[jadenfix]

Valid JSON Pointers through an empty root key currently resolve incorrectly. For example, a schema whose value property references #/ should use the schema under the root's "" key; instead a valid integer value can be rejected as an object. #//child also drops the initial empty key.

Two changes are needed for that consumer path:

• Preserve empty leading reference tokens, while retaining an empty fragment as the whole document.
• Assign stable unique names to referenced validators. The root and #/ otherwise sanitize to the same function name and overwrite one another. Ordinary non-colliding names remain unchanged.

The 16 focused tests cover local and handler-loaded references through both compile() and compile_to_code(), empty keys, ordinary/root/array references, tilde and percent escapes, name collisions, and URI alias reuse. On the baseline, 15 fail and the ordinary-reference control passes.

Validation:

• Python 3.14 and the supported floor Python 3.10: full CI command pytest -m "not benchmark": 4,525 passed, 12 deselected, 449 xfailed, 936 non-strict xpassed. Unchanged baseline: 4,509 passed with the identical deselection/xfail/XPASS counts.
• Installed wheel on Python 3.10: real schema/data files through compiled validators and generated Python modules pass 18 checks, including valid values and both child/root type rejection. The baseline rejects the same valid file input.
• Four ordinary schemas produce byte-identical generated code before and after the change.
• Changed resolver module: Pylint 10/10; git diff --check passes. Whole-project Pylint retains six identical baseline diagnostics in unrelated modules; it is separate from the required CI pytest command.

This is a focused reference-resolution fix, not a claim of complete JSON Pointer conformance. No external schema network requests were made; handler coverage uses in-memory schema data. Hosted CI and other platforms remain pending.

AI assistance: Codex prepared the patch, tests, and description. A separate Codex agent reviewed the final resolver and generator call sites, independently ran all 16 focused tests, and executed the installed-wheel consumer. This is agent review, not human review.