feat(desktop): name the computer an agent lives on — core (device identity, split from #6259)

[mfethe1]

Supersedes the conflict-free backend half of #6259 ("name the computer an agent lives on"). The mention-UI half of #6259 conflicts semantically with #6337/#6401 (agentProvenance) and needs a product decision, so this PR carries only the parts that merge cleanly onto current main.

What this adds

• device_identity.rs (+551): opaque 32-hex device_id minted per install, user-editable device_label (validated: non-empty, ≤32 chars, control/format chars rejected), stable across restarts, persisted atomically
• Agent event projection stamps device_id/device_label on kind:30177 records for local backends only — a provider-backed agent publishes device: null (it does not run on the owner's computer)
• RelayAgentInfo gains optional device_id/device_label (#[serde(default)]), projected from the directory
• DeviceNameSettingsCard UI (Settings → Agents), mounted before PreventSleepSettingsCard
• listRelayAgents moves from tauri.ts to tauriRelayAgents.ts with the new fields

Why split from #6259

1. origin/main moved 38+ commits since feat(desktop): name the computer an agent lives on #6259's last rebase; all three textual conflicts are mention-UI files (MentionAutocomplete.tsx, useMentionSendFlow.ts, AgentsSettingsPanel.tsx) where feat(desktop-messages): keep agents addressed across messages #6315/fix(desktop): distinguish duplicate agent devices #6337/fix(desktop): simplify duplicate agent provenance #6401 landed a competing disambiguator. Resolving them textually would ship two disambiguators in one suggestion row.
2. This branch is a fresh three-way integration, not a cherry-pick. Files were taken from a git merge-tree of feat(desktop): name the computer an agent lives on #6259 × origin/main (preserving upstream's 0.5.17→0.5.18 bump and custom_harness_dir in lib.rs), with the panel wiring applied semantically to current main rather than copying feat(desktop): name the computer an agent lives on #6259's stale whole-file version.
3. Fixes a latent defect in feat(desktop): name the computer an agent lives on #6259 itself: three test-only RelayAgentInfo literals in relay_directory.rs were never updated for the new fields. #[serde(default)] does not fill struct literals, so cargo check --tests fails on feat(desktop): name the computer an agent lives on #6259's head 4e0764d4b (verified: exit 101, 3× E0063). This branch adds device_id: None, device_label: None to those three sites (detailed on feat(desktop): name the computer an agent lives on #6259: feat(desktop): name the computer an agent lives on #6259 (comment)).

Verification (all on this branch's head, macOS 15.6 arm64, rustc 1.95.0)

Gate	Result
cargo fmt --all -- --check	pass
cargo check --locked --tests (tauri)	pass, 0 errors
device_identity:: suite	16/16
managed_agents::agent_events:: suite	16/16
tsc --noEmit (desktop)	pass, 0 errors
agentDeviceLabel.test.mjs	4/4

Negative control on the core invariant: flipping only the BackendKind::Provider arm of the device stamp to current() fails exactly one test — projection_omits_the_device_for_a_provider_backend ("a remote body must not be given this computer's id") — with 15/16 passing. Restored, 16/16. So the local-vs-provider gate is behaviorally pinned, not just compiled.

Independently reproduced by @ScaleLeanChris on #6259 (macOS): agentDeviceLabel.test.mjs 4/4, focused Tauri device-identity tests 13/13, TS checking clean — the files exercised there are unchanged by the rebase.

Scope / limitations

• Desktop only; no mobile changes.
• The mention-UI disambiguation (device label in @ autocomplete) intentionally remains unlanded pending reconciliation with agentProvenance (fix(desktop): distinguish duplicate agent devices #6337/fix(desktop): simplify duplicate agent provenance #6401).
• Future work noted in-code: a provider-backed agent still has a custody device (the install holding its secret) distinct from its execution location — distinguishing them needs a protocol change, out of scope here.

[mfethe1]

DCO repaired tree-identically at 6c46c272b (only the required Signed-off-by trailer changed; tree hash still matches the previously reviewed implementation). The DCO check now passes and GitHub reports the PR mergeable; review is the remaining gate.

This core is also the dependency for a follow-up owner-attested agent-directory capability/model surface (REG-4), which I will propose separately after this lands to keep each review focused.

[mfethe1]

DCO is repaired at 6c46c272b. The fix touched only the Signed-off-by trailer — the implementation tree is byte-identical to the previously reviewed version, so no re-review of the code itself should be needed. DCO Check is now passing and there are no failing checks on the head.

Ready for maintainer review whenever there's bandwidth; happy to rebase or address requested changes.

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is ac5a18697c8294e9237f505a2e01ec6fc374849a...8a8ad8c7d58ab19046b77fce0ec565cac6a405d6.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review 8a8ad8c7d58ab19046b77fce0ec565cac6a405d6 to authorize a new review.
Any previous review applies only to its recorded range.

[mfethe1]

Rebased onto current main (0dbd036f5). One conflict in managed_agents/types.rs: this PR extracts the harness DTOs to harness_catalog_types.rs, while #4625 added effort_canonical_values + effort doc updates to their old location. Resolution keeps the extraction and ports upstream's field/docs into the moved file — no behavioral change to this PR's feature. DCO and authorship preserved; desktop cargo test --lib 3146 passed / 0 failed, pnpm typecheck and all check:* guards green.

[mfethe1]

Rebased onto current main (ac5a186). Conflict resolution mirrors #6259: kept the type extraction into relayDirectoryTypes.ts and ported upstream's status: "unknown" union member into the extracted RelayAgent. Local gates: typecheck clean, 5968 tests pass / 0 fail. Authored locally; DCO preserved.