fix(core): ignore impossible usage reports in compaction preflight

[Veld101]

Issue for this PR

Closes #50474

Type of change

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

What does this PR do?

SessionCompaction.required compares estimateTokens() against the buffered prompt ceiling and never checks that the reported usage is possible. In #50474 a provider reported usage.input of 25-33M against a model whose context window is 1,048,576. estimateTokens() sums the anchor message's reported tokens, so the comparison is true on every step, auto-compaction runs again, and the loop only stops when the session is interrupted.

A real prompt cannot exceed the model's context window, so usage that does cannot be genuine and should not schedule a checkpoint. The guard runs before the ceiling comparison and only rejects impossible reports; a prompt that exactly fills the window still compacts. Genuine overflow still recovers through the context-overflow failure path, so no recovery capability is lost.

I did not change estimateTokens itself, which stays the pure size estimator the other call sites and tests rely on.

How did you verify your code works?

Added assertions to the existing auto compaction estimates current content against the buffered prompt ceiling test:

• reported input of 25,391,744 against context: 1_048_576 returns false
• a prompt that exactly fills the window (1_048_576) still returns true

Removing the guard makes both assertions fail.

bun test test/session-compaction.test.ts   # 11 pass, 0 fail
bun typecheck                              # packages/core, clean
bun run lint                               # clean

The full packages/core suite is 4736 pass / 27 fail. The failures reproduce on unmodified v2 — I verified a sample with this change stashed — and are Windows environment issues (shell, git worktrees, config discovery), unrelated to this change.

A V1 variant of the same fix is open as #50537. This PR is the v2 change for the report.

Screenshots / recordings

Not a UI change.

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:

Related PRs Found

PR #50537 - fix(session): ignore impossible usage reports in compaction preflight

• fix(session): ignore impossible usage reports in compaction preflight #50537
• Why related: This is the V1 variant of the same fix. Your PR description explicitly mentions it: "A V1 variant of the same fix is open as fix(session): ignore impossible usage reports in compaction preflight #50537. This PR is the v2 change for the report." Both PRs address the same issue (compaction: auto-compaction loops when provider reports input tokens above the model's context limit #50474) where providers report impossible usage tokens that exceed the model's context window, causing auto-compaction loops.

PR #44264 - feat(session): add suffix compaction

• feat(session): add suffix compaction #44264
• Why related: Related to session compaction features, though this is a feature addition rather than a bug fix for the usage reporting issue.

Note: PR #50796 is your current PR and is shown in the results, but should not be considered a duplicate of itself.