Add Vue LSP and enable eslint for .vue files. - #1952

* docs(core): plan V2 config env sources for swxtch-egress (anomalyco#1952)

Plan for swxtchio/swx-abbe#1952.

Root cause: the V2 Config layer (packages/core/src/config.ts) reads only
the global config directory and discovered project files. It ignores
OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT. The jail supplies its whole
config through those two variables (Dockerfile.jail sets OPENCODE_CONFIG,
opencode_server.py renders OPENCODE_CONFIG_CONTENT), so the V2 catalog
never sees the swxtch-egress provider or a configured default model.
V1 (packages/opencode/src/config/config.ts) reads both, which is why V1
turns run on the egress provider.

Change:
- Config takes explicit file and content sources, defaulting to
  Flag.OPENCODE_CONFIG and Flag.OPENCODE_CONFIG_CONTENT, with a
  layerWith seam for tests (mirrors Global.layerWith and upstream's
  Config.Options file/content from 43c0838).
- Precedence matches V1 and upstream: global directory, then the
  OPENCODE_CONFIG file, then project files, then project .opencode
  directories, then OPENCODE_CONFIG_CONTENT last.
- V1-format sources migrate through the existing loader path.

Tests:
- Config loader orders explicit file and content entries.
- A jail-shaped V1 config supplied as content reaches the catalog and
  SessionRunnerModel resolves the named swxtch-egress model to an
  OpenAI-compatible route at the proxy URL.

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* fix(core): load env-supplied config in V2 config

The V2 Config layer read only the global config directory and discovered
project files, ignoring OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT. The
OpenCode jail supplies its whole config through those variables, so the
V2 runner's catalog never held the swxtch-egress provider or the
configured default model: a named egress model failed with
ModelUnavailableError and an unnamed one fell to another provider
(swxtchio/swx-abbe#1952).

Config now takes explicit file and content sources, defaulting to the
two flags, and orders them as V1 does: global directory, explicit file,
project files, .opencode directories, then inline content. V1-format
sources migrate through the existing loader path. Config.nodeWith names
the sources directly for tests.

Tests drive the real loader: precedence of every source, and the jail's
V1 config (placeholder file plus rendered content) through the config
provider plugin and catalog to SessionRunnerModel, for a named model and
for the configured default against a competing provider.

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* fix(core): default V2 sessions to config-declared providers

Review round 1: with the env-config fix alone, a location-only V2
session in the jail (create_session_v2 names no model, and the jail
config sets none) still resolved a credential-free catalog model such
as opencode/hy3-free ahead of the config-declared swxtch-egress model,
so egress turns still failed.

V1's Provider.defaultModel takes its default from the config-declared
providers when no model is configured. The V2 catalog now does the
same: the config provider plugin records the declared providers in
config order, and Catalog.model.default picks the newest available
model of the first one that has any, before its newest-available
fallback. An explicit session model or configured `model` still wins.
The preference registers on the internal catalog under the plugin's
scope, because the public plugin catalog API has no such hook.

Tests: the jail path through the production location runtime (every
internal plugin, the recorded models.dev corpus, the jail config via
OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT) resolves a location-only
session to the egress model; and the preference skips an unavailable
declared provider and beats a newer undeclared model. Both are red
without the preference.

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* docs(core): plan V2 config env sources for swxtch-egress (anomalyco#1952)

Plan for swxtchio/swx-abbe#1952.

Root cause: the V2 Config layer (packages/core/src/config.ts) reads only
the global config directory and discovered project files. It ignores
OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT. The jail supplies its whole
config through those two variables (Dockerfile.jail sets OPENCODE_CONFIG,
opencode_server.py renders OPENCODE_CONFIG_CONTENT), so the V2 catalog
never sees the swxtch-egress provider or a configured default model.
V1 (packages/opencode/src/config/config.ts) reads both, which is why V1
turns run on the egress provider.

Change:
- Config takes explicit file and content sources, defaulting to
  Flag.OPENCODE_CONFIG and Flag.OPENCODE_CONFIG_CONTENT, with a
  layerWith seam for tests (mirrors Global.layerWith and upstream's
  Config.Options file/content from 43c0838).
- Precedence matches V1 and upstream: global directory, then the
  OPENCODE_CONFIG file, then project files, then project .opencode
  directories, then OPENCODE_CONFIG_CONTENT last.
- V1-format sources migrate through the existing loader path.

Tests:
- Config loader orders explicit file and content entries.
- A jail-shaped V1 config supplied as content reaches the catalog and
  SessionRunnerModel resolves the named swxtch-egress model to an
  OpenAI-compatible route at the proxy URL.

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* fix(core): load env-supplied config in V2 config

The V2 Config layer read only the global config directory and discovered
project files, ignoring OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT. The
OpenCode jail supplies its whole config through those variables, so the
V2 runner's catalog never held the swxtch-egress provider or the
configured default model: a named egress model failed with
ModelUnavailableError and an unnamed one fell to another provider
(swxtchio/swx-abbe#1952).

Config now takes explicit file and content sources, defaulting to the
two flags, and orders them as V1 does: global directory, explicit file,
project files, .opencode directories, then inline content. V1-format
sources migrate through the existing loader path. Config.nodeWith names
the sources directly for tests.

Tests drive the real loader: precedence of every source, and the jail's
V1 config (placeholder file plus rendered content) through the config
provider plugin and catalog to SessionRunnerModel, for a named model and
for the configured default against a competing provider.

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* fix(core): default V2 sessions to config-declared providers

Review round 1: with the env-config fix alone, a location-only V2
session in the jail (create_session_v2 names no model, and the jail
config sets none) still resolved a credential-free catalog model such
as opencode/hy3-free ahead of the config-declared swxtch-egress model,
so egress turns still failed.

V1's Provider.defaultModel takes its default from the config-declared
providers when no model is configured. The V2 catalog now does the
same: the config provider plugin records the declared providers in
config order, and Catalog.model.default picks the newest available
model of the first one that has any, before its newest-available
fallback. An explicit session model or configured `model` still wins.
The preference registers on the internal catalog under the plugin's
scope, because the public plugin catalog API has no such hook.

Tests: the jail path through the production location runtime (every
internal plugin, the recorded models.dev corpus, the jail config via
OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT) resolves a location-only
session to the egress model; and the preference skips an unavailable
declared provider and beats a newer undeclared model. Both are red
without the preference.

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* fix(core): run V2 turns only after the location's plugins boot

A Location boots its internal plugins in a forked fiber, and nothing on the
turn path waited for it. The jail's adapter creates a session and prompts it at
once, so the drain that boots the Location starts its first turn while the
catalog is still empty or only partly built. On the jail image running
v1.18.32-swxtch.774d90c7 the first turn failed with ModelNotSelectedError (or
ModelUnavailableError when the egress model is named), or was sent to a bundled
model such as nano-gpt/..., and the egress provider received nothing.

The internal plugin layer now exposes a readiness signal that completes when
its boot ends, whether it loaded every plugin or failed, and every runner turn
waits on it before it selects the agent, model, tools and guidance.

Refs swxtchio/swx-abbe#1952

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

* docs: record #350 in CHANGESET

Claude-Session: https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp