CL-6469: agents list to spec

[TheGreatAxios]

Summary

Brings apps/web/src/pages/agents-page.tsx (Agents list only) up to the mock spec's §5.1:

• Table columns: checkbox · Agent (name/slug two-line cell) · Description (.opt → hidden lg:table-cell) · Status · Model (.opt) · Runs · 7d (.opt, numeric)
• Bulk row selection via @corbits/react-ui's useListSelection/SelectionCheckbox/BulkActionBar (PR CL-5849: Return 409 not 500 when an agent can't resolve an inference source #33) — shift/cmd range select, header select-all/indeterminate, floating bulk bar with Archive only
• Top-bar page action renamed to primary "New agent" (#Plus), per §2.3's page-action inventory (was "Create")

Update after 2nd peer review (both blocking items fixed)

1. Partial-failure bug in bulk archive fixed. onArchiveSelected used Promise.all, which rejects on the first failure even though the other archive calls already succeeded server-side — a mid-batch failure left the roster showing stale state and the toast implied nothing happened. Replaced with archiveDefinitions (Promise.allSettled) + archiveResultToast, both exported and unit-tested: the definitions/runs queries are now invalidated regardless of outcome, and the toast reports an honest count ("Archived 4 of 5 — the rest failed" / all-succeeded / all-failed).
2. prettier --check now passes on the touched files (bunx prettier --write run on agents-page.tsx and agents-page.test.tsx).
3. Dropped the now-unused Trash export from packages/icons (was left over from before Delete was removed from the bulk bar).

Update after 1st peer review

• Bulk Archive now really works. onArchiveSelected calls the real setAgentDefinitionStatus(tenantId, id, "stopped") — the same PUT route the single-agent Archive button on agent-detail-page.tsx already uses. Duplicate/Move/Delete are dropped from the bulk bar entirely: batch duplication would need slug-collision handling the detail page's single-agent duplicate never had to solve, and Move/Delete have no backend primitive at all.
• Status tones fixed. Adopted @corbits/react-ui's own RUN_STATUS_TONE convention: Running → info (blue) with a pulsing StatusDot (live prop), Idle → success (green). Blocked (danger)/Archived (neutral) were already correct.

Real-data notes

• Status is derived from real fields: a stopped definition always reads Archived; a deployed one reads Running if any of its top-level runs (listTopLevelRuns) is running, Blocked if any is error, else Idle — grounded in WorkflowDefinitionResponse.status + WorkflowRunResponse.status, not invented states.
• Model is fetched lazily per row via getAgentCapabilities, the same plain fetch-effect pattern (not react-query) the existing detail rail already uses — a load/fetch failure degrades that row's cell to — instead of blocking the table.
• Runs · 7d counts a definition's top-level runs with createdAt inside the trailing 7 days.
• Kept the existing "Workbenches" side-panel summary (on row selection) untouched — not one of the spec's table columns, but real, tested functionality this task wasn't asked to remove.

Known follow-up (not blocking, flagged for a ticket)

AgentModelCell fires one uncached getAgentCapabilities fetch per row on mount via a plain useEffect — on a real-size roster every render refires every row's request (no batching, no cache). Needs either a batch capabilities endpoint or a react-query cache keyed by (tenantId, definitionId) before this scales past a handful of agents.

Test plan

• bunx prettier --check on touched files — clean
• bunx eslint on touched files — clean
• bun run typecheck (apps/web) — clean
• bun run build (apps/web) — clean
• bun test ./src ./test (apps/web) — 786 pass, 0 fail (full existing suite, no regressions)
• bun test (packages/icons) — 3 pass, 0 fail
• New tests: archiveDefinitions/archiveResultToast unit tests covering partial/full success/failure, plus the earlier agentRosterStatus/runsInLast7Days unit tests, populated/empty render, Running's live-dot rendering, selection checkboxes + header select-all, no bulk-bar leakage at zero selection, "New agent" top-bar action, workbench-link side panel still works
• NOT verified: no live-stack/browser check was done (per instructions — owner is testing live on :3000). Bulk-bar Archive's real network round-trip (including an actual partial-failure scenario against the live hub), the Model column's real network latency/skeleton timing, and Status against real multi-instance data are unverified beyond render/unit-test level.
• NOT verified: agent detail page (agent-detail-page.tsx) — out of scope, untouched, owned by another unit.

Do not merge — flagging for peer review per task instructions.