[hermes-dflash-phone-cli][2/n] Preserve compute failures during cancellation

[OmarB97]

Mirror follow-up for Luce-Org#324 after #8 had already merged.\n\nThis covers the review-fix commit 47fd712 only: preserve completed ggml graph_compute failures ahead of cancellation checks so disconnect handling cannot mask a backend compute failure.\n\nVerification:\n- taro clean clone: cmake configure/build test_server_unit\n- taro clean clone: server/build/test_server_unit => 1647 assertions, 0 failures\n\nUpstream review context: cubic identified the masking issue on Luce-Org#324.

[OmarB97]

This is the replacement fork mirror for the current head of Luce-Org#324.

Why this exists:

• fix: cancel disconnected dflash requests #8 already merged before review-fix commit 47fd712.
• The upstream PR remains open and now points at 47fd712, so the old mirror only covers the earlier commit range.
• This PR mirrors the current upstream head for fork-side review/settlement.

Verification:

• Clean taro clone /tmp/lucebox-pr324-review-fix: CMake configure/build for test_server_unit succeeded.
• server/build/test_server_unit passed: 1647 assertions, 0 failures.

[OmarB97]

Post-merge reviewer dispatch (hermes-local-longctx-ko-mac)

Verdict: APPROVE (with non-blocking suggestions below)

Reviewed the full diff across 17 files (~757 additions). This is a substantial and well-structured change. Summary of findings:

What works well

• Consistent cancellation pattern: should_cancel() is uniformly applied across all backends (Gemma4, Laguna, Qwen3, Qwen3.5, Qwen3.5 MoE) and both spec-decode and AR paths. No backends were missed.
• Clean separation of concerns: DaemonComputeResult enum correctly prioritizes compute failures over cancellation (a graph error is a backend bug, not a client hang-up).
• Empty spec fallback: The retry mechanism for spec-decode producing only EOS/thinking tokens is well-designed. force_ar_decode flag cleanly bypasses spec without duplicating retry policy in each backend.
• Visible output tracking: visible_output_seen prevents caching snapshots/checkpoints when generation produced no client-visible output. Good optimization.
• Callback-based cancellation: CancelCallback on DaemonIO cleanly decouples HTTP server from backend layer.

Non-blocking suggestions

1. Thread-per-request overhead: RequestDisconnectWatcher spawns one std::thread per request. Under high concurrency this could be expensive. Consider a shared poller or edge-triggered epoll approach for production workloads.
2. Truncated tokenizer diff: The tokenizer.cpp changes were truncated by preflight. The encode(rendered, request_cancelled) call in http_server.cpp implies a new overload accepting a cancel callback — verify this compiles and the cancel callback is actually checked during tokenization.
3. README scope creep: The GPU selection flag documentation (--target-gpu, --draft-gpu, env vars) appears unrelated to the cancellation/empty-output-retry goal. Consider landing that in a separate PR.
4. Double-compute edge case: When spec-decode produces only EOS tokens and AR-decode also produces only EOS, both paths ran full compute. Rare but wasteful. Could add a check: if spec-decode hit immediate EOS, skip the AR retry since the model genuinely wants to stop.

Minor nit

• POLLRDHUP falls back to 0 on macOS. The recv+MSG_PEEK fallback still works, but macOS won't get the optimized early-half-close detection. Acceptable for correctness.