Hivemind-driven maintenance sprint for ht-llama.cpp. Pick items in triaged order; cross off as done; append "saved for later" at the bottom.

Triage criteria (per ~/IDLE.md)

1. Obviously wanted (lowest risk of decision drift)
2. Low-hanging fruit
3. How much we want it
4. Difficulty

Active sprint — top 5

• CI health on origin/ht — build-sycl.yml has been failing on every recent commit (09b2124fb, b0daec55b). Investigate whether it's a pre-existing upstream issue or a real ht regression. If pre-existing, file a tracking task; if our problem, fix.
• Tests for --fit-print-plan (PR feat(fit-params): --fit-print-plan emits per-device byte plan as JSON (#66 step 2 prep) #72 added the flag with no test coverage). Add a smoke test that verifies the JSON output format. Either in tests/ as a small new test, or as a shellcheck-style script.
• Docs review on tools/server/README.md — the heierchat decoupling left some pointers that may be stale. Verify all cross-repo links resolve.
• Bug-hunt sweep on tools/server/server-models.cpp — 1500+ LOC of router state machine handling subprocess lifecycle, the area where heierchat-500s incident lives. Look for obvious logical bugs (race conditions, missing error paths, off-by-one on reserved-style accounting).
• Profile/bench scout for DFlash speculative decoding — per IDLE.md HPC clause: identify the hottest path in llama-speculative-simple against gemma-4-31B target. Once titan rolls, scripts/bench-dflash.sh --target Q8_0 measures throughput; understand which hotspot we'd target if accept rate stays at ~8%.

Saved for later

• DFlash + --parallel >1 per-seq_id (Task chore(sync): upstream master → ht (114-commit sync 2026-06-13) #107) — substantial refactor, defer until enforce-mode router (Router co-loads same-device-pinned models and OOMs: --models-max fit decision ignores per-device VRAM #66 step 5) lands
• Logit forward-parity test (Task Make a tag/release ggml-org/llama.cpp#111) — needs 6 GB z-lab safetensors download + PyTorch env
• Implement Router co-loads same-device-pinned models and OOMs: --models-max fit decision ignores per-device VRAM #66 steps 3-5 (Task Won't compile on a MacBook Pro M1 8 GB ggml-org/llama.cpp#123) — substantial, awaiting Markus's review of feat(fit): expose per-device byte plan from common_fit_params (#66 step 1) #69/feat(fit-params): --fit-print-plan emits per-device byte plan as JSON (#66 step 2 prep) #72 foundation PRs first
• Windows test-quantize-perf SEGFAULT (Task android port of llama.cpp ggml-org/llama.cpp#124) — needs Windows env to reproduce
• TBQ KV-cache CPU SEGFAULT (Task Fix build for Android ggml-org/llama.cpp#125) — superseded by issue TBQ KV-cache (tbq3_0 / tbq4_0): 65-73x perplexity regression — recommend mark experimental #70 (TBQ KV-cache experimental marking, PR chore(args): mark tbq3_0 / tbq4_0 KV-cache as experimental (closes #70) #71)

Out of scope (not maintenance)

• Anything that adds new features without an issue / RFC backing it
• Reviewing PRs from non-approved authors (none currently; all 4 open PRs are by marksverdhei)

Created per ~/IDLE.md guidance. Will close when sprint cleared.