fix(fetch): preserve UTF-8 when truncating fetched content - #4025

Fetching a large UTF-8 page can fail with Response content is not valid UTF-8 when the byte limit lands inside a character. For example, 102399 ASCII bytes followed by 界 trigger the error even though the response is valid UTF-8. Formatting the response as Markdown can also split a character at the output limit.

Read a small, bounded lookahead to validate the character at the boundary, then keep a complete UTF-8 prefix. This fixes both fetch and sub-agent web_fetch while preserving errors for malformed input. The existing truncation notice now reflects whether the response was actually cut, including when HTML extraction makes the result shorter; a complete response exactly at the limit no longer gets a notice.

Added 12 local HTTP regression cases covering two-, three-, and four-byte characters, Markdown output, exact limits, and malformed input. Kept upstream’s refreshed cassette and updated two recorded requests to include the truncation notices for the plain-text and HTML fetch results. The recorded responses were preserved and replayed locally.

Validation on macOS with Go 1.27.0, rebased onto 3b1f8604f0027cf5fa47f29a27655ef1aeed8d8f (head 4eeedcdd2a7230d380370ab76299a3bf5af230c1):

• The UTF-8 regressions fail on the unpatched upstream helpers and pass with this fix.
• go test -race ./internal/agent/tools -run TestFetch -count=1 and the DeepSeek fetch cassette replay passed.
• go mod tidy left the module files unchanged and go build -race ./... passed. The full go test -race -failfast ./... rerun passed. The first run reported a race in the unchanged Docker MCP backend test; it did not reproduce in five targeted or six full backend-package runs on upstream.
• CI-pinned golangci-lint v2.13.1, log-capitalization lint and git diff --check passed.

The rebase preserves the newer upstream fetch error handling and its tests. No dependency, workflow or unrelated source changes are included. GitHub Actions results are separate from these local checks.