fix(har): embed content that is not valid UTF-8 as base64 - #43524

Pre-existing flakes

• [msedge] › mcp/annotate.spec.ts:417 › should cancel browser_annotate when the MCP client disconnects — failed in 3 of 498 prior runs for the same project and Windows Edge bot, including this run for #43173.
• [msedge] › mcp/http.spec.ts:132 › http transport browser sigint — failed in 6 of 498 prior runs for the same project and Windows Edge bot, including this run for #43432.
• [playwright-test] › web-server.spec.ts:121 › should run web server with PLAYWRIGHT_TEST=1 environment variable — the same Windows Server 2025 bot had the same timeout and Expected: 0 / Received: 1 failure on unrelated main revision 0d7c141 in this prior run; it passed the other 165 of 166 prior runs.

Overall assessment

The failures do not look caused by this PR. A clean rerun would resolve the two uncertain results; without one, the available history is not enough for a green verdict.

Pre-existing flakes

• [playwright-test] › playwright.reuse.spec.ts:460 › should reset tracing — timed out with the same Expected: 0 / Received: 1 failure on unrelated #43405. It failed in 2 of 170 recent runs on this Windows runner and passed in 168.
• [playwright-test] › playwright.spec.ts:303 › should respect headless in modifiers that run before tests — timed out with the same Expected: 0 / Received: 1 failure on unrelated #43405. It failed in 2 of 170 recent runs on this Windows runner and passed in 168.
• [playwright-test] › playwright.unhandled.spec.ts:35 › should produce unhandledRejection when page.route raises — timed out with the same Expected: 1 / Received: 0 failure on unrelated #43405. It failed in 2 of 170 recent runs on this Windows runner and passed in 168.

Uncertain

• [playwright-test] › expect.spec.ts:943 › should support mergeExpects with builtin matcher overrides — the nested run timed out, but this is its first failure in 170 recent runs on this Windows runner. The HAR-only diff does not reach it, but there is no unrelated failure that proves it is a flake.
• [playwright-test] › reporter-json.spec.ts:360 › should report parallelIndex — reported workerIndex 0 instead of 2, but this is its first failure in 170 recent runs on this Windows runner. The PR does not touch reporter or worker allocation code, but there is no unrelated failure that proves it is a flake.