feat(capability): LLM parser + TUI Ecosystem live tables + how-to docs (Phase 4.4)

[flg77]

Summary

Closes the Phase 4 ladder. After this PR:

• The LLM can request capability invocations by emitting `[SKILL: id {json}]` / `[MCP: server.tool {json}]` markers in its response — the agent's task loop parses them, dispatches each through `CognitiveCore.invoke_skill` / `invoke_mcp_tool` (so Cat-A A-017/A-018 fire), and folds the outcomes into `TASK_COMPLETE`.
• The TUI Ecosystem screen replaces its two roadmap placeholders with live DataTables of every loaded skill + MCP server, sourced from the same registries the agent uses.
• Operators have two step-by-step how-to docs explaining how to add either kind of capability and what governance hooks fire.

What's in the box

`acc/capability_dispatch.py` (new)

Function	Role
`parse_invocations(text)`	Strict regex extraction of every marker. Order-preserving across `[SKILL:...]` and `[MCP:...]`. Single-line per marker, JSON args validated via stdlib `json`. Malformed args captured as `args_error` (not raised) so one bad marker doesn't drop neighbours.
`dispatch_invocations(invs, core, role)`	Runs each parsed marker through `CognitiveCore.invoke_skill` / `invoke_mcp_tool`. Every exception path funnels into `InvocationOutcome.error` — one bad marker never raises out of the dispatcher.

`acc/agent.py`

`_handle_task` now parses + dispatches after `process_task` returns (no-op when the task was blocked or output is empty), and folds the `(kind, target, ok, error)` tuples into the new `invocations` field of the `TASK_COMPLETE` payload. Full result dicts remain in the LanceDB episode (reachable by `episode_id`); the bus payload carries summary metadata only.

`acc/cognitive_core.py`

`build_system_prompt` now documents the marker grammar inside the "Available skills" / "Available MCP servers" blocks added in PR 4.3, so the LLM knows the exact shape the dispatcher will recognise.

`acc/tui/screens/ecosystem.py`

Two roadmap `Static` placeholders replaced with live DataTables:

Table	Columns
SKILLS	Skill / Version / Risk / Requires
MCP SERVERS	Server / Transport / Risk / Tools

Risk cells carry Rich colour markup (LOW=green, MEDIUM=yellow, HIGH=red, CRITICAL=bold red), matching the Compliance screen. Empty registries render a "no skills loaded — see howto-skills.md" guidance row instead of a blank table. Lazy imports keep the TUI startup time unchanged when `acc.skills` / `acc.mcp` are absent (e.g. in the minimal CLI image).

`docs/howto-skills.md` + `docs/howto-mcp.md` (new)

Six-step walkthroughs covering: directory layout → manifest → adapter (skills only) → role wiring → TUI verification → LLM marker invocation. Each includes the marker grammar, A-017 / A-018 enforcement detail, CRITICAL oversight semantics, and troubleshooting matrices.

Marker grammar

```
[SKILL: <skill_id> {}]
[SKILL: <skill_id>] # args default to {}

[MCP: <server_id>.<tool_name> {}]
[MCP: <server_id>.<tool_name>] # args default to {}
```

Single-line per marker. Tool names allow dots so nested namespacing (`fs.read`) is preserved. Malformed JSON is logged + skipped, not raised.

Smoke tests (all green locally)

• `parse_invocations` extracts 3 mixed (skill + MCP) markers in source order.
• Bad JSON: captured as `args_error`, not raised; marker not dispatched.
• `dispatch_invocations` happy path returns the adapter dict.
• `dispatch_invocations` denied path: A-017 block surfaces as `InvocationOutcome(ok=False, error='SkillForbiddenError: A-017 blocked ...')`.
• `build_system_prompt` includes the marker grammar examples.
• `EcosystemScreen` + helpers import cleanly; `_risk_cell` renders correct Rich markup for every level.

Existing suite: 139 unit tests (config + role-store + guardrails + compliance, with the lighthouse-OpenSSL-impaired `TestEd25519Validation` class deselected) pass unchanged.

Test plan

• `from acc.capability_dispatch import parse_invocations, dispatch_invocations` imports cleanly.
• `pytest tests/test_config.py tests/test_role_store.py tests/test_guardrails.py tests/test_compliance.py --deselect tests/test_role_store.py::TestEd25519Validation --no-cov -q` → 139 passed.
• Boot the TUI on a coding_agent collective and navigate to Ecosystem — confirm SKILLS table shows `echo` and MCP SERVERS table shows `echo_server`, both with green LOW risk cells.
• Walk through `docs/howto-skills.md` end-to-end — add a trivial `word_count` skill, confirm it appears in the table after agent restart.

Phase 4 delivery summary

PR	Subject
4.1	Skills foundation (#? merged)
4.2	MCP foundation (#6 merged)
4.3	Role config + Cat-A A-017/A-018 (#7 merged)
4.4	LLM parser + TUI live tables + docs (this PR)

After this lands, the cell can synthesise organelles in-house (Skills) and admit symbiotic bacteria (MCP servers), the membrane enforces both via Cat-A, and the operator can see both surfaces live from the TUI.