ci: native-roots-aarch64 is red on main — #7330's invoke refusal makes 09_try_catch_roots uncompilable in five of its steps

[proggeramlug]

native-roots-aarch64 cannot pass on current main. Four of its steps run
09_try_catch_roots through the explicit statepoint bridge, and since #7330
that module refuses to compile — by design.

Nobody has seen it because GitHub's queue has not started a gc-native-roots
job in hours; every run on main and on open PRs is sitting in queued.

Reproduced locally at e98dfdb25

macOS/aarch64, cargo build --profile perry-dev -p perry -p perry-runtime-static -p perry-stdlib-static with the workflow's own RUSTFLAGS:

01_nursery_churn: ok        05_closure_capture: ok
02_survivor_promotion: ok   06_string_retention: ok
03_cross_gen_writes: ok     07_array_grow_evacuate: ok
04_dead_after_deep_stack: ok 08_map_set_sidetables: ok
09_try_catch_roots: COMPILE FAILED
statepoint forced-evacuation matrix: 8/9

thread 'perry-main' panicked at crates/perry-codegen/src/function/precise_roots.rs:722:17:
perry: native-root lowering cannot yet express a safepoint on an `invoke` —
`<indirect-or-unsupported>` in @perry_fn__09_try_catch_roots_ts__roundTrip has 1
live root(s) across it. … PERRY_RS4GC=1 handles invokes, but needs
PERRY_LLVM_CLANG pointing at a version-matched LLVM 22 …

Same result on aarch64-unknown-linux-gnu, so it is not host-specific.

The steps that break

All are in native-roots-aarch64, all use set -euo pipefail, so each dies at
the perry invocation:

1. Probe matrix, statepoint mode, forced evacuation — globs *.ts and
   asserts pass == total; total is derived from the glob (9) precisely so it
   cannot go stale, so it can only be satisfied by 09 compiling.
2. Safepoint-only contract changes codegen (differential) — globs *.ts
   and runs --statepoint-report=json on each.
3. Probe matrix, safepoint-only contract in strict mode — globs *.ts.
4. Probe matrix under both non-default walkers — globs *.ts.

And one more that is not a glob:

5. Root-pressure report fails closed (--statepoint-report) hardcodes
   probe=benchmarks/gc_ratchet/probes/09_try_catch_roots.ts. That step exists
   to assert plain_stack_maps == 0 and statepoint_fallbacks == 0, and it now
   cannot run at all.

native-roots-rs4gc-aarch64 is unaffected — RS4GC handles invokes.

Why this is awkward rather than obvious

09_try_catch_roots was added by #7314 because nothing in the suite contained
a try at all — 0 of 8 probes. Dropping it from the explicit-bridge matrix
gives that coverage straight back up, and #7330's own changelog is explicit that
"this leaves no working statepoint path for try-carrying code on a default
toolchain".

So the options are not equal:

• Assert the refusal (the statepoints-refuse-x86 pattern): the bridge must
  refuse 09 with that message, and 09 must still pass under RS4GC. Keeps the
  gate red if the refusal ever silently becomes a rootless binary. Cheapest, and
  it does not pretend the coverage exists.
• Skip 09 in the bridge arm. Same effect on CI, but nothing then notices if
  the refusal regresses.
• Teach the bridge to form a statepoint from an invoke — the statepoint
  becomes an invoke, with gc.result and the relocates in the normal successor.
  fix(gc): the statepoint bridge must refuse an invoke it cannot root (#7327) #7330 calls this "real work"; it is also the only option that restores the
  coverage rather than documenting its absence.

I have deliberately not reshaped the gate, since which of these is right is a
call about how long try-carrying code stays uncovered.

Note on the fifth step

Whatever is chosen, step 5's report probe has to move to a module the bridge can
compile, or its assertions are simply not running — which is CLAUDE.md's fourth
failure mode (the gate runs, its subject never did).