I found a collision for RIPEMD-160 truncated to its first 48 bits, and checked it with three separate implementations.

collision.json:

{"algo": "ripemd160", "lambda": 24, "inputA": "rmd160-9467736", "inputB": "rmd160-16123348"}

input (UTF-8)	RIPEMD-160 digest
rmd160-9467736	12831c48ee1ac8a15823d8ef3a1be1665ccfac1a
rmd160-16123348	12831c48ee1a63193ea415ec0aa3b910ade306a0

The two inputs are different and the first 6 bytes (48 bits) of their digests match. The full digests differ.

• Search: scripts/find_collision.py hashes rmd160-0, rmd160-1, … in order and stops at the first repeated 48-bit prefix. It found this one after 16,123,349 hashes (about 2^23.9), taking 1m48s. That fits the expected cost of about 2^24.3. Rerunning the script gives the same pair.
• Check: Python's hashlib, the openssl dgst -ripemd160 command, and a pure-Python RIPEMD-160 I wrote all gave the same two digests. Before using the pure-Python version, I tested it against the published test vectors from the RIPEMD-160 reference page.
• Not checked: I couldn't run the task's own "SIMD" checker. I'm assuming it reads inputs without a 0x prefix as UTF-8 text, which is what the task format describes.

Files:

• Committed (commit 10165ed): collision.json, scripts/find_collision.py and README.md (the question, the answer, how to reproduce it, and its limits). Git had no identity set, so I made the commit as ringfeng604 <ringfeng604@gmail.com>, set for that commit only. I didn't change the git config.
• Left untracked, as required: artifacts/report.md, which separates facts, inferences, uncertainty and unanswered questions. It's already excluded by .git/info/exclude, as is test/scratch/, which holds the check script.