refactor(coderd/x/chatd): unify LLM model-call construction behind a resolver - #28145

…d turn

Add modelCallSpec/resolveModelCall as the single pipeline from a call
spec to a ready model client: config selection, call-config parse,
route and identity resolution, client construction with debug
recording, and provider-option derivation.

Migrate prepareGeneration and the computer-use substitution onto the
resolver, delete resolveComputerUseModel, and turn resolveChatModel
into a transitional wrapper for the remaining quickgen callers.

…-label flows to resolver

Title (auto, manual, override), whole-chat summary, and turn-status label
flows now resolve their model through resolveModelCall with purpose-specific
specs. fantasy.ObjectCall envelopes come from resolvedModelCall.newObjectCall.
Deletes titleGenerationProviderOptions, newQuickgenDebugModel, and the
transitional resolveChatModel wrapper. Summary and status-label calls keep
their historical provider-option omission as declared policy.

…le debug paths to resolver

Compaction overrides build through compactionOverrideSpec, advisor
overrides through advisorOverrideSpec with a typed
modelCallConfigParseError preserving the soft fallback on corrupt
options JSON, and manual-title debug rebuilds through
manualTitleDebugSpec. Deletes newDebugAwareModel,
buildCompactionOverrideModel, and compactionOverrideProviderOptions.

… templates

resolvedModelCall.newCall is now the only production fantasy.Call
construction point. chatloop.GenerateAssistantOptions carries a
CallTemplate instead of ModelConfig plus ProviderOptions, compaction
options carry a prebuilt SummaryCall with tool use forbidden, and
chatadvisor.RuntimeConfig carries a CallTemplate that runs clone
before mutating provider options. The chat-model compaction summary
keeps its historical provider-option omission via
compactionSummaryOverrides.

…E.md

Comment audit: delete restating comments, shorten verbose ones, and fix
the advisor fallback description. Simplify findings: parse model config
options once in resolveModelCall and pass the typed CallConfig to
newModel, and stop grafting the chat config onto the computer-use
resolution. Deslop findings: rename runChatResult.StatusLabel to
StatusLabelCall and lock the resolver's provider-option policy with
TestModelCallShapeProviderOptionPolicy instead of a misleading fixture.

Replace the providerOptionPolicy enum with an omitProviderOptions bool,
replace callOverrides and compactionSummaryOverrides with newCall and
newCompactionSummaryCall, pass newObjectCall arguments directly, and
turn callPurpose into a plain log string. No behavior change.

…ners

Fold standard-turn and compaction-summary template assertions into the
reasoning-effort clamp test, manual-title provider-option assertions
into the synthetic API key test, and status-label envelope assertions
into the structured status-label test. Replace the end-to-end summary
fixture with a focused resolver-level omission test that also covers
summaryObjectCall. Delete modelcall_shape_internal_test.go.

… data

Cleanup-gate follow-up: replace the newCompactionSummaryCall method plus
caller-side ProviderOptions mutation with a compactionSummaryCall free
function that takes the provider options directly.

…spec options

Delete the purpose-specific spec constructors and the per-flow policy
flags (omitProviderOptions, debug tri-state, defaultMaxOutputTokens,
routeOverride). Every resolved call now derives provider options,
defaults MaxOutputTokens, and wraps for debug recording when chat debug
is enabled; clients are always constructed with the configured model
string so gateway validation sees the configured name. Debug helpers
only create and finalize run records instead of rebuilding models.
Callers describe calls with inline modelCallSpec literals whose fields
are plain data: model source, requested effort, and chatd-scoped
routing.