Merge train: #9794, #9780, #9774, #9814, #9756, #9796, #9775 (all previously conflicting)

[proggeramlug]

Merge train: #9794, #9780, #9774, #9814, #9756, #9796, #9775 — seven PRs that had been stuck on conflicts, all resolved here.

The conflicts, and what each actually needed

#9756 presented as 43 conflict hunks across 13 files of GC internals and needed zero hand-merging. Its description opens with "Stacked on #9755 (its commit is included; review only the last commit)", and #9755 landed in train #9817. Comparing commit subjects against main showed 6 of its 8 commits already there; picking the 2 genuinely missing ones applied clean. Reading the PR body first would have saved the most work of anything today.

#9814 was the one that needed real judgement. boxed_primitives.rs had a mid-function conflict where neither side alone was correct: #9794 adds boxed_string_wrapper_utf16_len, #9814 makes string indices virtual and roots the length install — and #9814's own descriptor_state.rs calls #9794's helper. --ours dropped the virtual indices (its test then crashed with TypeError: Cannot assign to read only property '0'); --theirs dropped the helper (unresolved import). The merge keeps #9814's rooted install plus #9794's helper, and drops install_string_wrapper_indices, which virtual indices make obsolete.

That merge then created a real bug, caught by the suite rather than the compiler: it left two synthesize paths, and #9814's string_wrapper::has_index early-return short-circuited above the stored-descriptor table probe, so Object.defineProperty on a wrapper index could never win. Synthesis now happens only in the fallback below the table. All five string_wrapper tests pass.

#9796 — union for two independent regex.rs additions, then #9801 (landed since) had added lazy.rs call sites using the old (String, String) cache key against #9796's new (Arc<str>, Arc<str>). Four stale sites in total, each surfacing at a different stage: compile, test-compile, and gate.

#9775 — main already has try_import_meta_require, so HEAD wins that file; its inprocess.rs work is kept.

#9794, #9780, #9774 — a gc/mod.rs entry-list union, a two-comment concat, and the census hot-TLS conversion respectively. check_thread_locals is now green on every file for the first time.

File-cap splits

Four files crossed the 2000-line gate: descriptor_state.rs → descriptor_state/gc_scan.rs; regex.rs → regex/program_key.rs and regex/replace_expand_fancy.rs; inprocess.rs → inprocess/optimize_emit.rs; lower/tests.rs → lower/tests/ui_widget_add_child.rs. Each verified in both feature configurations — a --features regex-engine build passes while the default one fails when a moved item sits behind that gate.

The raw_handle_debt --no-raise-vs invocation then correctly refused to let the new replace_expand_fancy.rs start with a ceiling: a brand-new file must have zero bare reads. Rather than grandfather it, all four moved sites became with_mut_ptr — preserving the deliberate per-iteration re-read after an allocating js_string_from_str — so total debt dropped by four instead of relocating.

Gate verdicts

SHAPE_CACHE_YOUNG's verdict was deleted because #9756 removes that log — the inventory failing a stale entry is it working as designed. NEVER_MATCH (Arc<regex::Regex>, a Rust-allocator program) classified. PASS1_MARKED's window re-pinned after #9794's diagnostics touched gc/mod.rs and gc/policy.rs: both gain declarations and counters only, and the census bracketing in step_mark_propagation / step_sweep is unchanged.

Validation

64/64 lint gates; perry-runtime, perry-codegen, perry-hir, perry-stdlib, perry-transform, perry-ext-http — run serially (RUST_TEST_THREADS=1).