feat(desktop): unified AgentDialog create entry point with create intents (Phase 1B.2)

[wesbillman]

Phase 1B.2 — unified AgentDialog create entry point

Second slice of Phase 1B (UI unification, #centralize-personas-and-agents). Routes both create entry points on the Agents surface through a single AgentDialog that owns the create intent:

Intent	What it creates	Form rendered
definition_start (default)	definition + linked, spawned instance (personaId) — today's quick-start	PersonaDialog (create mode) with "Start agent after create" toggle on
definition	definition only — new behavior, see below	same form, toggle off
instance	standalone keyed managed agent	CreateAgentDialog, unchanged

Stacked on #1624 (B1 shared sections). Wire shapes (CreatePersonaInput / CreateManagedAgentInput) untouched — frontend recomposition only; independent of #1623 (A2).

Composition: pass-through router, not a new form

AgentDialog (~110 lines) is an intent router. The definition family renders PersonaDialog itself in create mode (via a new 2-line createFooterSlot prop — footer-left is empty in create mode; edit's import button owns it otherwise); the instance intent renders CreateAgentDialog unchanged. A new standalone form would have duplicated ~400 lines that B3 deletes anyway; physical consolidation of the forms into one file is B3's job. Consequence: B1's shared sections are consumed transitively (both forms already use them).

Entry point implies intent — "New agent" opens the definition family, "Custom agent" and the open-create-agent event open the instance form. There is no in-dialog mode picker; don't read AgentDialog expecting one.

Sequencing stays hook-side

usePersonaActions.handleSubmit(input, intent?) keeps ownership of create-then-instance (it already holds the mutations, feedback surfaces, and SecretRevealDialog plumbing). The definition intent is the existing create branch minus the instance step; upstream (runtime guard, avatar resolve) and downstream (instance step, error taxonomy, dialog close, pending reset) are untouched. Return type widens to Promise<boolean> so the unified mount can close on success — legacy callers ignore the return.

definition_start preserves this hook's semantics (harnessOverride: true, hard-error on unavailable runtime, resolveManagedAgentAvatarUrl before create feeding both records) — not handleStartPersona's card-start mapping, which diverges (conditional harnessOverride, resolvePersonaRuntime fallback) and stays untouched on the card path. Converging the two definition→instance mappings is a real cleanup but belongs to a visible future slice, not this PR.

Intent defaults pin legacy behavior

resolveCreateIntent defaults to definition_start: PersonaDialog's duplicate path still calls handleSubmit without an intent until B3, and create/duplicate both always quick-start today. A unit test pins the default so duplicate can't silently become definition-only. The toggle-on dialog copy is derived from createPersonaDialogState and deepEqual-pinned against it — it cannot drift from the flow it replaces.

Divergence flags (per the B1 rule: each intent preserves the flow it replaces)

• lockedRuntimeReset stays per-form by construction: definition intents replace PersonaDialog-create → "full"; instance replaces CreateAgentDialog → "provider-only". The flag choice is a consequence of the replaces-rule, not a judgment call.
• Definition-only is genuinely new behavior — today's UI has no way to create a definition without starting an instance ("Create an agent profile and start its managed agent instance"). This is the Phase-1 "definition without instance" goal, disclosed as new rather than dressed as recomposition. Skip-step-2 in the hook; notice copy "Created X."
• Runtime-required guard kept as-is for definition-only: create mode still requires choosing an available runtime, even though a definition-only record arguably shouldn't need one. Relaxing it is a product question (below), not smuggled in.

Both-paths-alive (weaker form for one path — explicit)

• CreateAgentDialog: fully alive and reachable — it is the instance form, byte-identical.
• PersonaDialog create mode: in-tree but unreachable via UI (edit/duplicate/import mounts remain live). Rollback = one-line revert of the AgentsView mount wiring; openCreate is intact in the hook, unwired.
• Mount exclusivity: create no longer sets personaDialogState (prepareCreate split out of openCreate primes feedback/runtime-load without it), so the old and new PersonaDialog mounts — which share testids until B4's rename — can never coexist. Comment at the mount site names the assumption.

Testids

The definition form reuses PersonaDialog's existing testids/ids (persona-dialog, persona-dialog-submit, #persona-runtime, …), so the existing create-path e2e specs run unchanged — verified locally. Renames are B4's vocabulary flip. New: agent-dialog-start-toggle.

Open product questions for @wesbillman (none block this PR)

1. Does "Persona" survive as the catalog name? (B4)
2. Create-flow default: quick-start vs definition-first? (currently quick-start = today's de-facto behavior; one constant to flip)
3. Should definition-only creation still require choosing an available runtime? (guard kept as-is here)

Verification

• tsc, biome (full src/), file-size/px-text/pubkey checks clean
• 2061/0 unit tests (5 new: intent resolver default pin, toggle mapping, copy derivation pins)
• smoke e2e -g "agent|persona" 56/56 locally; full suite in CI
• Cross-reviewed by Pinky (scope pre-pass + line-by-line on the diff, this thread)