fix(build-cache): register the two unkeyed codegen env vars (main CI is red) - #9044

…his PR's knob

Merging current main into this branch REVERTED PerryTS#9044: commit e3164ee removed
`PERRY_BOX_CAPTURE_ENTRY_CELLS` and `PERRY_GUARDED_PREINLINE_MAX_IR_BYTES` from
BUILD_CACHE_ENV_VARS along with PerryTS#9044's changelog fragment. The branch predates
that fix, so the commit was built over a stale tree and carries the removal as
an intentional-looking deletion -- which a merge then honours.

That alone would have re-reddened main: the assertion lives in a bin-crate unit
test, so its failure stops the whole `perry` test binary compiling and every
open PR's cargo-test job goes red.

This PR also adds a third codegen knob, PERRY_PACKED_LOOP_NUMERIC_ACCUMULATOR,
without registering it. It is a cache INPUT, not an exclusion: with it on,
`s += arr[i]` lowers to an inline fadd instead of
`js_dynamic_string_or_number_add`, so the two settings emit different code and
must never share a cached object.

All three registered; fragment renumbered 0000 -> 9060.

…ic proof — s += arr[i] at node parity (was 5.3×) (#9060)

* codegen: reduce accumulators earn the stable-packed fast clone's numeric proof

`for (let i = 0; i < arr.length; i++) s += arr[i]` — the most common reduce
shape in JavaScript — ran 5.3x slower than node, and BOTH halves of the reason
were invisible to profiling alone:

1. Inside the fast clone, `s += arr[i]` still lowered `+` through
   `js_dynamic_string_or_number_add` (25% of the isolated loop): the loop
   guard proves the ELEMENT is raw f64, but the accumulator's own writes are
   circular for every whole-function numeric fact, so the add had one unproven
   operand. The element-shape clone already solved this with its
   `numeric_accumulator` (preheader tag test = the induction base case; every
   in-clone write numeric-preserving = the step). This ports that design:
   `collect_numeric_accumulators` admits plain, uncaptured, unboxed locals
   whose every body write is numeric with all leaves provable in-loop
   (fail-closed fixpoint; nested closures not descended — their captures are
   boxed and excluded anyway), the fast preheader tag-tests each one and takes
   the slow clone on any non-Number, and the fact rides
   `StablePackedLoopFact::numeric_accumulators`, scoped to the fast-clone
   lowering exactly as the element facts are.

2. With the add fixed, the clone was STILL dead: the accumulator's
   per-statement shadow CLEAR — `js_shadow_slot_set(slot, 0)`, emitted
   precisely BECAUSE the stored value is a proven non-pointer — failed
   `fast_clone_call_free`, and the admission arm then emits an UNCONDITIONAL
   branch to the slow preheader while still calling (and discarding) the
   guard. Timing shows slow, lldb on the guard shows "admitted", the IR shows
   a perfect fast body: nothing points at the terminator. `js_shadow_slot_set`
   is a bounds-checked TLS store (`gc/roots/shadow_stack.rs`) that cannot
   allocate, collect, or revoke a layout — which is precisely what the two
   call-free clone scans exist to exclude — so `is_gc_unsafe_call` now exempts
   it, for both this tier and the element-shape tier.

The accumulator machinery lives in `stmt/stable_packed_accumulator.rs` (the
2,000-line file gate). `PERRY_PACKED_LOOP_NUMERIC_ACCUMULATOR=0` restores the
old lowering; the scan exemption is unconditional (it is a factual
classification, not a policy).

Isolated reduce loop (Mac, 1k elements): 5341 -> 993-1040 ns = node parity
(node 1009). wolf-ecs: +-0.08%, neutral. Differential vs node identical:
string accumulators (concat preserved via the slow clone), mixed-element
arrays (guard declines numeric mode), NaN/-0, in-loop reassignment to string
(admission declines), multiple accumulators, Math chains, update-form
counters. Kill switch output-identical.

Claude-Session: https://claude.ai/code/session_019WVcWKmYsUBnnFB7nBgbBJ

* fix(build-cache): restore #9044's registrations and register this PR's knob

Merging current main into this branch REVERTED #9044: commit e3164ee removed
`PERRY_BOX_CAPTURE_ENTRY_CELLS` and `PERRY_GUARDED_PREINLINE_MAX_IR_BYTES` from
BUILD_CACHE_ENV_VARS along with #9044's changelog fragment. The branch predates
that fix, so the commit was built over a stale tree and carries the removal as
an intentional-looking deletion -- which a merge then honours.

That alone would have re-reddened main: the assertion lives in a bin-crate unit
test, so its failure stops the whole `perry` test binary compiling and every
open PR's cargo-test job goes red.

This PR also adds a third codegen knob, PERRY_PACKED_LOOP_NUMERIC_ACCUMULATOR,
without registering it. It is a cache INPUT, not an exclusion: with it on,
`s += arr[i]` lowers to an inline fadd instead of
`js_dynamic_string_or_number_add`, so the two settings emit different code and
must never share a cached object.

All three registered; fragment renumbered 0000 -> 9060.

---------

Co-authored-by: Ralph Küpper <ralph@skelpo.com>