Merge train: #9912, #9917

[proggeramlug]

Merge train: #9912, #9917.

#9917 against #9970

#9917 collapses the three RegExp.prototype.test proof slots into one CanonicalTestSite behind a single HotKey — one TLS address resolved per call instead of three. It was written before #9970 (landed in #9978) gated that surface for the no-default-features product build, so taking its side wholesale would have re-broken that build. Merged so both hold: the unified record, carrying #9970's cfg on the struct, its impl, the thread_local! and TEST_ACCESSOR_KEY_BIT.

Three consequences reconciled rather than left to surface later:

• object_static_prototype_known_non_meta is perf(intl): Segments view — one realm record for the test proof, fixed-slot cursor fields, no duplicate pointer validation #9917's, and its only caller is regex-engine gated, so it is dead in a product build. Same gate.
• The three _SLOT holder entries added for perf(intl): take the view mode's canonicality proof and UTF-8 re-validation off the per-grapheme path #9893 are stale — those statics no longer exist. Deleted, along with a duplicate TEST_WALKS entry. The replacement REGEXP_PROTOTYPE_TEST_SITE needs no entry: the gate's walk reaches it through scan_canonical_test_site_roots_mut, which scan_object_cache_roots_mut calls and reg_scanner! registers. Verified, not assumed.
• segments_view's cursor-number store carried GC_STORE_AUDIT(NUMBER), and NUMBER is not an accepted class — the marker never satisfied the gate. POINTER_FREE is the class that states what its comment argues.

#9912 turns the segment-view lowering on by default

Its blocker is resolved: the runtime entry points it needs landed in #9870 (#9888). Measured on claude-code, the tier with #9893's levers is −14 % CPU and −30…−42 MB peak RSS with identical output. PERRY_SEGVIEW=0 remains as an opt-out, deliberately, because a switch that changes emitted code needs an off position that does not require rebuilding the compiler.

It also adds a test asserting the default really is ON — written because a typo making the predicate always-false would leave every direct-call test passing while the tier silently never fired in a real compile.

A note on how one fix here was found

The first validation run of this train died on ENOSPC. perry-codegen reported RC=101 with zero tests, and the -D warnings all-targets gate failed alongside it. It would have been easy to write both off as disk collateral. Re-running them individually on a healthy disk showed codegen was collateral (1948 passed, 0 failed) and the warnings gate was not: #9917's test nests an unsafe block inside one that already covers it, which only -D warnings on test targets rejects. An ENOSPC run mixes spurious failures with real ones, so each has to be re-run to tell them apart.

Validation

run_lint_gates: 3 of 80 FAILED; 0 ENOSPC, all three accounted for: public-baseline is pre-existing on main (verified on a pristine worktree, red since 2026-07-29), and the two API docs gates trip on this session's CARGO_TARGET_DIR override — with the binary where regen_api_docs.sh expects it, regeneration succeeds and the drift check is clean.

suite	passed	failed
perry-runtime	3313	0
perry-codegen	1948	0
perry-hir	630	0
perry-stdlib	132	0