Merge train 214: poll stride (removes the probe hack), dynamic-call ceilings, generator yields, timer ref state, BigInt bitwise (v0.5.1592)

[proggeramlug]

This train lands #10494, #10532, #10537, #10538 and #10540 as v0.5.1592, on 193889dd93. Ten source commits, each verified to preserve its patch-id and authorship.

• perf(regex): run the pre-search safepoint poll on one search in 64 (#10166) #10494 (#10166) — the pre-search safepoint poll runs on one search in 64 instead of every one: −438 instructions per call, about −9.1% on a hoisted .test(). The poll cannot cancel (nothing in production constructs EngineError::Cancelled) and does no stepping in practice — identical cycle_starts, completions and steps across 48M allocation-free calls — but it is strided rather than removed, because it keeps the one thing measurement could not rule out: the option of servicing a due collection from a loop that allocates nothing.
• fix(codegen,runtime): remove the argument-count ceilings on dynamic calls #10532 (#10420) — removes the argument-count ceilings on dynamic calls; more than 8 trailing args used to clamp to 9.
• fix(transform): suspend on yields in nested generator loop headers #10537 — suspends on yields in nested generator loop heads.
• fix(runtime): never evict a scheduled timer's ref state (#10447) #10538 (#10447) — never evicts a scheduled timer's ref state.
• fix(codegen): stop typing BigInt-capable bitwise results as int32 #10540 — stops typing BigInt-capable bitwise results as ints.

#10494 also removes a debug hack from main

main was carrying this at perex_runtime.rs:339, in place of the poll()? call:

// PROBE ONLY (#10166 poll experiment) — NEVER MERGE. Prices the
// pre-search safepoint poll by removing it. Unsafe by construction: ...

It reached main inside an unrelated guard commit — an uncommitted probe edit that git checkout -B carried onto a branch and git add -A swept in. Nothing caught it: the removal is functionally invisible (the poll cannot cancel and does no stepping, so every test and gate stays green), and the hunk is commented, which makes it read as deliberate to a reviewer. A comment saying NEVER MERGE is not a gate — and a commented hack survives review better than an uncommented one.

This PR replaces that hunk with poll_on_stride(poll)?, so the hack is removed and the poll restored in the same change, at the stride that was actually reviewed. Verified directly: no PROBE ONLY or NEVER MERGE remains anywhere under crates/.

Worth recording about #10494's own test, because it is a good pattern: the expected poll count is a literal, not SEARCHES / PRE_SEARCH_POLL_STRIDE. The derived form was self-consistent at every stride including 1, so it passed unchanged with the feature disabled — it asserted the arithmetic, not the behaviour. Both sabotage directions are run rather than reasoned: forcing stride 1 fails the pinned stride, and removing the call fails the count.

Validation

Validated head 16f4bf4417. Five-package release build pinned and hash-verified, and re-verified after the gap run.

• Crate suites: codegen 1571, runtime 3999, stdlib 139, hir 433, transform 137, cli 1139. The runtime count rises 3988 → 3999, which is the positive check that these PRs' new tests are registered and running rather than silently absent.
• All nine preflight gates pass. The holders gate is the one to note: fix(runtime): never evict a scheduled timer's ref state (#10447) #10538 removes the TIMER_HANDLE_KINDS entry while perf(regex): run the pre-search safepoint poll on one search in 64 (#10166) #10494 adds two, so the meaningful check is that the gate passes on the merged tree — 407 classified — not on either branch alone.
• Gap: filters timer, regex, generator, yield, bigint, bitwise, dynamic. Five of seven clean.

Reds attributed

Two fixtures A/B'd against main's own artifact set, both identical on both arms (node=1, main=0, train=0) with distinct build stamps per arm: test_dynamic_import_data_10104 and test_jwt_sign_dynamic_alg. That mattered here rather than being a formality — #10532 changes dynamic-call dispatch, so a verdict measured on binaries predating it would not transfer. test_timer is listed verbatim in run_parity_tests.sh's SKIP_TESTS on main (crypto.randomUUID() differs).

Two runtime failures, both release-profile artifacts rather than regressions. heap_generation::a_free_or_move_outside_every_scope_is_caught_in_debug_builds is main's long-known one, and copy_slot_decode::sabotaged_remembering_arm_is_refused_by_the_coverage_cross_check cannot pass under --release by construction: its observable is the restore_surviving_dirty_coverage cross-check, gated behind #[cfg(debug_assertions)] at copying.rs:1033. Confirmed under [profile.gcaudit] (release codegen with debug-assertions = true) on the previous train: 4 passed, 0 failed.

Held out of this train

#10530 — test_gap_cron_cronjob goes pass -> parity_fail, and it is not in gap_snapshot.json, so it is expected to pass. That is a test_gap_* fixture inside CI's actual gate scope, so it needs an explanation before it can ride.

#10539 — its cargo-test failure is commands::run::entry::tests::local_runtime_discovery_uses_install_layouts, in a file the PR does not touch (it is confined to fs/* and util_syserr.rs). perry bin tests are known to mutate global PATH, so this looks like interference rather than the PR's own defect — but that wants confirming, not assuming.

Before merging, the pushed head and unchanged main are checked again. After merging, the rewritten commits are checked for preserved authorship and the main tree must match the validated train exactly.