fix(#124): HAR recording returns 0 entries - add response_body via getResponseBody

[lekt9]

Adds request_id and response_body fields to HarEntry. New fetchResponseBodies helper loops entries after event flush, calls Network.getResponseBody for each, stores body. handleHarStop wires this in before final serialisation. All 17 har+client tests pass.

[justrach]

Review

Good addition — response bodies are the missing piece in our HAR capture. The implementation is clean. A few suggestions:

✅ What's good

• request_id and response_body fields properly owned/duped/freed
• fetchResponseBodies() is a clean post-processing step
• Existing tests updated with new fields

⚠️ Suggestions

1. Filter response bodies — Currently fetchResponseBodies() fetches bodies for ALL entries indiscriminately. PR #122 had a smarter filter that's worth adopting:

   // Skip non-useful responses
   if (entry.status < 200 or entry.status >= 400) continue;
   if (std.mem.endsWith(u8, entry.url, ".js") or std.mem.endsWith(u8, entry.url, ".css")) continue;
   // Only fetch JSON, HTML, plain text
   if (std.mem.indexOf(u8, entry.mime_type, "json") == null and
       std.mem.indexOf(u8, entry.mime_type, "text/html") == null and
       std.mem.indexOf(u8, entry.mime_type, "text/plain") == null) continue;

   Without filtering, a page with 50+ entries (images, fonts, CSS, JS) will make 50+ serial CDP calls, adding significant latency to /har/stop.

2. Duplicate doc comments — There are doubled /// Add a manually observed request/response and /// Serialize current entries to HAR 1.2 JSON format. comments in har.zig. Remove the duplicates.

3. response_body type — You use []const u8 (always allocated). PR #122 used ?[]const u8 (optional) which is cleaner — only allocate when we actually have a body. This avoids allocating empty strings for entries where we skip body fetch.

4. Rebase needed — har.zig on main now has OwnedEvent struct changes and different field layout. Should be a straightforward rebase.

🔧 After addressing

This is good to merge once the filtering and optional body type are in.

[justrach]

Can you update this PR to build and test against Zig 0.16.0? The active release/CI path is now targeting 0.16.0, so it would help to rebase and rerun on that toolchain before review.

[justrach]

Re-reading this against current main:

• The EventBuffer cross-arena fix here is the right shape — copies events into the long-lived allocator instead of holding a dangling request-arena pointer, with a regression test. Independent of the response-body work.
• Adding request_id + response_body to HarEntry is the feature half. Two questions before this is mergeable:
  1. Is anything actually calling Network.getResponseBody and threading the result into addEntry? The struct field is added, but I don't see the CDP roundtrip wired up in the diff — without that the field would always be empty. Did I miss it, or is it intentionally a follow-up?
  2. Response bodies can be megabytes; do you want a size cap on what gets pulled, or a --har-body-max-bytes flag?

If you'd rather just land the cross-arena fix now (which is uncontroversial) and split the response-body capture into its own PR, I'd merge the EventBuffer part today.