Picking up — most of the in-flight PRs (13 open) and issue #84 are gated on Markus review. Triaged sprint for things that don't need his sign-off:

• Bug-hunt sweep round 4 — Round 1 (server-models.cpp → PR fix(server): guard two mapping[] accesses against silent default-insert #77), round 2 (server-context.cpp → folded into PR fix(server): per-slot byte cap on context checkpoints (closes #67) #68), round 3 (server-queue.cpp → PR fix(server): clean up orphaned results when removing batch of task ids #83). Next adjacent server-* file. Candidates: `server-http.cpp`, `server-tools.cpp`, `server-chat.cpp`, `server-task.cpp` (only touched for Responses API: misleading error on context overflow, must communicate token limit exceeded #19, no scrub). Pick the one most likely to have concurrency / lifetime surface — `server-http.cpp` runs the response queue + chunked content provider, that's the prime suspect.
• Docs review round 2: `tools/server/README-dev.md` — internal docs. Verify the build/run instructions, endpoint internals notes, and the `/tools` route doc (which I saw README cross-link to without checking). Frequent target of drift since it ages outside CI.
• `scripts/bench-dflash.sh` post-master-sync drift check — the script defaults to Q8 (PR scripts(dflash): switch default bench target to Q8_0 + --target flag #65 era). Check it still resolves model paths, that the `--target` arg works, and that the JSON output schema matches what the bench parsers expect.
• Stale task cleanup in task list — task fix(server): return 400 (not 500) for truncated tool-call args in input (#19) #109 is "in_progress" but Markus hasn't bounced; chore(sync): upstream master → ht (114-commit sync 2026-06-13) #107/chore: gitignore UI node_modules and .svelte-kit build artifacts #110/Make a tag/release ggml-org/llama.cpp#111/Won't compile on a MacBook Pro M1 8 GB ggml-org/llama.cpp#123 are pending and gated on review. Set fix(server): return 400 (not 500) for truncated tool-call args in input (#19) #109 to "blocked-by" or leave a memory note.
• Inspect remaining unblocked memory items — `feedback_endurance_waivers.md` flagged the rebase-pain pattern; re-read the memories tagged feedback to make sure none of my recent autonomous moves drifted into a documented anti-pattern.

For later / saved (unchanged from epoch #81):

• DFlash + parallel >1 (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 + Markus signal
• Logit forward-parity (Task Make a tag/release ggml-org/llama.cpp#111) — needs 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 review on 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
• Add Native Gemini API Compatibility to Server #5 (Gemini API compat) — bigger surface, future epoch
• feat: analytical fit for large MoE models (avoid repeated model loads) #11 (analytical fit for MoE)
• Steering hints: mid-inference context injection via overlapping activations #17 (steering hints) — research-level
• TBQ KV-cache (tbq3_0 / tbq4_0): 65-73x perplexity regression — recommend mark experimental #70 (TBQ KV experimental marker via chore(args): mark tbq3_0 / tbq4_0 KV-cache as experimental (closes #70) #71) — awaiting review

Blocked on Markus / snoop-kube:

• 13 open PRs (fix(server): per-slot byte cap on context checkpoints (closes #67) #68, feat(fit): expose per-device byte plan from common_fit_params (#66 step 1) #69, chore(args): mark tbq3_0 / tbq4_0 KV-cache as experimental (closes #70) #71, feat(fit-params): --fit-print-plan emits per-device byte plan as JSON (#66 step 2 prep) #72, ci(sycl): drop auto-triggers (workflow has zero jobs upstream) #74, test(fit-params): smoke for --fit-print-plan JSON output #75, docs(server): drop dead chat.{mjs,sh} links + fix embedding example path #76, fix(server): guard two mapping[] accesses against silent default-insert #77, perf(dflash): gate top-5 logits diagnostic behind LLAMA_DFLASH_DEBUG #78, ci(cann): drop auto-triggers (workflow has zero jobs upstream) #80, fix(server): Responses API emits incomplete status on truncation (#19) #82, fix(server): clean up orphaned results when removing batch of task ids #83, docs(server): drop phantom --talker-model + --code2wav-model rows #85)
• CI audit: 11 self-hosted workflows queue forever on this fork #84 self-hosted runner policy (just filed)
• Task fix(server): return 400 (not 500) for truncated tool-call args in input (#19) #109 titan rollout