runtime: js_regexp_exec keeps the subject borrow (and its Captures) live across the result-array, capture-string and groups allocations

[proggeramlug]

Follow-up to #8428 (fixed in #8446), same borrowed-heap-slice class as #8423.

Problem

#8446 closed the live window in js_regexp_exec — the one whose collection
point is user JS (ToLength(Get(R, "lastIndex")) running a coercible
lastIndex's valueOf). What it deliberately did not close is the
allocation-point window that follows.

After the match, crates/perry-runtime/src/regex/exec.rs keeps using str_data
(the &str into the subject's inline WTF-8 payload) and the regex::Captures /
fancy_regex::Captures borrowing it, across every one of these allocations:

• crate::array::js_array_alloc(caps.len()) — the result array
• js_string_from_str(m.as_str()) per capture — so capture N+1's as_str()
  reads a borrow that capture N's allocation could have invalidated
• crate::object::js_object_alloc(0, 0) + js_object_set_field_by_name — the
  named-capture groups object
• set_exec_array_metadata(arr, str_data, …) in the fancy arm — allocates the
  "index" key string, then copies input out of the borrow
• set_exec_array_indices / set_exec_array_indices_fancy
  (regex/exec_array.rs) — these allocate an array and a [start, end] pair per
  capture, and call byte_index_to_utf16_index(str_data, …) between the
  allocations

crates/perry-runtime/src/regex/match_string.rs (js_string_match, the
non-global fancy arm and the standard non-global arm) has the identical shape.

Why this is latent rather than live

The only collection points in this window are the allocations themselves, and
gc_check_trigger's nursery-churn arm calls force_full_scan() unconditionally
since #7682 — a conservative stack scan makes the copying minor ineligible, so
nothing moves. That is the same "closed by accident" property string/alloc.rs
documents for #7213 and #8423's write-up spells out, and it is exactly the
property the moving-GC roadmap keeps eroding. These sites should not depend on it.

Suggested fix

The pattern already exists in this module: regex/match_all.rs's
materialize_match_all_results runs a Phase 1 (borrowing, no JS allocation)
snapshot into owned Rust data, then a Phase 2 that allocates only from the
snapshot (see OwnedMatchAllData and its doc comment, which names this exact
hazard from the 2026-07-09 audit wave 1). Give exec and js_string_match the
same two-phase shape.

Two sub-decisions worth measuring rather than guessing, which is why this is not
folded into #8446:

1. Owned String per capture vs. byte ranges. match_all pays a
   to_string() per capture. exec is hotter (tokenizers, routers, ajv), so the
   cheaper shape is to snapshot Vec<Option<(byte_start, byte_len, utf16_len, flags)>>
   in phase 1 and build each capture with string::string_copy_range — which
   roots its source and re-reads after the allocation — for zero extra copies.
   Needs compute_utf16_len_wtf8 / bytes_have_lone_surrogate on the phase-1
   side so the WTF-8 flags stay exact.
2. set_exec_array_indices{,_fancy} can be fixed in place without touching
   exec's structure: precompute every byte_index_to_utf16_index into a
   Vec<Option<(f64, f64)>> before the first js_array_alloc, then allocate
   from that. Self-contained and worth doing first.

Also fold in: the fancy arm's .input (set_exec_array_metadata) copies the
subject out of the borrow after allocating the "index" key; the standard arm
already re-boxes the rooted subject via set_exec_array_metadata_groups_fresh.

Validation

• The witness shape is gc::tests::runtime_roots::regexp_last_index (added in
  fix(runtime): coerce lastIndex before borrowing the regex subject #8446): plant the collection, assert the subject was live and moved, then
  assert the captures. For this issue the collection has to be injected at an
  allocation rather than in a callback — force_next_general_arena_alloc_slow
  • GcTriggerThresholdTestGuard::make_arena_trigger_due (see
    runtime_roots/string_slice.rs) with the alloc-point scan pinned off
    (ConservativeScanDisabledGuard), otherwise the minor is non-moving and the
    test is vacuous.
• Sabotage-check it: the test must fail against the pre-fix code, or it is
  proving nothing.
• Watch for a regression on regex-heavy workloads if option 1 lands as owned
  Strings.

Workflow

PR = code + tests + changelog.d/<PR>-<slug>.md; no version bump.