fix(chat): preserve final answers after tool calls

[1052326311]

Problem

A read-only request such as “inspect the project files, then describe the app” completes its tools but shows no visible answer. The final ordinary-text segment is provisionally stored as thinking, and the end-of-stream branch only promotes it when hadToolCalls is false. A reproduced turn persisted zero content characters and 286 thinking characters.

Change

Promote the final provisional ordinary-text segment after successful completion even when earlier segments invoked tools. Preserve explicit reasoning and tool-confirmed intermediate text, leave deferred-error segments alone, and use the existing thinking_to_text SSE event. Remove the exact buffered segment instead of blindly trimming the thinking suffix. Preserve thinking markers until the channel router classifies them, and track complete-message catch-up offsets per SDK message so earlier reasoning cannot swallow a later non-streamed answer. Offsets use raw SDK text before sanitization, preventing repeated tails when a split term such as package.json expands only in the complete message.

Validation

• Seven focused regression tests pass, covering post-tool answers, direct answers, explicit reasoning, deferred errors, trailing native reasoning, prior content, and conservative handling of unmatched text.
• Seven event-processor integration tests pass in the production dependency environment, covering SDK tool round-trips, tagged reasoning, subsequent tools, complete-only final text, split thinking markers, and catch-up offsets when sanitization changes text length.
• Existing Supabase/MCP recovery guards: three tests pass.
• Full API type-check in the production dependency environment reports 11 pre-existing errors in src/index.ts and src/routes/preview-proxy/proxy-handler.ts (the affected source is unchanged from main); no changed reply module reports an error. This is not a clean full type-check.
• Official Docker production API build passed. The final 17 tests also pass in an isolated container using the production dependencies.
• Real-provider read-only browser validation passed: the model read a project file, the final ordinary answer arrived through thinking_to_text, onDone fired, and the answer remained visible after reauthentication/history reload. The persisted record has 185 content characters, 26 reasoning characters, and tool calls; its final marker occurs once.
• An independent Chromium HTTP/2 PING transport failure was observed during deployment validation. A site-local HTTP/1.1 compatibility setting restored a clean live SSE run; this environment-specific setting is deliberately excluded from the PR. Multi-user collaboration and broader provider coverage have not been manually exercised.

No automatic history migration is included in this change.

[1052326311]

Superseded by #59, which consolidates this final-response fix with the related lifecycle/session repairs in one clean commit. Closing this duplicate to avoid overlapping merges.