Add calibrated native fixed-limit service evidence

[CompleteDotTech]

Scope

• Add a portable local authenticated service probe that regenerates an exact 32,768-branch / 65,536-aggregate request with the pinned tokenizer and submits measured one-token-over requests at the fixed limits.
• Publish a compact, content-free receipt linking the trained and calibrated checkpoint hashes, exact/overflow behavior, 255-choice and ten-level shape evidence, and the retained six-stratum marker failure.

Evidence

• Pinned calibrated checkpoint SHA-256 a731903c8e91adc6206ecf60325b6ebbf4cdedd1349f0ab831c85069cf37df3d, verified exposure chain: two optimizer steps at both exact limits.
• RTX 3060, torch 2.10.0+cu128, 70% process cap. Exact local service request: HTTP 200, 19 answers, 65,536 input tokens, 67.42 s, 4,666,471,936 peak allocated bytes. Aggregate 65,537 and branch 32,769: HTTP 422 with correct limit details.
• Six-stratum numerical and overflow checks passed 6/6 each; marker decisions remain 6/18 against the unchanged 80% threshold. Issue Prove trained native 32k branch and 64k request behavior #4 remains open.

Validation

• Portable runner rerun on the pinned calibrated checkpoint: HTTP 200 / 422 / 422, private receipt SHA-256 008e6f21b48a108acaeeb0b38a2e138608ae84240100385e02848871d8080f69.
• python -m pytest -q: 858 passed, 31 skipped.
• Compileall, critical Ruff, JSON parse, staged source audit (214 files), and diff check passed.

No model weights, requests, credentials, or private paths are included in the source change.