…irect routing

Deprecate the AIGatewayRoutingEnabled deployment option as a no-op
instead of hard-deleting it, so deployments with the flag in their YAML
config or CLI flags don't break on upgrade. Delete the direct routing
implementation, collapse the resolvedModelRoute discriminated union
into aiGatewayModelRoute, remove the dead providerKeys cascade, delete
the preferredShortTextCandidates quickgen function, and simplify the
advisor override error handling.

The option will be fully removed in a future release.

…re-chatd tests

PR #26658 only migrated full-server tests to AI Gateway routing; this
migrates the bare-chatd tests that build *chatd.Server directly via
chatd.New/newActiveTestServer/newTestServer/newInternalTestServer,
which PR #26862 left without a transport factory once direct routing
was deleted.

- 78 tests wired to an existing mock provider URL via
  chattest.NewMockAIBridgeTransport, mirroring the pattern already
  used in 10 passing tests.
- TestProviderSwitchSanitizesAndRestoresPEToolHistory and
  TestComputerUseSubagentToolsAndModel each drive two providers in
  one chat; added a small providerRoutedTransportFactory that
  dispatches by provider name instead of extending chattest's
  single-target mock.
- 6 TestSpawnAgent_* tests had an unrelated bug: they call
  runSubagentTool directly and never set
  aibridge.WithDelegatedAPIKeyID on ctx, so they failed before
  transport wiring was ever reached. Added the missing ctx line,
  copying the established sibling pattern.
- newInternalTestServer gained one new
  withInternalTestServerTransportFactory option, mirroring its
  existing With* options, for the two internal-package tests that
  needed it.
- TestRegenerateChatTitle_PersistsAndBroadcasts and its idle variant
  exposed a real call-sequence change: PR #26862 deleted the
  pre-lock resolveUserProviderAPIKeys call these tests' gomock
  expectations still modeled, and title-generation routing now goes
  through GetAIProviderByID instead of
  GetAIProviders/GetAIProviderKeysByProviderIDs. Rewrote the mock
  expectations and AIProviderID wiring to match the current code
  path; chatd.go itself is untouched.

No production code changed.

…iring sweep

- Restore the outgoing-request model assertion that was lost when
  TestRegenerateChatTitle_PersistsAndBroadcasts and its idle variant
  were rewritten to use a synthesized RoundTripper response instead
  of a real chattest.NewOpenAI server. The previous rewrite returned
  a canned response without checking what model was actually
  requested.
- Drop the PR-number reference from the explanatory comment in both
  tests; state the current routing invariant instead.
- Use the existing withSubagentDelegatedKey helper consistently at
  the two TestSpawnAgent_ComputerUse* call sites instead of spelling
  out aibridge.WithDelegatedAPIKeyID inline, matching the other four
  call sites added in the same change.

…ta, not chat models response

Replace codersdk.ChatModelsResponse.AIGatewayDisabled with a new
ai-gateway-enabled meta tag on the SPA shell HTML, mirroring the
existing HideAITasks/tasks-tab-visible mechanism exactly. This is a
deployment config boolean gating an AI feature for all users, not
just deployment admins, so /api/v2/deployment/config (which requires
ResourceDeploymentConfig read) isn't a usable home for it; the
embedded-metadata mechanism already serves this exact purpose.

- listChatModels drops its AI-Gateway-disabled special case and
  falls through to its normal catalog logic unconditionally, same
  as before this branch added the special case.
- AgentsPage reads the new metadata key, defaulting to enabled (not
  disabled) when the tag is unavailable in dev/Storybook, so the
  banner doesn't show by default outside production.
- AgentsPageView's own prop contract and its AIGatewayDisabled story
  are unchanged; only the data source moves.
- Removed now-dead ai_gateway_disabled fixture lines forced by the
  type change across 5 story/test files.

…extual notice when AI Gateway is disabled

Previously the top-of-page banner in AgentsPageView was the only
signal that AI Gateway was disabled deployment-wide, and it didn't
disable anything. Since this session's listChatModels simplification
made /models reflect DB config regardless of runtime gateway state,
a deployment with providers/models configured but AI Gateway off
would show a chat input that looks usable but cannot actually work.

- AgentSetupNotice gains an aiGatewayDisabled branch, checked before
  its existing provider/model checks (it can be true even when the
  catalog has entries), reusing the existing admin-aware messaging
  convention instead of a disconnected page banner.
- AgentChatInput forwards the flag into its existing setup-notice
  logic.
- Both composer entry points (AgentChatPage for existing chats,
  AgentCreateForm for new chats, reached via AgentsPage and
  AgentCreatePage) now fold aiGatewayDisabled into their existing
  disabled-state checks and read the ai-gateway-enabled embedded
  metadata added earlier this session.
- Removed the now-redundant top-of-page banner and its story from
  AgentsPageView; AgentsPage no longer needs the metadata read it
  only fed.

…isabled

Two crashes when AI.BridgeConfig.Enabled is false (chatDaemon stays
nil):

- The git sync worker captured api.chatDaemon.PublishDiffStatusChange
  as a method value unconditionally. A method value on a nil pointer
  receiver is itself non-nil, so the worker's own nil check on its
  callback never caught it; calling it later dereferenced the nil
  receiver and panicked.
- api.chatDaemon.Close() was called unconditionally during shutdown,
  immediately dereferencing the nil receiver in
  closeInflightAdmission().

Only bind the git sync worker's publish callback when chatDaemon
exists, and guard the shutdown Close() call. Added
TestAIGatewayDisabledStartupAndShutdown, which panics without this
fix (verified by reverting it locally) and passes with it.

Also adds Test_warnDeprecatedChatAIGatewayRoutingFlag, covering the
three explicit ValueSource cases plus the default/absent cases, per
review feedback.

Addresses coder-agents-review P0 and P3 findings on this PR.

…scan

warnDeprecatedChatAIGatewayRoutingFlag looped over the whole option
set with a continue to find one flag by name; serpent.OptionSet
already ships ByFlag for exactly this lookup. Also drops one of two
test cases (Default, None) that exercised the identical no-op
branch of the deprecation switch.

Per ponytail review.

- Drop the redundant NonNilChatDaemon subtest; a real non-nil
  chatDaemon's PublishDiffStatusChange is already exercised in
  coderd/exp_chats_test.go. Keep only the NilChatDaemon case, which
  is the actual non-obvious behavior this guard protects.
- Drop the deprecated-option existence/description assertions added
  to TestChatAIGatewayRoutingEnabledDefault; overkill for an upgrade-
  safety no-op contract.
- Collapse AgentSetupNotice's aiGatewayDisabled admin/non-admin
  branches into one message. Unlike the provider/model branches,
  there's no in-app settings page for this deployment-level flag for
  either audience, so the isAdmin distinction didn't gate any real
  capability difference.
- Rename useAIGatewayDisabled to useAIGatewayEnabled, returning the
  metadata value directly instead of inverting it internally, so the
  hook name matches the ai-gateway-enabled metadata key it reads.
  Callers invert locally (!useAIGatewayEnabled()) where they need
  the disabled polarity. Also drops a redundant .available check in
  the hook itself, since registerValue only ever sets available:true
  alongside a defined value.

…ed-flag-and-dead-direct-routing

- Collapse a vestigial flex-row wrapper div in AgentsPageView left over
  from a removed banner; restore sm:flex-row on the outer layout div.
- Drop three comments that only restated the code they annotated.
- Tighten requireChatDaemon's doc comment: it no longer covers all
  chat HTTP operations now that pure reads (getChat) are exempt.

…ed-flag-and-dead-direct-routing

Resolves conflicts with #26877 (drop chat_model_configs provider column):
- Delete model_routing_direct.go entirely (dead code path this PR removes).
- Drop `Provider` field from database.ChatModelConfig literals (column dropped).
- Replace `route.providerHint()` calls with `string(route.Provider.Type)`
  (method removed alongside the resolvedModelRoute union we collapsed).
- Keep our single-`candidate` shape in quickgen; drop the multi-candidate
  fallback fragments from main.
- Fold main's added `GetAIProvider*` mock expectations into our test wiring
  where needed; preserve our aibridgeTestAIProvider helper (which sets
  Provider.Name, required by newAIGatewayModel validation).