feat(ui): WP 7.0 admin UI overhaul + AI Connector support (closes #220)

[pluginslab]

Summary

• UI overhaul rebuilt on @wordpress/components primitives — Tabs/Cards/HStack/VStack across every surface (Settings, Abilities, Plugin Abilities, Chat, Composer, Model status). Custom SCSS down from ~3,600 → ~1,450 LOC (much of the remainder is token-aware overrides for radius flattening + chat bubbles).
• WP 7.0 AI Connectors wired end-to-end as a third provider option alongside Local (WebLLM) and Remote (OpenAI-compatible). New Connectors PHP class with GET /v1/connectors (list + model discovery via modelMetadataDirectory()) and POST /v1/connectors/chat/completions (forwards through AiClient::generateTextResult using structured Message / MessagePart DTOs). Persisted provider auto-loads on next page boot.
• Context-window auto-derive for remote/connector models via KNOWN_REMOTE_CONTEXT_SIZES longest-prefix table + new wpAgenticAdmin.contextWindow filter (@wordpress/hooks). The old per-provider context input is gone from the UI.
• Multi-select bundle picker in the composer (Dropdown + aria-checked selected state). Tools from every selected bundle are union'd into the filter; plugin namespaces are all scoped.
• Inline model status pill lives in the composer toolbar now (Spinner + percent + Cancel while loading, status dot + name + 3-dot menu while ready). The big status Card under the page chrome is gone.

Pre-PR review fixes

Ran a 3-angle code review (line-by-line, removed-behavior, cross-file tracer) → 1-vote verify, 10 findings. All addressed before this PR:

1. bundleId === 'content-create' regressed under multi-select → key off bundleIds.includes('content-create') and thread bundleIds[] through the orchestrator
2. pluginAbilitiesManager.scopeToPlugin only scoped first namespace → accept array, merge abilities
3. Connector silently dropped params.tools → warn once per session
4. PHP flattened messages demoted system role → AI Client PromptBuilder + structured Message DTOs (with usingSystemInstruction)
5. includes() false-matched gpt-4 vs gpt-4o → startsWith on the bare id (strips provider/ prefix), longest-prefix wins
6. Connector URL derived via brittle regex → use new connectorsRestUrl / restRoot localized fields
7. No size/rate caps on paid-provider proxy → MAX_MESSAGES=50 (413), MAX_TOTAL_CHARS=100k (413), RATE_LIMIT_PER_MINUTE=30 per user via transient (429)
8. Connector couldn't be aborted → forward params.signal to fetch
9. useModelStatus snapshot ignored in-flight load → read isLoading / loadProgress so mid-load mounts show the spinner immediately
10. Legacy agentic_admin_remote_context_size silently dropped → honour as fallback + one-shot deprecation log pointing to the filter

A follow-up commit also handles a runtime bug surfaced in manual testing: leading assistant turns (e.g. the welcome card) were sent as history; Anthropic rejects that. We now shift them off before building the prompt.

Test plan

• Local model: load Qwen 3 1.7B, send a few messages, confirm tool selection still works
• Remote model (Ollama / OpenAI-compatible): same
• Connector (Anthropic): pick a connector + model, send chat, confirm response
• Connector cancel: hit Stop Generation mid-flight, confirm the request aborts
• Connector limits: paste > 100k chars or > 50 messages → expect 413
• Multi-select bundles: pick 2+, send a message, confirm both bundles' tools are available
• Plugin abilities bundle: install a plugin that registers abilities, pick that bundle, confirm scoping works
• Settings → Context Window: confirm per-model context inputs still work; remote context auto-derives
• Fresh page mid-load: navigate away + back while a model is loading, confirm the pill shows the spinner immediately instead of "Not loaded"

🤖 Generated with Claude Code

[ivdimova]

Code review — 2 findings before merge:

[1] Rate limit TOCTOU race condition — includes/class-connectors.php:443–458

The check_rate_limit() method does get_transient → check >= 30 → set_transient( $count + 1 ). Two concurrent requests from the same user (multiple tabs, double-click, retry) can both read count = 29, both pass the check, and both increment independently — bypassing the cap. Since this rate limit exists specifically to protect against paid-API over-billing, the race window matters.

Suggested fix: set the transient with add_transient on first write (returns false if the key already exists), or wrap the read/increment in a short-lived mutex transient so concurrent requests queue rather than race.

---

[2] Silent exception swallow in list_connectors() — includes/class-connectors.php:~225

AiClient::defaultRegistry() is caught and sets $ai_registry = null on any exception, causing the endpoint to return wp_supports_connectors: true with an empty connector list. The UI would show the Connectors provider option but nothing selectable — no error surfaced to the operator or the user.

Suggested fix: add error_log( '[wp-agentic-admin/connectors] registry init failed: ' . $e->getMessage() ) in the catch block so operators can diagnose transient failures without changing the existing graceful-degradation behaviour.