Stop allocating a scratch vector for every bytecode &key call

[dg1sbg]

parse-key-args in the bytecode VM builds a SimpleVector to collect keyword arguments, then immediately pushes its contents onto the VM stack and drops it. The vector never outlives the opcode. It is also allocated unconditionally — before the lcc_nargs > more_start guard — so a call that passes no keywords at all still pays for it.

The cost is 24 + 8n bytes per call for a callee with n &key parameters, charged whether or not the caller passes any keywords:

callee	called as	before	after
(defun k1 (a &key x) ...)	(k1 1)	32 B	0 B
(defun k2 (a &key x y) ...)	(k2 1)	40 B	0 B
(defun k3 (a &key x y z) ...)	(k3 1)	48 B	0 B
(defun k6 (a &key p q r s u v) ...)	(k6 1)	72 B	0 B
(defun kopt (a &optional o &key x y) ...)	(kopt 1)	40 B	0 B
string=	(string= "abc" "abd")	56 B	0 B

Measured with gctools:bytes-allocated over 200000 warmed calls.

The change

The parameter slots are the destination anyway, so push them first and write each argument directly into its slot.

The index identity is derived rather than assumed: the loop being deleted pushed key_id descending, so key_id 0 ended up on top, and stackref is documented in threadlocal.h as "0 is most recently pushed" — hence slot n is key n. This matches the comment at cleavir/compile-bytecode.lisp:879, "the leftmost key is the last pushed".

The odd-argument-count check is hoisted above the pushes so that error path cannot observe them. The unrecognized-keyword throw still can, which is safe: the VM stack is a registered GC root (allocateRootsAndZero, gctools/threadlocal.cc), sp is a plain C local rather than vm._stackPointer, and both ~VMFrameDynEnv_O and VMFrameDynEnv_O::proceed restore _stackPointer on unwind. Slots are initialized to unbound before anything that can collect runs.

Peak stack use is unchanged — exactly key_count slots are pushed either way, just earlier.

The long-operand form at the second parse_key_args case gets the same treatment for consistency. It takes more than 127 &key parameters to reach, since key_count_info is (key_count << 1) | aokp.

Tests

A reversed or off-by-one slot index would bind the wrong parameter with no error at all, so the new tests pin the slot identity rather than the allocation:

• every argument permutation of a 3-key callee, plus each keyword alone and none
• leftmost-wins on duplicate keywords (CLHS 3.4.1.4) — the VM gets this by scanning right to left and overwriting, so only the write destination changes here, not the loop order
• &allow-other-keys in the lambda list and as a runtime argument, plus the unknown-keyword and odd-argument error paths
• 16 parameter slots live across a collection

Verification

Built from scratch on this branch (koga + ninja), no local modifications to any dependency:

• test-boehm — 1965 successes, no unexpected failures
• test-boehmprecise — 1967 successes, no unexpected failures
• ansi-test-boehm — 21936 tests, 27 failures, no unexpected failures

macOS arm64, LLVM 22.1.8.