(ai_dir, disk_floor)
 
     # -- Move site-packages --
     emit_progress(92, "Installing packages...")
     site_packages_dir = get_site_packages_dir(venv_path)
-    staging_sp = os.path.join(staging_dir, "site-packages")
 
     try:
         if os.path.isdir(staging_sp) and site_packages_dir: