feat(#1943): extensionEvents vocabulary — separate from hookEvents (extension-system surface) (#1946)

* feat(#1943): extensionEvents vocabulary — separate from hookEvents (extension-system surface)

* fix(#1943): re-export VALID_EXTENSION_EVENTS from gen-capability-registry (test import path)

* fix(#1943): regenerate capability-registry.cjs + add changeset fragment

* fix(#1943): import VALID_EXTENSION_EVENTS from validator, not gen-capability-registry (golden parity)

docs/adr/1239-gsd-embeddable-orchestration-engine.md

@@ -93,7 +93,8 @@ Phase A is **implemented** and Phases B–E have landed, so this ADR is **Accept
 - **`PROTOCOL_VERSION`** is an integer starting at `1`, **distinct** from the package `version` / `engines.gsd` semver (the `version`/`protocolVersion` overlap, resolved).
 - **`negotiateHostCapabilities(host, engine?)`** performs the in-process `initialize` exchange and enforces the trust-boundary invariant `effective ⊆ host-declared ∩ engine-known`: an undeclared axis or an unknown / higher-`protocolVersion` value is **never** trusted — it degrades to the most-restrictive known value (fail-closed), never throws.
 - **`degradationFor`** is the typed Full/Degraded/Absent ladder table; **`profileOf` + `PROFILE_BASELINES`** classify each descriptor into `programmatic-cli` (9 hosts: claude, opencode, cursor, cline, hermes, qwen, kilo, trae, kimi), `declarative-cli` (7 hosts: codex, gemini, antigravity, augment, codebuddy, copilot, windsurf), or `ide` (defined as a baseline; no installed host yet — VS Code lands in Phase D).
-- **Overlap resolutions (explicit):** `commandStyle` (GSD emission style, retained) ⊥ `commandSurface` (host surface type); `hookEvents` dialect ⊥ `hookBus` ownership (a host with `hooksSurface:none` may still be `hookBus:host` — e.g. opencode); the `opencode-subset` `hookEvents` value remains reserved for the Phase D OpenCode hook-dialect consumer; `runtimeCompat` (feature→host) stays an independent override, orthogonal to these runtime→engine axes.
+- **Overlap resolutions (explicit):** `commandStyle` (GSD emission style, retained) ⊥ `commandSurface` (host surface type); `hookEvents` dialect ⊥ `hookBus` ownership (a host with `hooksSurface:none` may still be `hookBus:host` — e.g. opencode); `runtimeCompat` (feature→host) stays an independent override, orthogonal to these runtime→engine axes.
+- **`extensionEvents` vocabulary (amendment, #1943).** The OpenCode extension-system event subset is a SEPARATE descriptor field + closed vocabulary, **not** a `hookEvents` value. `hookEvents` is the *managed-hook* dialect only (`claude`/`gemini`) — the event names GSD writes into a declarative host's settings.json. `extensionEvents` is the *plugin/extension-system* event surface an imperative host exposes: `{ opencode, pi, none }` (OpenCode ~25 plugin events; pi ~30 fine-grained events; `none` = the host exposes no extension surface and the engine owns the bus, e.g. VS Code). The former `opencode-subset` `hookEvents` value was this concept misfiled; it is now `extensionEvents: opencode`. Keeping them separate preserves the `hooksSurface:"none" ⇔ no-hookEvents` invariant (OpenCode declares `extensionEvents`, not `hookEvents`). Resolved by `extensionEventSurfaceFor` in `src/host-integration.cts`, validated by `VALID_EXTENSION_EVENTS` in `capability-validator.cjs`.
 
 **Every per-host axis value is documentation-sourced, with citations.** Each of the 8 axes for all 16 installed CLIs was determined from that CLI's authoritative documentation (Context7 + the official dev docs/source), never inferred. The full per-CLI, per-axis matrix — value, source, and an evidence quote — is recorded in [`docs/reference/host-integration-capability-matrix.md`](reference/host-integration-capability-matrix.md), the deployment source-of-truth that Phases B–E build on. Where a CLI's docs genuinely do not state an axis, the descriptor carries the explicit `undocumented` sentinel (which `negotiateHostCapabilities` fail-closes on) rather than a guessed value — 22 such markers exist today, each with its search trail in the matrix. Two findings corrected this ADR's original appendix matrix: (1) current OpenAI **Codex** docs document slash-commands, so its `commandSurface` is `slash-file`, not `prose-only`; (2) several hosts run non-Node runtimes (opencode & kilo on **bun**; hermes & kimi on **python**; antigravity on **go**), so the `runtime` axis vocabulary was widened to `node|bun|sandboxed-web|python|go|rust|electron|other`. The documented `embeddingMode` split (9 imperative / 7 declarative, above) likewise reflects each CLI's real plugin/extension API, not a profile assumption.
 
@@ -120,7 +121,7 @@ A plugin is a JS/TS module exporting an `async` function that returns a **hooks
 | 1 Command | slash-file commands projected to the xdg command dir (`gsd:`-namespaced); plugin may also surface entrypoints as custom `tool()`s and drive `tui.command.execute` | `commandSurface: slash-file` | none (full) |
 | 2 Dispatch | `mode: subagent` / `@`-mention; `subtask` is **synchronous-only** | `dispatch: { namedDispatch:true, nested:true, background:false, subagentToolkit:'full' }` | no background → waves run inline (the #853 flatten rule) |
 | 3 Model | per-agent `model` field on the agent `.md`; no provider `sendRequest` | `modelMode: passive` | tier routing degrades to per-agent model field |
-| 4 Hooks | host `event` bus (~25 events) | `hookBus: host`; ADR-1016 dialect = **`opencode-subset`** | session/tool-scoped only — see gap below |
+| 4 Hooks | host `event` bus (~25 events) | `hookBus: host`; `extensionEvents: opencode` (Phase D / #1943) | session/tool-scoped only — see gap below |
 | 5 State | filesystem `.planning/` + config under xdg `~/.config/opencode`; `opencode-jsonc` permissions sidecar (`permissionWriter: 'opencode'`) | `stateIO: filesystem` | `configHome` write-confinement applies |
 | 6 Artifact | native Agent Skills + `@agent` subagents + slash commands | — | none (full) |
 
@@ -128,7 +129,7 @@ A plugin is a JS/TS module exporting an `async` function that returns a **hooks
 
 ### The load-bearing gap: the loop is phase-scoped, the bus is session-scoped
 
-OpenCode's bus fires on **sessions, tools, files, and permissions** — never on **workflow phases**. GSD's 12 loop extension points (`plan:pre`, `verify:post`, `ship:post`…) have **no event on this bus**. So the imperative adapter for OpenCode cannot drive the loop *from host events*; the engine must own phase sequencing internally and treat OpenCode's bus as a **subset hook surface** (exactly what the ADR-1016 `opencode-subset` dialect already encodes). Concretely:
+OpenCode's bus fires on **sessions, tools, files, and permissions** — never on **workflow phases**. GSD's 12 loop extension points (`plan:pre`, `verify:post`, `ship:post`…) have **no event on this bus**. So the imperative adapter for OpenCode cannot drive the loop *from host events*; the engine must own phase sequencing internally and treat OpenCode's bus as a **subset extension-event surface** — exactly what the `extensionEvents: opencode` vocabulary encodes (amendment #1943; formerly misfiled as a `hookEvents` value `opencode-subset`). Concretely:
 
 - **Steps, gates, and most contributions fire from GSD's own workflow/command invocation (point 1), engine-side** — not from the host bus. The plugin invokes `gsd-tools.cjs` (via `$` or the companion MCP server) and the engine runs the loop resolver.
 - **Only the contributions that align with a real host event bind to the bus.** The clean case is memory: a MemPalace-style capability's capture/recall already keys on `discuss:post`/`plan:post`/`verify:post`; those can *additionally* bind to `experimental.session.compacting` so memory persists across OpenCode's compaction — a concrete win the host gives us for free.