[transcendence-x]

Summary

Flash-Next turns can fail before the first generated token with QSA pooled hold 19843 rows; resizing to 8192 would cut them when reused cache metadata still describes a longer KV frontier. DSH Desktop surfaces this engine error as PI_AI_ERROR and ends the turn. Normalize pooled validity at the resize and extension consumers so newly completed blocks are recomputed, allow capacity planning to decline unsafe in-place resizes, and keep execution-stage would-cut refusals on eager verification without truncating live rows. Other exceptions still propagate.

Related: #592 and #605. Reproduction report: #592 (comment). This complements #650's state-setter fix; it does not change that setter or claim to cover every source of stale state.

Verification

Base: 9882703f3105363ddc37eca9f97aa09a1d387112 (main). The three affected source files match the app's original 2.12.2 runtime byte-for-byte.

• New deterministic regression file: 8 failures / 2 passes on unmodified main; 10 passes with this change. Covers the recorded pooled counts, active-prefix preservation, numerical pooled-key and fp32-mirror recomputation, other compression ratios, declined resize estimates and the eager execution fallback.
• python -m pytest tests/test_qsa_cache_resize_frontier.py tests/test_one_copy_reserve_resize.py tests/test_one_copy_admission.py tests/test_one_copy_conversation.py tests/test_qwen4_exp_qsa_cache.py: 96 passed.
• Installed development/server extras into a separate Python 3.14.3 venv. The required CONTRIBUTING command, python -m pytest tests/test_no_mlx_imports.py tests/test_public_cli.py tests/test_runtime_kpis.py, gives 343 passed / 14 failed, both here and on a clean archive of the unmodified base. The failing test IDs are identical: 8 test_one_shot_expect_python_validates_the_final_program cases, 3 test_one_shot_takes_its_prompt_from_a_pipe cases, test_one_shot_prefers_an_explicit_prompt_over_the_pipe, test_ask_reasoning_effort_reaches_the_prompt_encoder, and test_terminal_quickstart_max_uses_verified_max_session. The failures involve synthetic/mocked runtime imports; they are not reported as passing.
• python -m build: passed (sdist and wheel).
• scripts/fresh_venv_smoke.sh: passed no-MLX CLI survival checks.
• New test-file Ruff checks, Python compilation, git diff --check, repository hygiene scan and attribution scan: passed.

Benchmark Evidence

Correctness smoke coverage, not a throughput benchmark. The same functional patch was loaded by starting MTPLX.app from its UI, using Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed on an Apple M5 Max / 128 GiB / macOS 26.5.1, app Python 3.14.5, MLX 0.32.2, mlx-lm 0.31.3, on October 23, 2026 (Asia/Shanghai). Turbo, native MTP depth 3, ar_batch / agent, 512-token prefill chunks, SSD session cache on, context 262144, smart fan mode; requests used temperature 0 and thinking off.

Six OpenAI-compatible SSE requests all completed with CACHE_OK: 656 -> 68656 -> 68661 -> 31256 -> 15956 -> 660 input tokens. The continuation reused 68096 cached tokens; shortening the history exercised an actual promotion=resized to 32768 rows. No engine errors were recorded during this sequence. These synthetic requests verify recovery paths and cache reuse; the original private failing prompt was not replayed. No model weights, prompts, credentials or machine-specific paths are included in the change.