feat(v2): TASK-124 agent profile token parity

[lilyshen0722]

Summary

• align the agent-profile primary and footer CTAs with the ruled ink/on-ink token grammar
• replace the avatar hard outline with the shared halo focus treatment
• move the six profile 13px roles and three 850 weights onto the body/meta/label ladder
• make pod and memory rows stable transparent-border, fill-only hover rows

Scope

This PR is the agent-profile half of TASK-124. Lily's Your Team PR owns the shared v2.css team selectors; this branch only edits agents/v2-agent-profile.css and profile layout invariants.

Verification

• npx jest src/v2/__tests__/v2-layout-invariants.test.ts --runInBand --watchAll=false (107/107)
• npm run typecheck
• npm run build
• git diff --check

Real-browser 1440/390 gate remains with UX after the combined Phase B deploy.

[lilyshen0722]

Code gate PASSED at 41327d8ba94bec71b15d0d260f37b2af3322ba8d. Browser gate remains ux-lead's.

Method: detached checkout of the pushed head; each mutation reverted after; tree clean. Control 107/107.

The invariants discriminate

mutation	result
reintroduce a 13px under the profile prefix	1 red
revert the primary button to var(--v2-accent)	1 red
restore outline: 2px on the avatar focus	1 red
reintroduce an 850 weight	1 red
empty agents/v2-agent-profile.css entirely	3 red

That last row is the one I most wanted. The spec requires the invariant file to read the second sheet, and a guard that silently resolves nothing would pass every assertion above for the wrong reason — the exact vacuity that has bitten this stack repeatedly. Emptying the sheet reddens three tests, so the reader is real.

Independent check of the sheet itself

Counted directly rather than trusting the assertions: 13px 0, 850 0, outline: 2px 0, var(--v2-ink 4, v2-focus-ring 1. Those agree with what the invariants claim.

Scope note

This is the agent-profile half only — agents/v2-agent-profile.css plus its invariants, no v2.css changes. The stay-blue side of the ruling (selection states, dots, brand mark, progress bar) is not exercised by this diff, so nothing here can regress it; that assertion belongs with the surfaces that carry those selectors.