fix(provider): cap OUTPUT_TOKEN_MAX at 32K so usable() stays above zero

[NormalDudeBro]

Summary

\OUTPUT_TOKEN_MAX\ in the provider transform defaults to \Infinity\ when the \OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX\ env flag is unset.

For models whose \limit.output\ equals \limit.context\ with no \limit.input\ (e.g. \xai/grok-4.5: 500K/500K), \usable()\ in \session/overflow.ts\ computes:

\
usable = max(0, context - maxOutputTokens(model)) = max(0, 500K - 500K) = 0
\\

With \usable() = 0, \isOverflow()\ returns \count >= 0\ — true for every single message. Every finished step triggers auto-compaction, and the auto-continue step triggers another, repeating until the user aborts.

This restores the 32K output-reservation cap as the default (the fork's own compaction tests document \usable = context - output = 200K - 32K = 168K). The \OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX\ env flag remains as an explicit override.

Tests

• Regression test in \ est/session/compaction.test.ts\ for the grok-4.5 shape (context 500K, output 500K, no input limit): 31K usage must not overflow.
• \�un test test/session/compaction.test.ts\ — 49 pass
• \�un test test/plugin/cloudflare.test.ts test/session/llm.test.ts\ — 18 pass
• \�un typecheck\ (opencode package) — clean