fix(q): reject oversized credential strings

[belowzeroff]

What changes for users

.q.connect now rejects user or password strings longer than 127 bytes with a length error instead of silently truncating them and later reporting a misleading authentication failure.

Valid credentials and all existing valid .q.* calls keep their behavior.

Why

The credential buffers are 128 bytes including the terminating NUL. Previously, oversized credentials were truncated before the handshake, which could make a valid user input fail with an unrelated auth error.

Validation

• Rebased onto fresh master (92e0d91)
• make test: codec, CLI, server, malformed-frame, real-q interop, all client suites, poll, and push tests passed
• make test client regression includes a 128-byte username and expects length
• git diff --check: passed

This PR intentionally contains only the length validation and its regression test; the overlapping type, timeout, and handle checks are already in PR #10.

[protocolstardust]

Not merged as is: this overlaps #10, which landed on master today (e08ab96) and already covers most of it:

• bool / temporal atoms are no longer accepted as integers,
• user / password must be strings,
• timeout is validated (type and 0..INT_MAX),
• .q.send / .q.close require a non-negative i64 handle,
• closing an unknown handle is a handle error.

The branch now conflicts in embed/rayforce_q.c, and the new 01_connection.rfl lines duplicate what 06_errors.rfl / 07_auth.rfl assert since #10.

What is still new here and worth keeping

• A length error for a user/password longer than 127 bytes instead of silent truncation (which today ends in a confusing auth failure). This is a real improvement.

What I would not take over #10

• Accepting negative timeouts (INT_MIN..INT_MAX): fix(q): validate handles and wire protocol frames #10 settled on 0..INT_MAX for the builtin.
• Accepting i16/i32 handles: handles are the i64 .q.connect returns, so fix(q): validate handles and wire protocol frames #10's exact-type check is the tighter contract.

Please rebase onto master and reduce the PR to the length check plus one test line, then it can go in quickly. Leaving this open for that.

[belowzeroff]

Addressed:

• rebased onto current master
• removed the overlapping type, timeout, and handle changes already covered by fix(q): validate handles and wire protocol frames #10
• kept only the length error for user/password longer than 127 bytes
• reduced the regression coverage to one client test line

make test passes, including the new oversized-username case.