feat(hub): forward desktop bridge tool results instead of re-wrapping (E4)

[physercoe]

Vision-parity E4 (asymmetry #3). A screenshot taken through the hub reached a remote agent as base64 prose.

routeTunnelInvoke passed the desktop's reply through mcpResultJSON, which renders it as JSON inside one text block — so {"content":[{"type":"image","data":"<base64 PNG>"}]} arrived as text no model can look at.

It was never image-specific

Reading the far end widened the fix. dispatchHubInvoke answers a relayed call with the object callTool returns, and every bridge tool returns an MCP tool result — 16 textContent() calls plus the literal {content:[{type:'image',…}]} returns. The relay was re-wrapping an already-conformant result on every call; the image case is just where the extra layer destroys the payload instead of merely quoting it.

The relay now forwards a well-formed tool result verbatim — isError and extra keys included, since "what a local caller sees" is the point — and keeps the JSON wrapper for anything else. Local and relayed calls now return the same thing.

The guard errs toward the wrapper, deliberately

content must be a non-empty array of objects each carrying a string type.

That rejects {"content":[]}, which is a legal MCP result — a false negative costing a reader one layer of quoting. The other direction would hand the agent something malformed, and nothing distinguishes "an empty tool result" from "an arbitrary object with an empty list under that key". Losing nothing beats forwarding garbage.

Two things checked rather than assumed

• The Go fixtures were describing a desktop that does not exist. okBrowserEnvelope was being handed bare JSON ({"tabs":…}, {"clicked":true}) — a shape no bridge tool has ever returned. So the existing tests could not have caught the double-wrap and would not have noticed the fix either. The read-routing test now asserts against the real reply shape, and okBrowserToolResult names the distinction for the callers that still exercise the fallback on purpose.
• The rest of the chain forwards it. The host-runner gateway returns a tools/call result verbatim (gwIdentity only stamps _meta server info), so the image block survives desktop → hub → gateway → agent.

Verification

• Full go test ./... — exit 0, all 33 packages (server suite 290s); build + vet clean.
• New tests on both relayed classes: browser_invoke (image block survives; text stops being double-wrapped; five non-tool-result shapes keep the wrapper) and ui_screenshot (the path the asymmetry is named for), plus a 15-case table for the shape predicate including the {"content":[]} boundary it deliberately gets "wrong".
• Four mutations checked, all fail: removing the passthrough, dropping the shape guard, skipping the per-block type check, accepting an empty content array.
• Full lint set green, including lint-openapi (venv) and lint-desktop-tokens from the repo root. NUL + CJK scans clean.

Note for the merge

Branched off main, not off #563 (E3) — the code paths don't overlap. Both edit the plan's status line, so whichever lands second needs a one-line rebase there.

🤖 Generated with Claude Code