Text Worker v0.3.10

Text Worker v0.3.10

This release makes the worker practical to set up on a headless Linux server
and fixes vision detection for reasoning models. Job execution, charging, den
accounting, and payout settlement are unchanged from v0.3.9.

Operator Changes

• Terminal setup. grid-inference-worker --setup configures the worker
  without a browser. With --backend-url it runs without prompts: it reads the
  served model from the backend, proves it with a live completion, and saves
  the same configuration as the browser wizard. --concurrency, --grid-model,
  --backend-api-key, and --install-service combine with it. One connection
  without --api-key uses the secure Console approval link; parallel slots
  still need an advanced account key.
• Linux service install over SSH. --install-service now uses sudo in a
  terminal and the desktop pkexec prompt only when a desktop session exists.
  If neither works it prints the systemd unit and the commands to install it by
  hand, instead of failing with "Authentication cancelled".
• Backend URLs may be entered with or without a trailing /v1 in the
  wizard, command-line flags, and detection.
• Vision detection reads both reasoning and reasoning_content, gives
  reasoning models room to answer, asks templates that support it to skip
  thinking, and runs once per backend for all parallel connections, so one
  model can no longer advertise mixed text-only and vision capability.
• LM Studio context limits come from the loaded model instance rather than
  the catalog entry (#39).
• Taken dashboard port. In console mode the worker now exits with a clear
  error and a suggested --port when port 7861 is already used (for example by
  a second worker on the same machine), instead of stopping silently. The
  desktop window still opens the running dashboard. Output is line-buffered so
  setup messages print in order under pipes and journald (#42).

Release Qualification

• Source: main after #42, tagged v0.3.10. CI ran the Python suite on
  3.11–3.13 and the browser-script tests, built all four platforms, and
  assembled verified checksums, manifest, SBOM, and provenance.
• An operator ran the release source on a 2× RTX PRO 6000 host with vLLM
  serving a reasoning model at 256K context and 8 parallel slots. Before these
  fixes it was serving real Grid jobs as a systemd service; that run surfaced
  the defects fixed here. With the release source, prompt-free --setup
  detected vLLM, validated the model with a live completion and saved the
  configuration; a second worker on the same host exited with the new port
  error.
• Not covered: the frozen release binaries themselves were not separately
  re-run on that host before publication, Console-approval setup was not
  repeated, and Windows/macOS have CI build/runtime checks only.

Platform Trust

• Linux x64 and ARM64 are the only platforms exposed by the public /run
  release gate after the complete release envelope verifies.
• Windows is unsigned and macOS is not notarized. Their manifest state and
  installation warnings remain explicit; signing is recommended but not a
  publication blocker.
• Windows may show a SmartScreen warning. macOS Gatekeeper may require an
  explicit open/allow action. Verify the published checksum and provenance
  before running either download; do not disable OS protections globally.