app-server: add app/installed runtime snapshot API

[mzeng-openai]

Summary

• add the feature-gated v2 app/installed API backed by the committed connector runtime snapshot
• project tool-derived connector identities through effective global, workspace, local, managed, MCP-tool, and app-tool policy at read time
• keep app/list behavior unchanged and add the M4 installed-read telemetry seam

This is PR 4 of the connector runtime-state stack and is based on #31476.

Public contract

app/installed { threadId?: string, reload: boolean }
  -> { apps: Array<{ id: string, enabled: boolean, callable: boolean }> }

reload is required. reload: false performs a committed-snapshot/cache-only read. reload: true refreshes workspace policy and, when Apps are permitted, performs exactly one uncached hosted-connector refresh under the shared lock before returning the committed result.

The response deliberately omits Tool[] and all canonical connector metadata. IDs are sorted, deduplicated, non-empty connector IDs derived from non-synthetic /ps/mcp tools/list metadata. Consequently, zero-tool connectors are absent. This deliberately narrows the RFC draft and does not consume _meta.installedApps.

Canonical names, descriptions, branding, labels, and install URLs remain an M3 app/read / ConnectorMetadataStore dependency once that lands. Tool-level connector names and descriptions remain compact runtime/search metadata only.

Behavior and ownership

• ConnectorRuntimeManager remains the owner of persisted/committed raw runtime snapshots and adds a non-activating peek used only for failure telemetry.
• McpConnectionManager remains the owner of generic MCP filtering and publication; the read-time callable projection reuses its effective MCP tool-filter semantics.
• local and managed configuration is never persisted in the snapshot.
• disabled global/workspace policy skips the network tools refresh, retains prior identities, and projects them disabled/non-callable.
• failed refreshes return an error and retain the prior committed snapshot.
• the method is disabled by default behind apps_runtime_state_refactor; legacy app/list remains available with its existing response and forceRefetch behavior.

Telemetry

• new-path installed duration with reload, refresh disposition, outcome, and retained-previous labels
• installed response bytes, connector count, raw snapshot tool count, and snapshot age
• distinct workspace-policy error/disable dispositions

The strict refresh trigger/outcome metrics are emitted by the preceding refresh-manager PR. This does not duplicate turn-start availability analytics.

Testing

• just test -p codex-app-server-protocol (253 passed)
• just test -p codex-mcp (118 passed)
• just test -p codex-app-server app_installed (10 passed)
• just test -p codex-core callable_app_ids_apply_effective_mcp_tool_filter
• focused workspace-settings and effective app-enable tests were also run
• just fix for codex-mcp, codex-core, codex-app-server-protocol, codex-chatgpt, and codex-app-server
• just fmt

Coverage includes feature gating, reload false/true, empty and zero-tool results, synthetic identities, MCP/app policy, thread config, concurrent serialization, failure retention, context changes, atomic publication, and legacy app/list compatibility.