Polish community shared compute and agent setup

[sandro-sq]

Summary

Stacked follow-up to #580. Base: jimmy/mesh-share-wip, not main; this PR proposes merging the polish into #580's branch. It does not merge or mark #580 ready.

• Polish sharing status, model-fit options, community device inventory, names and model labels.
• Add community-agent examples that prefill the existing owner-reviewed creation dialog; no automatic creation. Media examples assume an authenticated buzz CLI and host ffmpeg.
• Exclude this machine's lingering serve advertisement from peer join tokens, and permit Share On to replace a connecting consumer without replacing a starting server.
• Publish chip/GPU label and rated memory while serving; do not publish hostnames or asset tags. These hardware facts are visible to community members.
• Add regression coverage for the affected native and frontend contracts.

Screenshot

User-supplied screenshot of community agent examples and the community mesh inventory. Visual evidence only, not proof of inference or restart acceptance.

Validation at 8d45e94ea9123259f53d1b6131ef16f235f96f81

Local macOS validation on October 6, before publication:

• Formatting, lint, TypeScript, design-system types/guards, frontend build and diff checks passed.
• Node integration: 192 passed.
• Full Vitest: 7,842 passed across 548 files.
• Default workspace/all-target Clippy and Mesh-enabled workspace/all-target Clippy passed with -D warnings.
• Default Rust workspace tests/doctests passed.
• Mesh-enabled full workspace tests/doctests passed with serial test execution (--test-threads=1), including 317 native tests and 55 Mesh tests; existing ignored tests remain ignored.
• Mandatory repository staged-file and pre-push groups passed. These groups were invoked explicitly because the installed corporate hook manager does not automatically invoke the repository's custom groups; corporate hooks were preserved and the actual push passed its organization check.
• Independent agent review found no concrete blockers (9/10 assessment). This is not a security audit or live inference acceptance.

Known local failures and acceptance gaps

Not an all-green validation claim. No assertions, timeouts, retries or error allowlists were weakened to obtain passes.

• Two parallel Mesh workspace runs failed on different unchanged host_command process-start tests: passes_exact_args_without_shell_interpretation and aborting_the_read_kills_descendants. The first passed in isolation; the complete serial workspace run passed. Concurrency/scheduling remains a hypothesis, not a confirmed fix.
• Full browser run: 5 passed, 1 failed, 1,273 did not run because measurement dependencies blocked subsequent projects. Failure: Chromium cursor-paging/virtualization scroll test, with ERR_NETWORK_IO_SUSPENDED resource errors.
• Focused Mesh/agent-model Chromium + WebKit run with measurement dependencies excluded: 6 passed, 2 failed. Chromium agent-model fixture navigation timed out; Chromium sharing fixture navigation/dependency optimization failed with Rolldown Invalid argument (os error 22) while writing a temporary dependency file. This does not establish fresh successful Chromium sharing behavior.
• Design-system browser run: 77 passed, 5 failed (Chromium status-badge geometry; toast recovery and message-gallery fullscreen in both engines). These failures are outside edited behavior, but have not been proven baseline-only or repaired.
• Hosted CI run 37555118320 passed at this head: both JavaScript shards, Rust/tool integration, native browser fixture, browser measurements and all 12 Chromium/WebKit journey shards; CI required, DCO, Semgrep and zizmor passed. Windows native validation was skipped (manual-dispatch-only). This establishes hosted automatic CI success, not resolution of local failures or Mesh-enabled/live-native acceptance.
• Fresh live sharing/inference, agent reply, quit/relaunch persistence acceptance and cross-platform packaging were not exercised in this validation pass. Human feedback is not represented as completion of the repository's formal acceptance checklist.

No browser cases were added or removed in this follow-up; existing journeys were attempted, and colocated tests cover the new state/prefill/grouping contracts. No performance claim is made. Opened for review at the author’s request after hosted CI passed. Remaining acceptance gaps are retained above; no review-completed attestation is included.