XERK-538: show 'QA Review' status while a QA/QA-delta subagent runs

[xerhab]

What

On the Sessions page (and the Dashboard host card), a session that is working because of a background QA pass now reads "QA Review" instead of the generic "N background agents".

• A qa subagent (first QA pass) or qa-delta (every pass after) in the session's live-agent list flips the card's status text to "QA Review".
• The session stays in the Active category — it is still working, so it does not move to Ready-for-review or Idle, and it fires no ready-for-review alert. Only the displayed label changes.
• A QA agent wins even when an ordinary agent (e.g. Explore) is running alongside it — a QA pass is the thing the operator is waiting on.

Why

XERK-538. When QA is running, the operator's own change is being adversarially exercised — a distinct thing to watch for versus an ordinary fan-out. The bare agent count didn't say that.

How it works

Detection keys on the live agent's type (the subagent_type), which the transcript scan already carries onto session.agents[].type (_scan_agent_entry in hub-agent.py) and the hub preserves through sanitizeLiveAgents (server.js). No new wire field, no change to the "Working"/ready-for-review determination.

Label mirrors updated:

• turma/public/sessions.html — QA_AGENT_TYPES + qaReviewing(s) helper; agentWorkLabel(s).
• turma/public/index.html — same helper; dashboard liveState background-agent branch.
• android/.../ui/CommonUi.kt — liveStateLabel(state, live) (Web ⇄ Android parity).

Glasses is unaffected — it renders a working glyph (!) with no text status label, so a QA session already shows correctly as working; there is no count text to mirror.

Deliberately out of scope (noted so it's a documented decision, not an oversight): chat.js's in-chat "Background agents…" working bar (shown inside an opened session's chat view) is a separate element and still reads the generic wording. XERK-538 targets the session card + dashboard host card; extending it to the chat bar can be a follow-up if wanted.

How verified

• cd turma && node --test → 1702 pass, 0 fail, including new QA-Review cases in sessions.test.js and dashboard-livestate.test.js (real jsdom DOM render asserting the card innerHTML), plus retargeted count/pluralization tests.
• Android: SessionsTest (Robolectric/JVM) with a new liveStateLabel reads QA Review for a QA agent case → BUILD SUCCESSFUL, 0 failures.
• Adversarial QA agent: PASS. It drove the real render code for all six claims, traced the subagent_type → type → sanitizeLiveAgents seam, and mutation-tested both platforms: neutering qaReviewing fails 3 node tests; neutering the Android QA branch fails the Kotlin test. No defects found. Not directly verified: on-device Android rendering (no emulator; Robolectric + pure-function only) — the standing PARTIAL basis for every Android UI change.