feat(tui): capability invocation telemetry on Performance screen

[flg77]

Summary

Wires the `TASK_COMPLETE.invocations` field PR-B added all the way to the operator's eye: the NATSObserver folds each entry into the CollectiveSnapshot, and the Performance screen renders two new panels — running per-(kind, target) totals + a recent-failures tail — so operators see which skills + MCP tools are firing, at what success rate, and what the latest failure was.

Architecture

```
Agent ── publishes ─▶ TASK_COMPLETE { …, invocations: [{kind,target,ok,error}, …] }
│
▼
NATSObserver._route_task_complete
│
▼
CollectiveSnapshot.record_invocation
├── capability_stats[k]: {total, ok, last_error, last_seen_ts}
└── invocation_log: FIFO tail (cap 50)
│
▼
PerformanceScreen.watch_snapshot
├── _render_capability_invocations → DataTable (sorted by total desc)
└── _render_capability_failures → Static (last 10 failures)
```

Files

File	Change
`acc/tui/models.py`	New `CapabilityInvocationStats` dataclass + `CollectiveSnapshot.capability_stats` dict + `invocation_log` FIFO + `record_invocation()` helper.
`acc/tui/client.py`	`NATSObserver._route_task_complete` now iterates `data["invocations"]` and folds each entry. Pre-PR-B agents (no `invocations` field) are a no-op.
`acc/tui/screens/performance.py`	Two new panels in the right column: `CAPABILITY INVOCATIONS` DataTable + `RECENT FAILURES` Static.
`acc/tui/help/performance.md`	Two new §sections describing the panels + colour legend.
`tests/test_capability_telemetry.py` (new)	13 cases.

Visuals

```
CAPABILITY INVOCATIONS (skill / MCP tool)
┌──────────┬───────────────────────┬───────┬──────┬──────────────────────────┐
│ Kind │ Target │ Total │ OK% │ Last error │
├──────────┼───────────────────────┼───────┼──────┼──────────────────────────┤
│ mcp │ echo_server.echo │ 142 │ 100% │ — │
│ skill │ echo │ 87 │ 98% │ schema mismatch │
│ mcp │ fs.read │ 23 │ 78% │ A-018 blocked │
└──────────┴───────────────────────┴───────┴──────┴──────────────────────────┘

RECENT FAILURES (latest 10)
14:32:18 mcp:fs.read coding-aabb
A-018 blocked tool 'fs.read'@'fs_server': mcp_server 'fs_server' not in role.allowed_mcps
14:31:55 skill:echo coding-aabb
schema mismatch: 'text' must be string
```

Tests (13 new, all green)

Group	Cases
`CapabilityInvocationStats`	initial state, fail + ok_rate
`CollectiveSnapshot.record_invocation`	first call, accumulation, kind:target keying, malformed drops, FIFO cap
`NATSObserver._route_task_complete`	folds invocations, missing field is no-op, non-list field is safe
`PerformanceScreen` Pilot	sort-by-total-desc, failures-only panel, empty-state guidance

Combined regression: 166 existing tests pass unchanged (config + role-store + guardrails + compliance + observer + prompt-channel + ecosystem + file-picker; TestEd25519Validation deselected for lighthouse OpenSSL platform limit).

Test plan

• `pytest tests/test_capability_telemetry.py -v --no-cov` → 13 passed
• `pytest tests/test_config.py tests/test_role_store.py tests/test_guardrails.py tests/test_compliance.py tests/test_observer_task_listener.py tests/test_prompt_channel.py tests/test_ecosystem_screen_pilot.py tests/test_file_picker_pilot.py --deselect tests/test_role_store.py::TestEd25519Validation --no-cov -q` → 166 passed
• Manual: `acc-tui` → press `5` → Performance. Submit a few tasks via the prompt pane (or have an agent emit a few `[SKILL: echo {...}]` markers) → confirm the CAPABILITY INVOCATIONS table populates with rows, OK% colours match the 95/80 thresholds.

Out of scope (planned follow-ups)

• TASK_PROGRESS streaming for prompt channels.
• Telegram + WhatsApp PromptChannel adapters.
• Real echo MCP server image + compose entry.