Clarify experiment bookkeeping in program.md

[Gabrieliam42]

Summary

• clarify that train.py is the only source file to edit
• explicitly allow results.tsv bookkeeping and run.log updates as experiment artifacts
• require the results.tsv row to be appended before any reset, checkout, or revert

Why

The previous wording could be read as forbidding the log and results bookkeeping that the experiment loop itself requires. This makes the intended workflow explicit without broadening what source files may be changed.

[elementalcollision]

Technically, program.md is also a source file to edit, as you can designate which dataset you want to experiment on. I don't necessarily view this as ambiguous as even the agent (in my case Opus and/or Sonnet) understands which files are immutable and which aren't.

[Gabrieliam42]

All right,
It's just that when I ran it with Codex on VS Code it didn't update results.tsv or run.log because it interpreted that train.py directive as a override restriction, so it ignored the other directives, therefore it didn't work as intended with Codex and GPT-5.4, probably Claude is more permissive. After I modified it as stated it began to work as the project intended.