Done. A real 48-bit SHA-256 collision, found locally and cross-verified.

Result

inputA	0x000000000017211c
inputB	0x00000000014060cc
sha256(A)	4e84dca19fa6·99674e41dc0ff0f92046577a96fd3bbc22ae6a036986dc268301
sha256(B)	4e84dca19fa6·7ad46fd0f73243c0081834a4767ce3fc91b0b4dade47ab574d2e
shared first 48 bits	4e84dca19fa6
hashes evaluated	20,996,301 = 1.252·2²⁴ (theory: 1.2533·2²⁴ — within 0.11%)

Deliverables (both untracked, as required)

• collision.json — 90 bytes, exactly the four specified keys, lambda as a JSON number
• artifacts/report.md — facts / inferences / uncertainty / open questions separated, each fact tied to a captured command output

Source (committed, df791a8): src/collide.py (search), src/verify.py (independent checker), README.md, .gitignore. Python stdlib only — no compiler existed in this environment (gcc/clang/make/cargo/go all absent), no third-party packages, no network needed at verification time, no submodules.

Checks run

• Three independent SHA-256 implementations agree bit-for-bit on both full digests: Python hashlib, the openssl dgst CLI (OpenSSL 3.0.13), Node.js crypto.
• Full digests differ — truncated collision only, not a SHA-256 break.
• Determinism: a from-scratch re-run produced a byte-identical file (cmp clean).
• The verifier is not vacuous: it rejects identical inputs, a non-colliding pair, algo:"md5", and an extra JSON key (exit 1 each), and handles the UTF-8 input form.

The one load-bearing assumption, flagged in the report: inputs use the 0x… hex form the task permits, so SIMD must hex-decode them to 8 raw bytes rather than hash the literal ASCII string. If its recomputation hashes the literal text instead, the check fails — hex is listed first in the task's own format spec, and no single pair can satisfy both readings.