fix: keep mapping-key guidance off native hot paths

[goblinmode2700]

What changed

• Restore src/encode.rs exactly to the published v0.3.0b4 native source.
• Rewrite the fixed non-string-key message through the existing Python-supplied encode-error constructor.
• Preserve the Issue 13 API behavior and the Python buffer-normalization membrane unchanged.

Why

The tag-time 0.3.0b5 guard established regressions in all four typed/functional encode endpoints while untyped encode remained non-inferior. The only remaining native source difference from b4 was the cold mapping-key error text inside encode.rs. This change removes that native layout perturbation entirely; the error factory is called only after an encode error has already occurred.

Performance evidence that rejected the prior candidate

R 4.4.2 analyzed 36 randomized baseline/current pairs across the preregistered 22-endpoint family using paired log process means and simultaneous Bonferroni intervals:

• typed encode @64: +6.7% [+5.2%, +8.3%], regressed
• typed encode @4096: +6.5% [+4.9%, +8.1%], regressed
• functional encode @16: +5.6% [+3.5%, +7.8%], regressed
• functional encode @4096: +6.8% [+4.6%, +9.1%], regressed

The lower bound exceeds the +3% margin for each endpoint, so these are established regressions. No unchanged-candidate rerun was performed.

Validation

• git diff v0.3.0b4 -- src: empty
• uv run pytest -q tests/test_native_encode_types.py: 40 passed
• make qualify: 488 Python tests passed, 11 skipped; 43 Rust tests passed; official TOON corpus 538/538; G2 zero-intermediate-container proof passed

The next performance decision belongs to the release workflow’s R-owned guard on verified release wheels.