other native entry points taking strings may mis-unbox SSO strings (as #11430/#11486 did)

[proggeramlug]

#11430 (fixed by #11468 and #11486) had one root cause, found in two places. Since #10762, String(n), `${n}` and n.toString() return small values as SSO strings: the characters live inline in the NaN-box, with no heap StringHeader behind them. A native entry that unboxes a string argument with bits & POINTER_MASK (0x0000_FFFF_FFFF_FFFF) and casts the result to *const StringHeader turns those inline characters into an address, then segfaults or reads garbage. Code that first checks tag == 0x7FFF (heap string only) avoids the crash but treats the SSO value as "not a string", so it returns a wrong answer.

Both sites found so far were fixed only locally:

• Buffer.write and the <encoding>Write family, plus indexOf / lastIndexOf / includes needles (fix(runtime): Buffer write/indexOf accept an SSO string; parameterised pg queries no longer segfault (#11430) #11468).
• Every crypto native: hash.update, the createHmac key, pbkdf2Sync / hkdfSync inputs, cipher update (fix(codegen,runtime): crypto natives accept an SSO string argument; hash.update(String(n)) no longer segfaults (#11430 follow-up) #11486, via the new js_ffi_arg_ptr).

The pattern is systematic and there are more candidates. This is a grep on main at c1d93bb, not an audit; each hit still needs a look to decide whether an SSO value can reach it.

Runtime and stdlib: masked bits cast to *const StringHeader (count per file, tests excluded)

file	hits
perry-runtime/src/proxy/put_value.rs	7
perry-runtime/src/json/replacer.rs	7
perry-runtime/src/json/stringify.rs	4
perry-runtime/src/object/buffer_dispatch.rs	4 (the hasOwnProperty / propertyIsEnumerable key and key-export format readers; #11468 fixed only the write arms)
perry-stdlib/src/fetch_blob.rs	3
perry-runtime/src/thread.rs	3
perry-runtime/src/string/concat.rs	3
perry-runtime/src/string/char_ops.rs	3
perry-runtime/src/array/subclass.rs, array/named_props.rs	3 each
perry-runtime/src/{map,set}.rs, value/dynamic_object.rs, symbol/constructors.rs, json/{stringify_shape_template,stringify_primitive_object,stringify_flat,raw_json}.rs	2 each

Heap-only tag checks (== 0x7FFF) in files that also read a StringHeader, which gives wrong answers rather than crashes: string/concat.rs, object/class_registry/class_meta.rs, array/{from_concat,generic,flat_clone,indexing_keyed}.rs, typed_feedback.rs, date.rs, proxy.rs, value/{dyn_index,dynamic_object}.rs, {map,set}.rs, typedarray/mod.rs, buffer/u8_codec.rs, and in perry-stdlib common/dispatch/sqlite.rs, fetch_blob.rs, webcrypto/util.rs.

Codegen: unbox_to_i64 passed to natives whose parameter is a string (the shape #11486 fixed in the crypto arms). The largest non-crypto users: lower_array_method.rs (28), expr/arrays_finds.rs (19), expr/logical_collections.rs (15), expr/url_main.rs (13), expr/instance_misc1.rs (13), expr/misc_methods.rs (12), lower_call/native/native_tui_layout_branch.rs (11), native_instance_branch.rs (8), lower_call/namespace_call.rs (7), expr/string_regex_proc.rs (7), expr/os_uri_dates.rs (7). Most of these unbox arrays or objects, which is fine. Only the ones whose runtime parameter is a StringHeader are affected, for example url_main.rs and string_regex_proc.rs.

Suggested guard. Add a ratchet in the style of string_payload_access_inventory.py that counts, per crate:

1. <mask> as *const …StringHeader casts (and (bits & POINTER_MASK) as usize as *const StringHeader) outside value/nanbox.rs.
2. (… >> 48) == 0x7FFF string-tag checks that are not paired with is_short_string() in the same function.
3. In perry-codegen, unbox_to_i64( whose result is passed to a runtime function declared with a string parameter. This would need a small table of which js_* params are strings; the runtime_decls modules already name them.

New string arguments should go through js_ffi_arg_ptr (scratch copy, for natives that only read during the call) or js_get_string_pointer_unified / OwnedStringBytes (a heap or owned copy). The baseline would lock in today's counts, so each fix lowers it and new code cannot add to it.

A cheap way to find the live ones is a differential sweep: take existing gap tests that pass a string literal to a builtin, replace the literal with String(<short number>) or a template, and compare against Node. #11468 and #11486 each had a program that crashed outright on main this way.

[proggeramlug]

Progress: #11627 landed. About 40 native entry points that mis-unboxed short (SSO) strings are fixed: crashes in Date.parse, child_process, JSON.parse with a reviver and URLSearchParams, hangs from short node:http event names, and wrong answers. A new lint gate, scripts/sso_unbox_inventory.py, keeps new entry points from regressing. Keeping this open for any remaining sites the inventory lists.