# FSST: the table and output capacity are part of the contract

Ten valid constructed inputs exercise the authors' actual FSST8 decoder at
the revision recorded in source.json. These are API controls, not trained-table
compression benchmarks, malformed-input fuzzing, or a replication of paper timings.

## Replay

Obtain fsst.h from the URL in source.json. Keep its MIT license notice.
With Python 3 and a C++17 compiler:

```sh
python3 run.py --header /absolute/path/fsst.h --output replay.json --check observations.json
```

The runner checks the header hash, compiles probe.cpp at -O0 and -O2, checks output
against independent expected bytes, checks canaries outside the supplied output
capacity and compares both executions with the retained observations.

## What is being checked

Table A maps code 0 to the bytes ram; table B maps it to rom. Both map code 1
to abcdefgh. Code 255 escapes the following literal byte. Both tables use binary
mode, not optional zero-terminated mode.

For bytes 00 ff 21, table A yields ram! and table B yields rom!: equal encoded
bytes without equal table context do not identify equal decoded values.
With table A and capacity 2, the written bytes are ra but the returned length is
4. A caller must not treat that length as the number of valid bytes in its buffer.
This is documented truncation behavior, not a newly discovered decoder defect.

The remaining controls cover zero capacity in binary mode, an escaped ff byte,
empty input, the four-code fast path, an eight-byte symbol and its truncated
prefix, and the escape fast path. None exercises a serialized table importer,
encoder training, untrusted malformed streams, Unicode collation or hardware I/O.

Source, compiler and architecture are recorded in observations.json. The two
compiler settings on one machine are not independent platform replications.
The fixture makes no throughput, compression-ratio or safety claim beyond these
valid inputs. Input/output examples and Hesela code: CC0-1.0. Upstream code: MIT.
