Sprint #2. Previous: #73 (closed, 5/5 done).

Sprint — top 5

• Triage stale issues — surveyed Rebase conflict: ht on master (2026-05-18) #51, ht-llama-webui Tauri shell: bind dual-stack on port 5173 to prevent IPv6 squatter hijack #50, webui: 120-error strategy-(b) Props re-derivation against post-#22937 upstream surface #45, WebUI: evolve into general-purpose AI workspace with virtual filesystem and artifact rendering #31. Rebase conflict: ht on master (2026-05-18) #51 is superseded by the 2026-06-04 history rewrite + master sync; ht-llama-webui Tauri shell: bind dual-stack on port 5173 to prevent IPv6 squatter hijack #50/webui: 120-error strategy-(b) Props re-derivation against post-#22937 upstream surface #45/WebUI: evolve into general-purpose AI workspace with virtual filesystem and artifact rendering #31 reference `webui` which lives in heierchat now. Close with cross-link comments; transfer scope if any of them describes work still relevant on the ht side.
• Verify Responses API: misleading error on context overflow, must communicate token limit exceeded #19 (Responses API context-overflow 500 → token-limit message) and chat/completions endpoint returns 500 when mmproj is loaded (Qwen3.5-27B VLM) #16 (chat/completions 500 with mmproj loaded) still reproduce on post-rewrite ht. If they do, close with repro recipe attached or open a fix PR. If they don't, close as fixed-by-rewrite or upstream-merge.
• Tests for PR fix(server): per-slot byte cap on context checkpoints (closes #67) #68 (per-slot checkpoint byte cap) — added a CLI flag + eviction logic without test coverage. Pattern from PR test(fit-params): smoke for --fit-print-plan JSON output #75 (`tools/fit-params/tests.sh`): drive llama-server with low `--ctx-checkpoints-max-mib` + long prompts, verify the `reason=count|bytes` log fires on the byte path.
• Bug-hunt sweep round 2 on `tools/server/server-context.cpp` — adjacent file to last epoch's server-models.cpp; same heierchat-500s territory. Look specifically at slot lifecycle, checkpoint create/restore, and the seq_id race surface that chore(sync): upstream master → ht (114-commit sync 2026-06-13) #107 touches.
• More CI workflow audit — last epoch fixed ci(sycl): drop auto-triggers (workflow has zero jobs upstream) #74 (build-sycl.yml). Other workflows may have similar zero-job-on-master triggers post upstream cost-cutting. Scan `.github/workflows/*.yml` for the same pattern.

Saved for later

• DFlash + --parallel >1 per-seq_id (Task chore(sync): upstream master → ht (114-commit sync 2026-06-13) #107) — gated on Router co-loads same-device-pinned models and OOMs: --models-max fit decision ignores per-device VRAM #66 step 5 + canary
• Logit forward-parity (Task Make a tag/release ggml-org/llama.cpp#111) — 6 GB safetensors download + PyTorch env
• 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) — awaiting Markus's review of foundation PRs 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
• TBQ KV-cache CPU SEGFAULT (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 / PR chore(args): mark tbq3_0 / tbq4_0 KV-cache as experimental (closes #70) #71
• test-quantize-perf SEGFAULT windows-x64 (android port of llama.cpp ggml-org/llama.cpp#124) — needs Windows env

Created per `~/IDLE.md`. Closes when sprint cleared.