fix(tui): flag likely cache busts accurately - #38727

What

Make the debug turn-token view warn about likely prompt-cache busts without flagging known cache resets or OpenAI accounting noise.

The TUI now:

• renders unexplained comparable drops as ! Likely cache bust
• uses warning styling rather than neutral or error styling
• suppresses OpenAI movements between 1,024 and 2,048 tokens, which can reflect cache accounting quantization
• compares cache usage only when provider, model, and variant match
• starts a fresh comparison baseline after completed compaction

Before / After

Before

Every lower cached-token count was labeled a definitive Cache bust, including normal OpenAI quantization, comparisons across model switches, and expected cache resets after compaction. An intermediate revision suppressed those false positives but made every remaining drop a subdued neutral observation.

Cache reuse dropped by 62,464 tokens from the previous step

After

Known OpenAI bucket movements remain silent. Model switches and completed compactions establish a new baseline. A larger decrease between matching, uncompacted model steps is highlighted without claiming certainty about its cause.

! Likely cache bust: 62,464 fewer cached tokens than the previous step

How

• packages/tui/src/routes/session/rows.ts carries model identity with cross-turn cache baselines, clears the baseline at completed compaction boundaries, and centralizes reuse-drop classification.
• packages/tui/src/routes/session/index.tsx compares each step against the preceding comparable cache sample and renders unexplained drops with warning wording and color.
• packages/tui/test/cli/tui/session-rows.test.ts covers OpenAI threshold boundaries, non-OpenAI drops, model/variant switches, cross-turn baseline projection, and completed compaction resets.

Scope

This changes debug-only reporting. It does not claim a proven cause, alter model requests, change cache policy, modify compaction, or change token accounting.

Testing

• bun run test test/cli/tui/session-rows.test.ts: 15 passed, 0 failed
• bun run test in packages/tui: 519 passed, 5 skipped, 0 failed
• bun typecheck in packages/tui: passed
• pre-push bun turbo typecheck --concurrency=3: 33 tasks passed
• inspected real GPT-5.6 Sol session accounting: benign 1,024/2,048-token bucket movements are suppressed, while materially larger collapses remain reportable