feat(q): log every request in the core's query log

[ihrfv]

How it looks from the user's side

With the core's query log on, a request that arrives over the q wire now leaves a row in .sys.querylog, as a native IPC request already does:

(.q.send h "(.sys.querylog.enable 1)")
(.q.send h "(+ 1 2)")
(.q.send h "(+ 1 2")                      ; a parse error
(.q.send h "(count (.sys.querylog))")

Before -- the q wire server never writes to the log, so the ring stays empty (test/rfl/server/querylog.rfl against a server built from master):

querylog.rfl:18: expected "2", got "0"  -- src: (.q.send h "(count (.sys.querylog))")
querylog.rfl:19: expected ""(+ 1 2)"", got """"  -- src: (.q.send h "(at (get (.sys.querylog) 'query) 0)")
querylog.rfl:30: expected "true", got "false"  -- src: (.q.send h "(like (last (get (.sys.querylog) 'query)) \"(upd ...) * B\")")
querylog.rfl FAIL (14 assertions)

After -- one row per evaluated request, with its duration (decode included), status and source:

querylog.rfl ok (14 assertions)

Why

The q wire server evaluates every request on the poll thread, beside everything else the runtime does. When a query waited seconds, nothing said which request held the thread: the native IPC server records one row per query in .sys.querylog, and the q wire server recorded none. A publisher's pushes ((upd; table)`) are exactly the requests a query queues behind.

What changed

• q_read_body brackets decode and evaluation with the core's ray_qlog_begin / ray_qlog_end, so each request is one row once .sys.querylog.enable is on. With the log off, ray_qlog_begin is one relaxed load and the source is never formatted.
• A string request is logged verbatim, cut at RAY_QLOG_QUERY_MAX (256 bytes). A list request, the shape a publisher pushes, is logged as its head and the frame's size, (upd ...) 132 B: formatting a pushed table back to source would cost more than evaluating it.
• A RESPONSE frame is data for a parked q_conn_send, not a query, and is not logged.
• test/rfl/server/querylog.rfl covers a sum, a parse error, the duration and a list request. q_server.h and docs/CHANGELOG.md (Unreleased, Added) describe the same behaviour.

Overlap with open PRs

Merges cleanly with #10. Conflicts with #8 on one line: both add an #include right after core/sock.h (core/timer.h there, core/qlog.h here), and keeping both resolves it.

Tests

git diff --check origin/master...feat/q-server-querylog
CC=clang make test
    # core master at 4eb3b10: codec and exchange selftests ok,
    # 3 server files passed (querylog.rfl 14 assertions);
    # real-q legs skipped, no q binary, as in CI
# before/after above: test/driver built from origin/master and from this
# branch, each serving test/rfl/server/querylog.rfl

[protocolstardust]

Reviewed and merged (f5cf1b5). Thanks, this is a real gap: the q wire was the only entry point that left no trace in .sys.querylog.

No CI ran on this PR (first-contribution gate), so it was run locally instead: core master 31f45a7, querylog.rfl 14 assertions, full suite green with #7, #8, #10 and #11 merged first, and the full rayforce binary built with the glue.

Review

• ray_qlog_begin / ray_qlog_end bracket decode + eval, so the row's duration is what the poll thread actually paid. RESPONSE frames stay out. With the log off, begin is one relaxed load and the source is never formatted.
• q_log_source for a list request: head symbol + frame size is the right trade; formatting a pushed table back to source would cost more than the push.
• After the merge with fix(q): validate handles and wire protocol frames #10 there is now an early return (unknown message type → deregister) between begin and end. ray_qlog_begin only touches the stack context, so nothing leaks today; if ray_query_measure_begin ever acquires global state, that path needs an end too.
• Cosmetic: a string request is cut at 256 bytes, which can split a multi-byte UTF-8 sequence. Same as the native IPC's rows, so fine for now.

The #include conflict with #8 was resolved by keeping both lines, as the description predicted.

[protocolstardust]

Reverted on master in 92e0d91 (git revert -m 1 f5cf1b5): maintainer decision, the q wire server should not write into the core's query log for now. The change itself was reviewed as sound (see above); nothing in the revert reflects on the code. The suite is green after the revert.