fix(opencode): honest context arithmetic for small and unreported model limits

[yanglinfang]

Issue for this PR

Closes #41372

Second slice of #44242 (being split into small PRs per review feedback; first slice #44895). #41372 is the umbrella issue for the series — if this merge auto-closes it before the remaining slices land, it should be reopened.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Fixes the context arithmetic that makes opencode unusable with small local models:

1. Unset limit.context disabled compaction entirely. Routers and local providers frequently report 0; sessions then ran straight into provider overflow errors. Now a conservative 32k usable window applies, shrunk further by a per-session cap learned from actual provider overflow rejections (logged once per session).
2. Unset limit.output was assumed to be 32k. usable = context − 32000 reaches 0 on any window under 32k, so every completed step counts as overflow and compaction re-triggers after every single step. The fallback is now window-proportional (25%, floor 1024), and each request additionally clamps its output budget to the headroom the estimated input actually leaves, so a long history can't push a request over the window.
3. The compaction reserve was effectively fixed at 20k — 36% of a 56k local window. Now proportional: min(20k, max(2048, 15% of window)). compaction.reserved config keeps absolute priority and windows ≥133k keep today's 20k unchanged.

How did you verify your code works?

New unit suites for the reserve formula, usable-window derivation, learned caps, and the per-request clamp; the changed upstream expectation (context: 0 → conservative window instead of never-overflow) is updated with both directions asserted. Full packages/opencode suite green on current dev.

End-to-end A/B on the same machine, model (qwen3:4b-instruct via ollama), config (context: 16384, output: 0), and coding task:

• without this fix: 30 compactions in 15 minutes — one after every step; the model re-reads the same file forever, zero files changed, killed at the cap
• with it: 9 steps, 0 compactions, task completed and verified in ~7 minutes

Screenshots / recordings

Evidence panels for the A/B run to be attached below (no UI change in the code itself).

before and after fix

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on the search results, I found one related PR (excluding the current PR #44898 which appears in results):

Related PR:

• feat(opencode): model capability tiers for small/local models #44242 - feat(opencode): model capability tiers for small/local models
  • This is the parent issue mentioned in the current PR's description as the origin of this work (being split into smaller PRs). The current PR (fix(opencode): honest context arithmetic for small and unreported model limits #44898) is explicitly noted as "second slice of feat(opencode): model capability tiers for small/local models #44242" after the first slice fix(opencode): deterministic plugin load order and hook error isolation #44895.

However, this is not a duplicate but rather a parent/predecessor PR as part of the same feature split. The current PR is a focused bug fix for context arithmetic that's part of a larger effort to support small/local models.

No duplicate PRs found