refactor(coderd/x/chatd): extract turn preparation into a deep TurnEnvironment module - #28757

Summary

Chat turn preparation was smeared across generation_preparer.go (926 lines), chatd.go (turnWorkspaceContext plus tool policy and system prompt assembly, ~1,200 lines), and generation.go, which consumed a 25-field generationPrepared bag with 8 correlated lookup maps and ~10 policy helpers on Server. Tool policy could only be tested by booting a full Server + database fixture, which is why ~950 lines of e2e tool-filtering tests existed.

This extracts one deep module, coderd/x/chatd/turn_environment.go: buildTurnEnvironment(ctx, server, input) returns a six-method turnEnvironment interface (Turn, ModelConfig, Prompt, Toolset, CompactionConfig, Close) backed by cohesive turn/model/toolset state. Workspace agent connections, skills and instruction lookup, plan/explore/subagent tool policy, system prompt assembly, deferred MCP tool search setup, and dynamic tool wrapping are hidden behind it. generation.go consumes only the interface; the policy helpers moved off Server.

Net diff: 12 files, +2,370/-3,414 (-1,044 lines) across production and tests.

Behavior

Behavior preserving by design: the generation decision still reads the user-visible message set unchanged while the provider prompt is built from the model-visible set, deferred MCP tool search is untouched semantically, tool admission/hook dispatch is untouched, and model-call resolution is wrapped, not moved. Remote dogfood UAT differentially compared this branch against main on the same deployment config: per-turn tools[] inventories captured at the provider boundary were identical across normal/plan/explore/subagent/workspace/dynamic/deferred-MCP/inactive-filter scenarios, and the system prompt was byte-identical.

Tests

• Deleted seven e2e tool-policy tests from chatd_test.go (~950 lines) that booted a full Server + DB fixture; their assertions are covered by direct table-driven unit tests on the module (TestActiveToolNamesForTurn, TestFilterExternalMCPConfigsForTurn, TestAllowedExploreToolNames, TestStopAfterPlanTools, TestStopAfterBehaviorTools, plus toolset metadata coverage).
• Kept TestSubagentChatExcludesWorkspaceProvisioningTools as the end-to-end smoke proving a turn still assembles.
• ARCHITECTURE.md gets a TODO(human) marker only, per repo guardrails.

Xum acted on behalf of @ibetitsmike: implementation, validation, remote dogfood UAT, and this PR were produced by an AI agent under his direction.