All compression and context management parameters expressed as fractions of
max_tokens. This is the single control surface for the compression pipeline:
changing max_tokens (e.g. switching model) rescales every derived limit
automatically with no other configuration change required.
Host-counted provider request overhead and reserves deducted before context rendering.
These values are input facts, so configuring them through the kernel journal makes replay use
the same hard prompt allowance as the original run.