perf(codegen): entry resolution reaches Any-typed callback params — __commonJS-shape calls 7.4 → 4.5 ns (typed parity) - #9109

What

Entry resolution (#9071) reaches Any-typed callback parameters — the
esbuild __commonJS/__esm factory shape, where every module-init callback
erases to Any and paid the full js_closure_callN dispatcher per call:

• the emission gate in emit_callee_binding_resolutions admits a PARAMETER
  whose hint is absent or Any (a binding hinted as something
  non-function stays out — its resolution could only return null);
• the consuming arm's gate in early_branches.rs accepts a callee with a
  resolved-target map entry even when the Function hint is missing — the
  arm's runtime behavior (checked unbox, null-guarded direct diamond, full
  dispatcher fallback) was always hint-independent; the hint only selected
  who enters.

A resolution over a non-closure value (a number, undefined, a bound
function, a generator) returns null at entry and every call keeps the exact
dispatcher path.

Numbers

Single-shape 50M-call probes, quiet Linux, same-build
PERRY_CALLEE_BINDING_RESOLUTION=0 A/B:

The Any shape reaches exact parity with the typed shape. This is the
"partial until the param-hint gate lifts" band flagged in #9105 and in the
claude-code startup profile split — the cc --help re-profile is gated on
this PR landing (plus #9105 and the #9106 fix).

Semantics

Differential vs node identical: non-function/undefined through the Any
param (TypeError at the call), the param REASSIGNED mid-body (collector
excludes; the new value is observed), bound functions and ordinary functions
(this === undefined) through the Any param, arity-mismatch and rest
arrows, capture mutation between calls, generator functions. Kill-switch build
output-identical; the #9105 corpora re-run clean (the one diff is the
documented pre-existing no-TDZ-for-globals TypeError vs ReferenceError
line, gate-independent).

Testing

• RUSTFLAGS=-D warnings cargo check (host excludes) — clean;
  perry-codegen 1830/0 — green.
• perry-runtime --lib: 1804 tests pass, then the suite ABORTS at
  object::reserved_floor::…::user_properties_read_back_through_the_get_path_at_scale
  (shapes.rs:1539 ShapeId/ObjectHeader disagreement, double panic) — a main
  regression filed as regression(main): shapes.rs ShapeId/ObjectHeader disagreement aborts the entire perry-runtime test suite (SIGABRT, ~1000 tests masked) #9108 that masks the suite tail for every branch;
  this PR contains no perry-runtime files.
• Pinned-worktree integration battery; note issue_8690 /
  issue_8773 carry one pre-existing failure each on current main
  (regression(main): wolf-ecs-shaped nested subclass loops lost their versioned fast clones (packed-guard sites 3 → 0) #9106, the versioned-loop-clone regression — window
  8b40634abd..f3f405271c, owned by the claudecode session), identical on
  base and branch.
• Lint: census, address-class, gc-store-site, file-size, raw-handle debt.

Summary by CodeRabbit

• Bug Fixes
  • Improved handling of direct calls through untyped or broadly typed callback parameters.
  • Enhanced compatibility with bundled module factory patterns.
  • Improved closure call performance by using faster resolution paths when safe.
  • Preserved existing runtime behavior with fallback safeguards for other callback types.